Classification of faults in friction stir processed composites using a machine learning and ensemble learning approach

Classification of faults in friction stir processed composites using a machine learning and ensemble learning approach | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Classification of faults in friction stir processed composites using a machine learning and ensemble learning approach Pragya Saxena, Arun Bongale, Satish Kumar, Rajesh Kodbal This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4834721/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 4 You are reading this latest preprint version Abstract Aluminium alloy based surface composites with hard reinforcement particles have wide scope in aerospace and automobile manufacturing industries. In this paper, the aluminium composites, manufactured by friction stir processing (FSP) with varying parameters are investigated for the faults occurred during fabrication process. It explores a machine-learning approach to detect defects of surface hybrid composites with an Al6061 alloy matrix, reinforced with copper and graphene nano-powders, using friction stir processing and a tungsten carbide tool on a milling machine. Multi-sensor time series data (vibration, force, and current) collected during fabrication, is preprocessed and labelled with normal and defective categories (e.g., pin break, brazing break, rough surface, no composite) using visual inspection. The important time domain and frequency domain features are extracted using different libraries in python. Thenafter, various types of feature selection techniques, viz filter, wrapper and embedded methods are implemented to select most relevant features. The selected subset of features from all selection methods used, are applied to different machine learning and ensemble learning classifiers and their performances are evaluated. The optimal combinations of the type of feature selection method and classifier used, are obtained for efficient classification of surface defects in composited formed by FSP. The real time monitoring and defect detection system can be developed in future for the composites developed by FSP using the developed models. composites defects feature selection sensor data machine learning classification Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 1 Introduction Friction stir processing (FSP) has recently been established as the advanced method not only to alter the exterior microstructure and characteristics of substance but also to create high-quality surface composites through localized plastic deformation without undergoing melting of the material. Thus, it is an environmentally friendly technique for manufacturing surface composites. FSP has recently been utilized to obtain refined grain microstructure in aluminium alloy-based composites by adding suitable reinforcements and other metal alloys. It is based upon the principles of Friction Stir Welding (FSW) developed by researchers at The Welding Institute, UK [ 1 ], for joining two metal pieces utilizing the heat caused by friction. The joining of dissimilar metals, ultrasonic-assisted, and underwater joinings is obtained much more efficiently with friction stir welding technologies. FSP is utilized significantly in various applications, such as aerospace [ 2 ] and the automobile sector, requiring improved physical, chemical, and surface mechanical properties of metallic materials. Recent studies indicate that fundamental mechanisms drive significant microstructural and textural changes during FSW or FSP of different metals and alloys, resulting in a promising technique for fabricating surface hybrid composites [ 3 ]. FSP consists of a specifically shaped tool with a shoulder integrated with a pin, which, in rotating condition, is moved in the direction of the surface of the metal workpiece as fixed on a machine bed. The pin, integrated into the shoulder, is pushed inside the workpiece so that tool shoulder comes in complete contact with the metal surface. The FSP tool is then passed over the specified lane over the workpiece surface to be processed. The thermal energy is produced due to the friction produced when the tool shoulder slides over the metal surface, ensures localized recrystallization on the workpiece surface [ 4 ]. The recrystallization helps refine the microstructure of the grain near the matrix surface in the traversed region. Recent research includes improving the mechanical, thermal, and microstructural properties of the composites manufactured by FSP, with a wide range of mechanical tests optimizing the process parameters [ 5 ]. The influence of various factors on the surface characteristics is beneficial in determining the ideal range of values for achieving the desired results. However, today’s industry demands high production in less time with fewer defects, which leads to the application of Industry 4.0 in FSW and FSP of metal alloys or composites. The application of a machine learning approach for real-time monitoring surface composite manufacturing using the FSP technique with the help of multi-sensor data analysis is very rare. The recent research in this field includes the condition monitoring of the FSW and FSP of the composites utilizing machine learning techniques with the collected sensor data. Machine learning methods significantly regulate the process parameters and monitor friction stir welding and processing techniques. As surveyed by Ammar H. Elsheikh et al. [ 6 ], the utilization of machine learning techniques in FSW was examined. The study focused on predicting joint characteristics, online process csontrol, tool failure analysis, and the application of optimization techniques with machine learning methods. The study encompassed various methodologies, including linear regression, kNN, random forest, ANN, SVM, fuzzy system, adaptive neuro-fuzzy inference system, and random vector functional link [ 7 ]. Shivraman Thapliyal et al. [ 8 ] utilized a machine learning-based model to investigate how input factors influence the physical characteristics during the copper welding by FSW. It was observed that the neural network obtained the maximum accuracy using a deep leaning model in predicting the ultimate tensile strength over kNN, decision tree, and information gain methods. Researchers are now attempting to employ real-time monitoring by integrating multiple sensors throughout the FSP. Bipul Das et al. [ 9 ] utilized the online torque sensor data recorded during the FSW process by using the discrete wavelet technique to identify the process faults. Also, the feature selection methods employ the chosen features in the support vector machine to predict the tensile strength of the joints formed by the FSW of AA1100. In another similar study by Sangyoung Yoon et al. [ 10 ], defects are identified in the laminated composites using deep learning by highly non-linear solitary waves. The several characteristics influence the classification efficiency of the deep learning algorithms which is investigated. It is observed that the proposed algorithm achieved high accuracies in recognizing the defects and locating them on the composite surfaces. These studies indicate the possibilities of utilizing machine learning and deep learning techniques to achieve online detection system for identification of defects in the composites. The applicability of machine learning (ML) classifiers, fitted and tested on different systems, is investigated by Yuxuan Chen et al. [ 11 ] for predicting the faults in air conditioners using sensor data. After selecting suitable features with feature selection techniques, they investigated the significant improvement in the accuracies of machine learning-based classifiers. In another work, Hossein Moosaei et al. [ 12 ] attempted to improve the classification accuracy by simultaneously implementing the feature selection and machine learning classification by the twin k-class support vector machine (SVM) algorithm. It was found that the method improved the classification accuracy significantly. An investigation conducted by Ayat Naji Hussain et al. [ 13 ] includes applying feature extraction techniques, feature selection techniques, and utilizing machine learning based classifiers to automatically classify the foot disorders to help in leg rehabilitation. It was obtained that the feature selection techniques help in modifying the classification accuracies of the machine learning models up to 99%. Wei Guan et al. [ 14 ] evaluated the distinctive patterns of change in the force signals as measured by force sensors for FSW joints, including metal joints that have defects and metal joints that do not have defects, using Fast Fourier transform to analyze the distortion in waveform in the traverse force helping in investigating the material flow and defect formation. Zhiwei Li et al. [ 15 ] introduced a technique for circular laser scanning by sensing many parameters in the 5-axis RFSW. The laser displacement sensor underwent rotational movement around the tool, generating a sequence of measured points to capture the three-dimensional form of the workpiece accurately. Using a laser displacement sensor decreases the required quantity of sensors and enables dependable process sensing. Sharda et al. [ 16 ], 2013, proposed a hybrid filter-type technique for selecting features that combines two methods, Relief-F and Pearson Correlation. The technique aims to reduce the complexity of the classifier and achieve high accuracy in classification algorithms. The result is a feature set with improved performance. The increasing popularity of IIoT tools and applications has led to a growing interest in online condition monitoring systems that offer remote access capabilities. Utilizing these tools to monitor production operations can result in improved performance and heightened efficiency. A group of researchers working on tool wear prediction utilized a tool wear dataset [ 17 ]. Various feature selection techniques, including the PCC, random forest (RF), PCC with RF and PCA with PCC, were employed. The same group of authors also utilized the dataset to estimate the Remaining Useful Life (RUL) of the cutting tool in milling [ 18 ]. The more the level of excellence in feature engineering conducted on the raw data, higher is the accuracy of the forecast. Adham E. Ragab et al. [ 19 ] presented the development of an online monitoring system utilizing Raspberry Pi and thermocouple and force sensors for the purpose of monitoring the peak temperature and axial force during spot welding using friction stir process (FSSW). They examined influence of tool rotation rpm, plunge rate, and time duration on maximum temperatures and axial forces during spot welding by FSW. The study by Santwana Gudadhe et al. [ 20 ] involved the classification of cerebral CT images by hemorrhage using various machine learning classifiers. A method is created by combining the features based on texture and transform to develop a combined feature-based approach. The proposed collaborative feature extraction technique enhances the accuracy of CT image classification compared to the features extraction based on texture and transform respectively. The transform-based feature approach utilizes the energy coefficients of CT images. A review work by Balachandar K et al. [ 21 ] focuses on the present level of comprehension and advancement of the input factors during the friction based joining and processing techniques. They focused that to determine the most likely causes of failure, many techniques are employed to gather data, such as vibration monitoring. Thermal imaging refers to the process of using specialized equipment to capture and display the infrared radiation emitted by objects, allowing for the visualization of heat patterns and temperature variations. Some researchers examined the additive manufacturing of multilayered Al6061 alloy using the friction stir powder additive manufacturing [ 22 ]. The goal was to modify physical and grain structural characteristics of the alloy by reducing the temperature variation in the direction of process. This was achieved by maintaining the substrate in close proximity to its designated temperature for artificial aging by utilizing an external heat source within a closed-loop system. In this paper, the machine learning approach is utilized for classification of surface defects occurring in composites manufactured by FSP. The multi-senosor data from vibration (tool, matrix), current (tool) and force sensors (matrix) mounted on the milling machine, is collected at a sampling rate of 1000 Hz with the help of a data acquisition system during the fabrication of surface composite samples by FSP. The raw sensor collected is subjected to preprocessing, i.e., data cleaning, eliminating noise, outliers, etc. The relevant features of the preprocessed data are extracted to characterize and reduce the dimension of data. Also, the ten most important features are selected using each of the feature selection techniques used. Figure 1 shows the proposed framework for classifying faults in the composite samples prepared by FSP. The paper is categorized into five sections. Section 1 provides an introduction to the framework of the defect classification of composites with related literature. Section 2 represents the material and detailed methodology, viz., sample preparation, FSP experimentation, and, Physical identification of surface defects in fabricated composites utilized for the preparation and analysis of composites. Section 3 represents the sensor data analysis including data acquisition and preprocessing; extraction of relevant features, selection of important features; fault classification models; and, evaluation of models. Section 4 represents the discussion of results, including; extracted features, selected features, performance evaluation using classification models, and comparative result analysis. Section 5 throws light on an important conclusion regarding the selection of optimal methods for the effective classification of faults in surface composites. 2 Equipment and Methodology 2.1 Experimentation Detail In the study, the surface composites of Al6061 alloy matrix with copper and graphene reinforcements are prepared by FSP on the CNC milling machine. FSP parameters selected are: Tool rpm = 1000 rpm, Tool feed = 45 mm/min, depth of cut = pin length, i.e., 5 mm [ 23 ]. The experimentation involves advancing FSP tool into the Al6061 alloy matrix having holes filled with reinforcement with traverse direction motion. The rubbing action between the tool shoulder and the matrix filled with reinforcement causes friction and the heat generation which produces surface composite with uniform distribution of the particles [ 24 ]. Various sensors pre-installed on the milling machine, are utilized, during FSP experimentation, i.e., composite fabrication, for analyzing specific parameters. These include two vibration sensors (spindle and table), one current sensor (spindle), and a dynamometer (table) with loads in X, Y, Z axes. The FSP setup for multi-sensor data collection is shown in Fig. 2 . 2.2 Signal Acquisition The composites prepared by FSP are likely to be subjected to various surface defects due to fabrication or machining errors. These defects are identified with the help of signals collected through various sensors, such as vibration sensor, current sensor, and dynamometer. Vibration sensors are sensitive to surface condition and any irregularities on surface may alter their amplitude resulting in detection of faults. The current sensor mounted on the machine spindle, determines the variation in the amount of current drawn by spindle providing insights into the stability of the operation and and indicate inconsistenies occurred due to defects. A table-based dynamometer, is placed between workpiece and machine bed, for collecting the signals for load in X, Y and Z directions to obtain slight variation in the axial and transverse loads [ 25 ] on the matrix plate during FSP which indicate presence of defects. When a defect occurs as the tool is passed over the matrix surface with reinforcement filled into the holes, the magnitudes of the cutting forces suddenly increase. Vibration levels might vary significantly during FSP. The vibration mainly affects the holding devices for tool or workpieces in multiple directions [ 26 ]. A current sensor is installed on the machine spindle and determines the current drawn by the tool during process. Also, the vibrations caused during the process are investigated in this study using accelerometers. Two vibration sensors, one on tool spindle, and another on the machine bed where matrix plate is fixed, are utilized for measuring vibration signals on the tool and on the composite being prepared respectively. The signal acquisition equipment are mentioned in Table 1 . A signal acquisition system by National Instruments, (NI USB-6210) has been used to collect the data for further processing. The data acquisition module acts as a medium that helps to digitize the analog signals collected from all the sensors so that they can be synchronized and analyzed simultaneously. One data file consists of the multisensory data during one pass (lap) of FSP tool travel on the matrix filled with reinforcement to form surface composites. Table 1 Details of Signal Collection equipment and sampling frequency S. No. Signal Mounting location Signal Collection Equipment 1 Vibration (Tool) Tool spindle VBR1/D0-3V 2 Vibration (Matrix) Machine bed VBR1/D0-3V 3 Current (Tool) Tool spindle Pico TA018 4 Dynamometer (Load cell X) Machine bed XEEPL FHC 031 5 Dynamometer (Load cell Y) Machine bed XEEPL FHC 031 6 Dynamometer (Load cell Z) Machine bed XEEPL FHC 031 2.3 Sensor data processing The sensor data collected is preprocessed, i.e., cleaned and organized, to structure data and remove unnecessary data. This includes the data being transformed or encoded into a more understandable form so that the computer can easily comprehend the multi-sensor signals as recorded. The data labelling includes defining the target variables for each sensor data. According to the visual inspection of the composites prepared, the data is categorized in terms of their defect condition. The reason for the occurrence of these surface defects is the manual sample preparation and experimentation errors. The data is split into features (X) and labels (y). The features refer to the multi-sensor data. The five different labels (categories) as the target variable (y) are selected as normal composite, pin-break composite, rough surface composite, no composite, and brazing-break composite. Figure 3 shows the raw sensor data signals for vibration (table) sensor for all five labels of composites. Normal composite refers to those samples where moderate levels of vibration, current and cutting forces with no significant changes in sensor readings are observed. Pin break type composites are those in which, after the FSP tool traverses, the pin is broken from the shoulder and is stuck into the matrix surface. It is caused by excessively high pin temperature due to the lack of a sufficient amount of release of material chips during the tool traverse. The break point refers to the sudden increase in current drawn by motor due to excessive load and a moderate rise in vibration amplitudes during the end of tool travel. In the rough surface type composites, an uneven surface composite layer is formed which includes voids and clusters of particles. The reason is the uneven surface contact between the tool and matrix surface during FSP due to excessive traverse speed resulting in lack of axial pressure on the matrix surface and improper recrystallization. High cutting forces occur throughout the tool traverse in the rough surface composites. a moderate rise in vibration amplitudes is observed during the end of tool travel in pin break and brazing break composites. No composite type are those in which the tool is not properly in contact with the matrix surface and the holes filled with reinforcement during FSP. This results in no or negligible dispersion embedding the reinforcement into the matrix and hence no manufacturing of composite occurs. This type of composite fabrication shows a significant drop in vibration, current and force signals. The brazing break type composites refer to the composites during fabrication of whom, the brazed tool (FSP tool is a steel body brazed with tungsten carbide shoulder and pin) undergoes failure as the rise in temperature in the tool body reaches the brazing temperature (450֯C approx.) during the tool traverse. There is a gradual rise and then a sudden drop in the vibration signals during the brazing failure. However, the current and load in z direction suddenly rise and drop at the breaking point. Figure 4 shows the detailed methodology adopted for fault classification of surface composites prepared by FSP. Also, the detailed explanation on the steps followed in the methodology, such as sensor data processing, feature extraction, feature selection, and classification of faults in the methodology are discussed in the current and sections further. 2.4 Feature Extraction The important characteristics of sensor data which reflect the distinct and significant characteristics of the process, are extracted for characterizing various aspects of data. The sensor data is acquired at each millisecond as the FSP tool is traversed on the matrix surface. Several significant characteristics of signals in the temporal and spectral domain are retrieved at every second from the preprocessed dataset for reducing the dimension of data and analyzing it effectively. The features are retrieved using standard python libraries including NumPy, SciPy, Pandas, scikit-learn, etc. The time domain and frequency domain features represent the variation in signals with time and frequency bands respectively. 2.4.1 Time domain features Time based or temporal characteristics are displayed with their functions in Table 4 . These features obtain the time-based characteristics of the signal such as its trend, variability, dynamics, etc. These features study the amplitude statistics, such as mean, median, variance, etc., which provides insights into the central tendency, spread and distribution, etc. The signal shape, magnitude, variation, frequency, stability, etc., are studied by these features. Table 4: Methods utilized in various types of feature selection techniques S.No. Time-domain Features Function 1 ECDF Empirical cumulative distribution function concerning time 2 Interquartile range Difference between the upper and lower quartile values of a time series signal. 3 Kurtosis Measurement of tailedness of a signal. 4 Maximum Calculation of highest amplitude of a signal. 5 Mean Calculates the average amplitude of signal. 6 Mean absolute deviation Measure of the variability of signal. 7 Median Middle value or mean of two middle values in a signal. 8 median absolute deviation Average distance of data from the median of the signal. 9 Minimum Determines the smallest utility in signal. 10 RMS value Root mean square of the signal. 11 Skewness Amount of deviation of the data from the sample mean, indicating the extent of asymmetry. 12 Standard deviation Deviation from the mean of a time series signal. 13 Variance Calculates Signal variance. 14 absolute energy Determines the true energy of a signal. 15 Distance Computes signal travelled distance. 16 Entropy Calculate the Shannon entropy of the signal. 2.4.2 Spectral-domain features The frequency domain or spectral features provide insight into the spectral characteristics of a signal. They reveal the energy distribution across various components related to frequency in the signal. These features obtain the significant frequency magnitude, dispersion, shape, variation, etc. These are displayed with their functions in Table 5 . Table 5: Methods used in various machine learning and ensemble learning models S.No. Frequency-domain features Function 1 FFT Mean coefficient Calculates the average amount of every frequency in the spectrogram. 2 Fundamental frequency The signal's lowest periodic waveform frequency. 4 LPCC Cepstral coefficients for linear prediction. 5 MFCC Calculates the Mel-frequency cepstral coefficients. 6 Maximum power spectrum Highest signal's spectral density and power. 7 Maximum frequency Highest signal frequency 8 Median frequency Calculates the centre point of the power distribution. 9 Power_bandwidth Bandwidth of the signal's power spectrum density. 10 Spectral centroid Represents Barycenter of the spectrum. 11 Spectral_decrease Quantifies reduction in amplitude of the spectra. 12 Spectral distance Computes the signal spectral gap. 13 Spectral entropy Calculates signal's spectral entropy using the Fourier transform. 14 Spectral kurtosis Measures dispersion relative to the average value. 15 Spectral positive turning Calculates count of positive inflection points in signal by FFT 16 Spectral roll-off Calculates frequency below to signal's total power is contained. 17 Spectral roll-on Computes the signal spectral roll-on 18 Spectral_skewness Quantifies the lack of symmetry relative to its average. 19 Spectral slope It calculates the gradient of frequency data. 20 Spectral spread Quantifies the extent to which the values in the spectrum deviate from their average value. 21 Spectral variation Calculates the degree of variability in the spectrum over time. 22 Wavelet absolute mean Calculates the wavelet scales' absolute mean value using the Continuous Wavelet Transform (CWT). 23 Wavelet energy Calculates the CWT energy for each scale of the wavelet. 24 Wavelet entropy Continuous Wavelet Transform (CWT) entropy of the signal. 25 Wavelet standard Calculates standard value of the CWT for each scale of wavelet. 26 Wavelet variance Calculates variance of each wavelet scale using CWT Table 6: Methods utilized in various types of feature selection techniques S.No. Feature Selection Type Methods used 1 Filter Methods Chi square method F_classif method Information gain Correlation Mean absolute difference Variance threshold 2 Wrapper Methods Forward feature selection (FFS) Recursive feature elimination (RFE) 3 Embedded Methods Lasso Regularization Random Forest Importance 2.5 Feature selection The technique for selection of features is the act of choosing very important relevant features among the extracted features which promotes to develop simpler models for further analysis. The reduced dimension of data helps to improve the accuracies of the machine learning classifiers with a shorter training time and reduced overfitting. In this study, three different types of feature selection techniques are utilized: Filter type, wrapper type and embedded type methods. 2.5.1 Filter methods These methods select the variables or features based on the attributes of the data and are independent of the machine learning approach. In these methods, the features are ranked based on their relevance and highest rank features are chosen to induce the classification models. The filter methods utilized in the study are: Chi square method, F_classif method, information gain, correlation, dispersion ratio, mean absolute difference and variance threshold methods. Among filter methods, Chi square method for selecting important characteristics is utilized to perform the chi square (χ2) test which measures the autonomy among two parameters and identifies most informative ten characteristics according to their importance with respect to the target parameter. The F_Classif method for feature selection is used for calculating the ANOVA F-statistic values for all the features extracted and selecting the top ten features based on those values. The information gain method is utilized to obtain mutual information between the features and their target variables by mutual_info_classif function. The features with ten highest scores obtained are selected for further analysis. The correlation method quantifies the linear interconnection of the features with target variable. The correlation coefficients are calculated and most significant ten features are selected. The Mean Absolute Difference (MAD) is used to calculate the mean absolute deviation or variability of each feature from the feature's mean value and features with ten highest variabilities are selected. Variance Threshold is used to eliminate features with low variance. In all of the filter methods utilized, the features are ranked based on the respective evaluation criteria (different for all methods) by eliminating the features having constant or quasi constant values. The filter methods provide simple and powerful methods to eliminate irrelevant, duplicated, correlated, and hence the redundant features in an easier and quicker way. 2.5.2 Wrapper methods These use the machine learning models for evaluating the quality for selecting the subset of features. It involves the training of new model on each feature subset and determines best performing feature subset based on machine learning algorithm. They are also capable to detect the interactions between the variables, so provide better prediction performance than that by filter methods. The wrapper methods used in this study are: forward feature selection (FFS) and recursive feature elimination (RFE) techniques. FFS begins with an empty model, adds features one by one, and evaluates the model, until the accuracy is increased. On the contrary, RFE begins with a model with all features, evaluates the model, eliminates features one by one based on least importance, until desired number of features is reached. 2.5.3 Embedded techniques The embedded techniques perform the feature selection during the training of data in machine learning classification algorithm. These methods are comparatively faster than the wrapper methods and achieve higher accuracies in comparison to filter methods with lower risk of overfitting of data. Embedded methods utilized in the study are: Lasso regularization and Random Forest Importance methods. Lasso regularization method can shrink some of the coefficients to zero so that some features can be eliminated. Random Forest Importance method involves building of random forest trees, calculating feature importance, and eliminating the unimportant features until the condition is achieved. Table 6 shows methods used for various the feature selection types discussed. 2.6 Fault Classification Feature classification refers to the categorization of features using machine learning. The selected features of the sensor data are classified using different machine learning classification algorithms (classifiers). Figure 5 shows the workflow of the classification of defects in FSP prepared composites using machine learning approach [ 27 ]. The multi-sensor data selected using feature selection techniques, which is in the form of features and their labels, divides itself into a training and a testing dataset. The machine learning classifier is fit with the labelled training data and its efficacy is tested with the unlabelled test dataset by comparing the predicted labels with the true labels of the test dataset. Also, there is scope for adjustment of different parameters of the ML model using the hyperparameter tuning [ 28 ]. Eight classifiers are utilized to classify the features selected by each of the selection methods as discussed individually. The Different machine learning (LR, NB, kNN, SVM, DT) [ 29 ], [ 30 ] and ensemble learning classifiers (RF, AB, GB) [ 31 ]–[ 33 ] categorize the selected feature data on dataset for training and validate it on the test dataset. Various feature selection methods and machine learning classifiers are utilized for identifying and categorizing the most important features and hence classifying the defects in fabricated composites. i. k-Nearest Neighbors method (kNN) The k-Nearest Neighbors is a non-parametric method which clusters the data and assigns a new label to the data on the basis of the closeness of the values according to the Euclidean distance between the data points to make classification of data. It calculates the distance to the training data using the selected distance metric.The hyperparameter tuning is used for adjustment of parameters, i.e., the number of neighbours and window size. kNN is a non-parametric method capable of handling multisensory data. ii. Support vector machine (SVM) Support Vector Machine method applied for multi-class classification, is capable to handle high dimensional data exhibiting non-linear complex relationships among data. Two approaches are mainly used: one versus one where all the pairs of classes are considered for training the model and one versus all where n number of output classes and n number of classification models are trained concurrently while accounting for the separation between the real class and the remaining classes. iii. Gradient boosting classifier (GB) Gradient boosting is an ensemble method that constructs models in a sequential manner. Every subsequent model is educated to rectify the mistakes committed by its previous models. The algorithm consists of initializing the model with a base estimator and then adding more models iteratively by computing errors. Then the model is trained to predict errors and the ensemble is updated with new model for classification and evaluation of the model. iv. Decision tree classifier (DT) The decision tree classifier is highly useful for classification of time series sensor data since it is capable of handling non-linear relationships of data and feature interactions. It is a non-parametric type of machine learning technique that is easier to interpret, robust to noise, and provide a measure of feature importance. However, it may promote overfitting of data during training and data instability, i.e., small variation in data may lead to different splits. These limitations can be overcome by using ensemble methods such as, random forest classifiers, etc. v. Random forest Classifier (RF) Random Forest is a type of ensemble method for classification that improves the accuracy of predictions by combining multiple decision trees trained on various groups of the data and using averaging. The random forest method combines predictions from many decision trees and generates the final output based on the majority vote among these predictions. It functions in two separate stages. The initial stage is creating a random forest by combining N decision trees. The second stage involves creating predictions for each independent tree that was constructed in the beginning stage. Random forest classifier is capable to effectively manage extensive datasets with a high number of dimensions. Also, it improves the precision of the model and mitigates the risk of overfitting the data. vi. Ada boost classifier (AB) The AdaBoost or Adaptive Boosting, is a magnifying approach employed as an ensembling technique among different machine learning approaches. The technique includes reassigning weightage to each instance, with bigger weights given to examples that were categorized erroneously. While the data is trained, it produces a variable number of decision trees. When first model is built, the improperly categorized data in the first model gets a priority. This imposes these data records to be provided as intake to the second model and so on. This method continues til it is designated by a specific quantity of base learners needed to generate. vii. Naïve bayes (NB) It is a widely used machine learning method employed for tasks involving classification, such as text classification. This algorithm is classified as a generative learning algorithm, as it models the input distribution for a certain class or category. This approach relies on the premise that the characteristics of the input data are statistically unrelated for the class, enabling the algorithm to produce fast and accurate predictions. viii. Logistic Regression (LR) It is a statistical technique and a machine learning method utilized for classification tasks, relying on the principle of probability. This method is employed when the dependent variable, also known as the target variable, is categorical in nature. Logistic Regression is a commonly employed method because to its great efficiency and little requirement for computer resources. It exhibits enhanced efficiency by eliminating variables that possess negligible or no correlation with the output variable. Table 7 displays the methods used for various types of classification models used. Table 7: Selected list of features by wrapper methods used with accuracy scores S.No. Classification Model Methods Used 1 Machine Learning Models LR NB kNN SVM DT 2 Ensemble Learning Models RF AB GB 2.7 Evaluation of Models The machine learning classifiers are utilized to categorize the training set of data and then classification is employed on the test data. The classification of the data as conducted by the model algorithm, is afterwards, validated from the true values with the help of evaluation of models [34] which suggests how the model has performed in terms of predicting classes. There are several measures which determine the model performance: i) Accuracy : This evaluates the model's accuracy by calculating the accuracy as the percentage of correctly anticipated instances to the total number of occurrences. Also, it can be determined as the rate of accurately recognized affirmative instances and true negatives to the overall positive and negative values. It is expressed as: $$\:Accuracy=\:\frac{Number\:of\:correct\:instances}{Total\:number\:of\:occurences}$$ …………………...(1) ii) Precision : It quantifies the precision of optimistic forecasts. The term "precision" refers to the precise ratio of accuracy predicted the ratio of positive cases to the overall cases or instances predicted as positive. It helps in determining the frequency at which a positive test accurately identifies a true positive. The term "positive predictive value" refers to the proportion of real positive results compared to false positive results. It is expressed as: $$\:Precision=\frac{True\:Positives\:}{True\:Positives+False\:Positives}$$ …………………….(2) iii) Recall : It is also called as sensitivity or the real positive rate. It evaluates the model's capacity to accurately identify and include all events that are classified as positive. The term "precision" refers to the ratio of correctly predicted positive instances to the total confirmed positive instances. It evaluates the test's ability to yield a positive result when the condition is present. It can be expressed as: $$\:Recall=\frac{True\:Positives\:}{True\:Positives+False\:Negatives}$$ …………………………….(3) iv) F1 score : It integrates the measures of precision and recall into a unified metric. It is especially beneficial in cases where there is an imbalanced distribution of classes. Also, it is said to be the harmonic mean of the recall and accuracy. It can be expressed as: $$\:F1=2*\frac{precision*recall}{precision+recall}$$ ……………………….(4) 3 Results and Discussion The sensor data recorded during the manufacturing of surface composites by FSP, includes six incoming signals, i.e., three cutting forces (X, Y and Z directions on matrix), one current (tool spindle) and two vibration sensors (tool spindle and matrix). The multi sensor signals are digitized for synchronized analysis with the help of the data acquisition system module. The raw sensor data is preprocessed and labelled with their defect condition as target variables. Then important features are extracted, and most important fewer features are selected for further analysis. The fault identification and classification of the surface defects in the composites is performed using different machine learning classifiers with their performance evaluation. The classification algorithms utilized include kNN, SVM, GB, decision tree, and random forest, ada boost, naive bayes, and logistic regression. The performance of all the classifiers used are compared using different evaluation metrics. The python programming in Jupiter notebook was used for extracting, selecting and classifying the sensor data. 3.1 Feature Extraction The features are extracted from the raw sensor data (six sensors) that can be processed while preserving the information in the original dataset using standard libraries in python. Different time (temporal) and feature (spectral) domain features are extracted at a sampling frequency of 1000 Hz and a window size of 1000 samples is selected for the extraction. Time domain features are extracted by ‘NumPy’ (numerical python) and ‘pandas’ libraries in python. Spectral domain features are extracted using ‘PyWavelets’, ‘scikit-learn’, ‘joblib’, ‘SciPy’ (Scientific Python) libraries, etc. Time domain features extracted in the present research are 16 for each raw sensor data. The frequency domain features extracted are 26 for each sensor data. Overall, 252 features are extracted which includes 96 time domain and 156 sprectral features for all six sensor data. All the extracted features, alongwith the label (condition of defect) are stored in the form of csv files where the features are considered as the independent parameter and label is taken as the target parameter. 3.2 Feature Selection The most important ten features relevant for defect classification are selected out of the extracted features dataset using the features selection techniques including various filter methods, viz., Chi Square method, F_classif method, information gain method, correlation method, dispersion ratio, mean absolute difference, and, variance threshold method. Also, various wrapper methods, such as, forward feature selection and recursive feature elimination methods, and embedded methods, such as lasso regularization, and random forest importance method are utilized for feature selection. 3.2.1 Feature Selection using Filter Methods Among all the filter methods used for feature selection, i.e., Chi square, F_classif, information gain, correlation, dispersion ratio, mean absolute difference and variance threshold method, It is obtained that among these, the information gain filter method achieves high performance accuracies with the machine learning classification models. This approach quantifies the quantity of data acquired about one variable through another variable. The information gain is well-suited for capturing complex dependency of features on the target variable, rendering it a good option for selecting the sensor data exhibiting linear as well as non-linear relationships. From the extracted time domain and spectral domain features, ten most important features are selected by calculating the mutual information (MI) of features with respect to the target variables. This calculated value refers to the reduction in uncertainty of target variable, given the feature. Then, the features are ranked according to their mutual information scores as calculated and the features with top ten MI scores are selected for further classification. Table 8 displays the record of ten most significant characteristics selected by each of the six filter methods used. Table 8 Selected ten features by each of the Filter methods S. No. Chi Square Method F_Classif Information gain Correlation Mean Absolute Difference Variance Threshold 1 1_ECDF Percentile_0 2_ECDF Percentile_1 1_Median 3_LPCC_1 2_Spectral distance 3_LPCC_1 2 1_ECDF Percentile_1 2_Mean 1_Root mean square 3_LPCC_11 4_Spectral distance 3_LPCC_10 3 1_Max 2_Median 1_ECDF Percentile_0 3_MFCC_0 3_Spectral distance 3_LPCC_11 4 1_Mean 2_Root mean square 1_Mean 3_Spectral slope 5_Spectral distance 3_LPCC_2 5 1_Median 2_Wavelet absolute mean_1 1_ECDF Percentile_1 3_Spectral centroid 1_Spectral distance 3_MFCC_0 6 1_Root mean square 2_Wavelet absolute mean_4 1_Min 3_MFCC_9 0_Spectral distance 3_MFCC_1 7 3_Histogram_5 2_Wavelet absolute mean_5 1_Wavelet absolute mean_8 3_LPCC_2 3_Power bandwidth 3_MFCC_10 8 1_Wavelet energy_7 2_Wavelet absolute mean_6 1_Wavelet absolute mean_7 3_LPCC_10 4_Histogram_5 3_MFCC_11 9 1_Wavelet standard deviation_7 2_Wavelet absolute mean_7 1_Wavelet absolute mean_6 3_MFCC_1 4_Power bandwidth 3_MFCC_9 10 1_Wavelet variance_7 2_Wavelet absolute mean_8 1_Wavelet absolute mean_5 3_MFCC_11 4_Histogram_4 3_Spectral centroid 3.2.2 Feature Selection using Wrapper Methods Among the two wrapper methods used, i.e., forward feature selection (FFS) and recursive feature elimination (RFE) for selecting the significant features for classification of sensor data, it is obtained that the features selected by FFS are more efficiently classified by machine learning. In FFS, one by one the features, which improve the model performance the most (highest accuracy scores), are added until ten most significant features are achieved. It is advantageous for the large set of data, is less computationally expensive, promotes incremental improvement, and prevents multicollinearity which is common issue in sensor data. Table 9 shows the ten selected features by both of the wrapper type methods with individual acuuracy scores. Table 9 Selected list of features by wrapper methods used with accuracy scores S.No. Forward Feature Selection Recursive Feature Elimination Feature Accuracy Score Feature Accuracy Score 1 1_ECDF Percentile 0.996 2_ECDF Percentile_1 1067.3 2 1_Mean 0.996 2_Mean 1107.4 3 4_ECDF Percentile_1 0.996 2_Median 1081.9 4 4_ECDF_0 0.996 2_Root mean square 1115.9 5 1_FFT mean coefficient_69 0.996 2_Wavelet absolute mean_1 1037.6 6 3_Wavelet variance_0 0.996 2_Wavelet absolute mean_4 1055.6 7 4_Wavelet absolute mean_0 0.996 2_Wavelet absolute mean_5 1077.3 8 4_Wavelet energy_4 0.996 2_Wavelet absolute mean_6 1091.1 9 5_MFCC_0 0.996 2_Wavelet absolute mean_7 1099.3 10 5_Spectral centroid 0.996 2_Wavelet absolute mean_8 1103.7 3.2.3 Feature Selection using Embedded Methods In this study, among embedded methods for selecting important characteristics, Lasso regularization and random forest importance methods are utilized to select ten very important characteristics of the dataset. It is obtained that the Random Forest Importance method is capable of handling large amount of data having non-linear relationships, prevents overfitting, and creates ranking of feature importances. Hence the features selected by this method achieve high accuracy with the machine learning classifiers. Table 10 shows the list of top ten features selected by both of the embedded methods alongwith the rank of importance for further classification. Table 10 Selected list of features by embedded methods used with accuracy scores S.No. LASSO Regularization method Random Forest Feature Selection Feature Importance Feature Importance 1 3_Spectral decrease 0.811674 3_Wavelet energy_0 0.013291 2 2_Histogram_6 0.615873 3_Wavelet standard deviation_0 0.011518 3 0_Spectral kurtosis 0.582926 1_Root mean square 0.011256 4 1_Spectral kurtosis 0.535394 1_ECDF Percentile_0 0.010952 5 2_Histogram_9 0.487217 1_Mean 0.010774 6 5_Spectral centroid 0.392809 1_Median 0.009951 7 5_Spectral decrease 0.374709 3_Wavelet variance_0 0.009383 8 1_Fundamental frequency 0.259443 1_ECDF Percentile_1 0.008665 9 0_Maximum frequency 0.158904 1_Wavelet absolute mean_4 0.008509 10 1_MFCC_4 0.079898 1_Wavelet absolute mean_8 0.008349 3.3 Performance Evaluation using Classification Models The selected features of multi-sensor data with all the techniques utilized for selecting features (filter, wrapper and embedded type methods), are provided to different machine learning and ensemble learning classifiers, i.e., KNN, SVM, GB, decision tree, random forest, Ada Boost, naive bayes, and logistic regression. For classification, the selected data, i.e., features (X) and labels (y) are divided into training (70%) and testing (30%) datasets. The labelled training dataset (X 1 , y 1 ) data is provided towards the machine learning models for fitting or training purpose. However, the unlabelled testing set (X 2 ) is provided for the performance evaluation, i.e., for validating the predicted label values with the true labels (y 2 ). 3.3.1 Model Performance by Filter methods of Feature Selection The evaluation of the accuracies of classifiers for the selected features by all the filter methods used suggest that the highest performance is obtained with the features selected by Information gain method of feature selection. The metrics for performance of the features selected by this method are shown in Table 11 . It is obtained that among all the classifiers used, the Gradient Boosting classifier achieves highest accuracy i.e., 0.99 for classification of the data. Figure 6 (a) shows the bar plot for comparing the accuracies of all the machine learning based classifiers used. Figure 6 (b) shows that ROC curve for classification model of the selected features is nearly perfect in distinguishing between positive and negative classes with no false positives or false negatives and AUC = 1. The reason for the excellent performance of GB classifier is that it builds the model sequentially by fitting new models to the residual errors made by previous models by improving mistakes of the other models, leading to high efficiency. Also, being ensemble of multiple trees, it captures non-linear relationships, can model complex patterns, handles overfitting, and involves feature interactions. Table 11 Performance of the classifiers of features selected by Information gain FS method S.No. Machine learning Classifier Accuracy Precision Recall F1 score 1 KNN 0.96 0.96 0.95 0.95 2 SVM 0.98 0.98 0.98 0.98 3 Gradient Boosting Classifier 0.99 0.99 0.99 0.99 4 Decision Tree 0.94 0.94 0.94 0.94 5 Random Forest 0.98 0.98 0.98 0.98 6 Ada Boost Classifier 0.85 0.87 0.84 0.84 7 Naive Bayes 0.91 0.91 0.91 0.91 8 Logistic Regression 0.98 0.98 0.98 0.98 3.3.2 Model Performance by Wrapper methods of Feature Selection The classification models are examined for their efficacies to classify the features selected by both the wrapper methods used, i.e., FFS and RFE, suggests that the highest performance is obtained with the variables chosen by FFS method. The metrics for the performance of the classifiers are shown in Table 12 . It is obtained that Random Forest classifier achieves maximum accuracy i.e., 0.99 for classification of the data. Figure 7 (a) shows the bar plot for comparison of accuracies by all the machine learning and ensemble learning classifiers used. Figure 7 (b) displays the ROC curve for classification model of the selected features which is good in distinguishing between positive and negative classes with no false positives or false negatives and AUC is 0.98 which is close to 1. The reason for the excellent performance of Random Forest classifier is that being an ensemble learning technique, it produces multiple decision trees and integrates the results obtained by all of them, perfroms averaging reducing the variance of the model leading to enhanced and precise forecasts in comparison to a single decision tree. Also, it employs bootstrap aggregating or bagging which includes training of every tree from the random data subset, while reducing overfitting and enhancing generalization. Table 12 Performance of the classifiers of features selected by FFS method S.No. Machine learning Classifier Accuracy Precision Recall F1 score 1 KNN 0.97 0.97 0.97 0.97 2 SVM 0.98 0.98 0.98 0.98 3 Gradient Boosting Classifier 0.98 0.98 0.98 0.98 4 Decision Tree 0.91 0.92 0.91 0.91 5 Random Forest 0.99 0.99 0.99 0.99 6 Ada Boost Classifier 0.67 0.70 0.66 0.64 7 Naive Bayes 0.92 0.92 0.92 0.92 8 Logistic Regression 0.88 0.89 0.86 0.86 3.3.3 Model Performance by Embedded methods of Feature Selection The multiple machine learning and ensemble learning classifiers are investigated of performance of the classification algorithms on the data features selected by both the embedded methods used suggest that the highest performance is obtained with the features selected by Random Forest Importance method of feature selection. The metrics to evaluate performance of the classifiers are shown in Table 13 . It is obtained that the classifiers used, the Random Forest classifier achieves maximum accuracy i.e., 0.96 for classification of the data. Figure 8 (a) shows the bar plot for comparison of accuracies of all the machine learning classifiers used. Figure 8 (b) shows that ROC curve for classification model of the selected features is good in distinguishing between positive and negative classes and AUC is 0.95 which is close to 1. The reason for the better performance of Random Forest classifier is that it is an ensembles different decision trees simultaneously and can model complex non-linear relationships and features interactions without requiring explicit specification. Random forest classification is especially useful for time series data, as acquired in the present study, where relationships can be highly non-linear and intricate. Table 13 Performance of the classifiers of features selected by RFI method S.No. Machine learning Classifier Accuracy Precision Recall F1 score 1 KNN 0.81 0.81 0.81 0.81 2 SVM 0.83 0.84 0.82 0.82 3 Gradient Boosting Classifier 0.93 0.93 0.93 0.93 4 Decision Tree 0.94 0.95 0.93 0.95 5 Random Forest 0.96 0.96 0.96 0.96 6 Ada Boost Classifier 0.43 0.43 0.43 0.42 7 Naive Bayes 0.65 0.75 0.65 0.62 8 Logistic Regression 0.80 0.78 0.78 0.75 3.4 Comparative Result Analysis This subsection discusses the comparison and analysis of efficiency of various machine learning classifiers utilized for the selected features with different feature selection techniques. Table 14 displays the comparative analysis of performance of several feature selection techniques used with different classifiers for the detecting and classifying the surface defects on the composites prepared by FSP. This analysis verifies that the chi square method for feature selection method performs better among other filter methods. Similarly, the forward feature selection method performs well among wrapper methods, and, random forest feature selection performs well among embedded feature selection methods. It is obtained that the ensemble type machine learning classifiers perfrom better than other classifiers. Table 14 Comparison of Model Accuracies of several feature selection methods for all classifiers. S. No. Feature Selection / Classifier KNN SVM GB DT RF AB NB LR 1 Chi-Square 0.88 0.68 0.87 0.90 0.90 0.50 0.74 0.50 2 F_classif 0.53 0.44 0.53 0.52 0.54 0.39 0.46 0.32 3 Information Gain 0.96 0.98 0.99 0.94 0.98 0.85 0.91 0.99 4 Correlation 0.57 0.58 0.64 0.65 0.66 0.50 0.52 0.55 5 Mean Absolute Difference 0.98 0.95 0.98 0.94 0.98 0.75 0.85 0.88 6 Variance Threshold 0.56 0.59 0.66 0.65 0.66 0.50 0.53 0.54 7 Forward Feature Selection 0.99 0.98 0.98 0.91 0.99 0.67 0.92 0.88 8 Recursive Feature elimination 0.99 0.98 0.98 0.91 0.99 0.67 0.92 0.88 9 LASSO regularization 0.92 0.93 0.93 0.89 0.94 0.43 0.81 0.90 10 Random Forest importance 0.96 0.77 0.95 0.94 0.96 0.61 0.80 0.66 4 Conclusion Investigation, includes fabricating Al6061 alloy composites with copper and graphene reinforcement via FSP and analyzing multi-sensor data to identify and classify surface defects. During each FSP process, data from vibration sensors, current sensors, and dynamometers are collected and processed. Features are extracted to reduce data size and highlight significant characteristics. These features are labeled based on five defect conditions—normal composite, pin break, rough surface, no composite, and brazing break composite—determined by visual inspection. Various feature selection methods and machine learning classifiers are employed to determine the key characteristics that have the highest significance. and classify the defects. The study focuses on the significance of techniques for selecting important features in enhancing accuracies achieved by classification models and aims to identify the optimal techniques, providing recommendations for researchers working with the dataset. The results reveal optimal combinations of machine learning classifiers and feature selection methods for classifying defects in FSP-prepared composites. The highest accuracies were achieved with the following combinations of feature selection and machine learning techniques: 1. Among all the Filter Methods used, Information gain feature selection in combination with Gradient Boosting ensemble technique achieves maximum accuracy. 2. Among all the Wrapper Methods used, Forward Feature Selection in combination with Random Forest ensemble technique achieves maximum accuracy. 3. Among all embedded methods used, Random Forest importance in combination with Random Forest ensemble technique achieves maximum accuracy. Abbreviations FSP – Friction Stir Processing FFS - Forward feature selection RFE - Recursive feature elimination LR – Logostic Regression NB – Naive Bayes kNN – k-Nearest Neighbours SVM – Support Vector Machines DT – Decision Trees RF – Random Forest AB – Ada Boosting GB – Gradient Boosting Declarations 5 Authors’ declaration This manuscript is the authors’ original work and has not been published elsewhere. All authors have checked the manuscript and have agreed to this submission. 6 Author Contributions Pragya Saxena: Writing – original draft, Methodology, Review and editing; Arunkumar Bongale: Review, Formal Analysis, Conceptualization; Satish Kumar: Project administration, Formal Analysis Conceptualization, Results and discussions.; Rajesh Kodbal: Review and Formal Analysis. 7 Funding No funding sources were provided for publishing the manuscript. References G. K. Padhy, C. S. Wu, and S. Gao, “Friction stir based welding and processing technologies - processes, parameters, microstructures and applications: A review,” J. Mater. Sci. Technol. , vol. 34, no. 1, pp. 1–38, 2018, doi: 10.1016/j.jmst.2017.11.029. V. C. Kale, “Aluminium Based Metal Matrix Composites for Aerospace Application: A Literature Review,” IOSR J. Mech. Civ. Eng. , vol. 12, no. 6, pp. 2278–1684, 2015, doi: 10.9790/1684-12653136. A. Heidarzadeh et al. , Friction stir welding/processing of metals and alloys: A comprehensive review on microstructural evolution , vol. 117. Elsevier Ltd, 2021. doi: 10.1016/j.pmatsci.2020.100752. S. A. Alidokht, A. Abdollah-zadeh, S. Soleymani, and H. Assadi, “Microstructure and tribological performance of an aluminium alloy based hybrid composite produced by friction stir processing,” Mater. Des. , vol. 32, no. 5, pp. 2727–2733, 2011, doi: 10.1016/j.matdes.2011.01.021. M. M. Tash and E. R. I. Mahmoud, “Development of in-Situ Al-Si/CuAl2 metal matrix composites: Microstructure, hardness, and wear behavior,” Materials (Basel). , vol. 9, no. 6, 2016, doi: 10.3390/ma9060442. A. H. Elsheikh, “Applications of machine learning in friction stir welding: Prediction of joint properties, real-time control and tool failure diagnosis,” Eng. Appl. Artif. Intell. , vol. 121, no. June 2022, p. 105961, 2023, doi: 10.1016/j.engappai.2023.105961. M. Sajid, A. K. Malik, M. Tanveer, and P. N. Suganthan, “Neuro-Fuzzy Random Vector Functional Link Neural Network for Classification and Regression Problems,” IEEE Trans. Fuzzy Syst. , vol. 32, no. 5, pp. 2738–2749, 2024, doi: 10.1109/TFUZZ.2024.3359652. S. Thapliyal and A. Mishra, “Machine learning classification-based approach for mechanical properties of friction stir welding of copper,” Manuf. Lett. , vol. 29, pp. 52–55, 2021, doi: 10.1016/j.mfglet.2021.05.010. B. Das, S. Pal, and S. Bag, “Torque based defect detection and weld quality modelling in friction stir welding process,” J. Manuf. Process. , vol. 27, pp. 8–17, 2017, doi: 10.1016/j.jmapro.2017.03.012. S. Yoon et al. , “Defect detection in composites by deep learning using solitary waves,” Int. J. Mech. Sci. , vol. 239, no. October 2022, p. 107882, 2023, doi: 10.1016/j.ijmecsci.2022.107882. Y. Chen, A. Ebrahimifakhar, Y. Hu, and D. P. Yuill, “Generalizability of machine learning-based fault classification for residential air-conditioners,” Energy Build. , vol. 295, no. June, p. 113263, 2023, doi: 10.1016/j.enbuild.2023.113263. H. Moosaei and M. Hladík, “Sparse solution of least-squares twin multi-class support vector machine using ℓ0 and ℓp-norm for classification and feature selection,” Neural Networks , vol. 166, pp. 471–486, 2023, doi: 10.1016/j.neunet.2023.07.039. A. Naji Hussain, S. A. Abboud, B. A. baki Jumaa, and M. N. Abdullah, “Impact of feature reduction techniques on classification accuracy of machine learning techniques in leg rehabilitation,” Meas. Sensors , vol. 25, no. September 2022, p. 100544, 2023, doi: 10.1016/j.measen.2022.100544. W. Guan et al. , “Detection of tunnel defects in friction stir welded aluminum alloy joints based on the in-situ force signal,” J. Manuf. Process. , vol. 71, no. July, pp. 1–11, 2021, doi: 10.1016/j.jmapro.2021.09.014. Z. Li, H. Zhao, X. Zhang, J. Dong, L. Hu, and H. Gao, “Multi-parameter sensing of robotic friction stir welding based on laser circular scanning,” J. Manuf. Process. , vol. 89, no. 92, pp. 92–104, 2023, doi: 10.1016/j.jmapro.2023.01.060. S. Sharda et al. , “A hybrid machine learning technique for feature optimization in object-based classification of debris-covered glaciers,” Ain Shams Eng. J. , vol. 13, no. 6, p. 101809, 2022, doi: 10.1016/j.asej.2022.101809. S. Sayyad, S. Kumar, A. Bongale, K. Kotecha, G. Selvachandran, and P. N. Suganthan, “Tool wear prediction using long short-term memory variants and hybrid feature selection techniques,” Int. J. Adv. Manuf. Technol. , vol. 121, no. 9–10, pp. 6611–6633, 2022, doi: 10.1007/s00170-022-09784-y. S. Sayyad, S. Kumar, A. Bongale, K. Kotecha, and A. Abraham, “Remaining Useful-Life Prediction of the Milling Cutting Tool Using Time–Frequency-Based Features and Deep Learning Models,” Sensors , vol. 23, no. 12, 2023, doi: 10.3390/s23125659. A. E. Ragab et al. , “Open-source real-time monitoring system of temperature and force during friction stir spot welding,” J. Eng. Res. , no. July, 2023, doi: 10.1016/j.jer.2023.08.020. S. Gudadhe, A. Thakare, and A. M. Anter, “A novel machine learning-based feature extraction method for classifying intracranial hemorrhage computed tomography images,” Healthc. Anal. , vol. 3, no. April, p. 100196, 2023, doi: 10.1016/j.health.2023.100196. K. Balachandar, R. Jegadeeshwaran, J. Lakshmipathi, and D. Saravanakumar, “Friction Stir Welding Tool Condition Prediction Using Vibrational Analysis through Machine Learning - A Review,” J. Phys. Conf. Ser. , vol. 1969, no. 1, 2021, doi: 10.1088/1742-6596/1969/1/012051. B. Chaudhary, N. K. Jain, J. Murugesan, and V. Patel, “Friction stir powder additive manufacturing of Al 6061 alloy: Enhancing microstructure and mechanical properties by reducing thermal gradient,” J. Mater. Res. Technol. , vol. 26, pp. 1168–1184, 2023, doi: 10.1016/j.jmrt.2023.07.270. P. Saxena, A. Bongale, S. Kumar, and P. Jadhav, “Investigation of microstructural and wear behavior of Al6061 surface composites fabricated by friction stir process using Taguchi approach,” Mater. Res. Express , vol. 9, no. 1, 2022, doi: 10.1088/2053-1591/ac4a2d. P. Saxena, A. Bongale, S. Kumar, and P. Jadhav, “Microstructural and sensor data analysis of friction stir processing in fabricating Al6061 surface composites,” Eng. Res. Express , vol. 5, no. 1, p. 015065, 2023, doi: 10.1088/2631-8695/acc158. Y. Y. Lim and C. K. Soh, “Effect of varying axial load under fixed boundary condition on admittance signatures of electromechanical impedance technique,” J. Intell. Mater. Syst. Struct. , vol. 23, no. 7, pp. 815–826, 2012, doi: 10.1177/1045389X12437888. M. Shah, V. Vakharia, R. Chaudhari, J. Vora, D. Y. Pimenov, and K. Giasin, “Tool wear prediction in face milling of stainless steel using singular generative adversarial network and LSTM deep learning models,” Int. J. Adv. Manuf. Technol. , vol. 121, no. 1–2, pp. 723–736, 2022, doi: 10.1007/s00170-022-09356-0. M. Liu, H. Li, H. Zhou, H. Zhang, and G. Huang, “Development of machine learning methods for mechanical problems associated with fibre composite materials: A review,” Compos. Commun. , vol. 49, p. 101988, 2024, doi: https://doi.org/10.1016/j.coco.2024.101988. L. Yang and A. Shami, “On hyperparameter optimization of machine learning algorithms: Theory and practice,” Neurocomputing , vol. 415, pp. 295–316, 2020, doi: https://doi.org/10.1016/j.neucom.2020.07.061. S. R. Anvari, F. Karimzadeh, and M. H. Enayati, “Wear characteristics of Al-Cr-O surface nano-composite layer fabricated on Al6061 plate by friction stir processing,” Wear , vol. 304, no. 1–2, pp. 144–151, 2013, doi: 10.1016/j.wear.2013.03.014. O. Iparraguirre-Villanueva, K. Espinola-Linares, R. O. Flores Castañeda, and M. Cabanillas-Carbonell, “Application of Machine Learning Models for Early Detection and Accurate Classification of Type 2 Diabetes,” Diagnostics , vol. 13, no. 14, 2023, doi: 10.3390/diagnostics13142383. S. Verma, M. Gupta, and J. P. Misra, “Performance evaluation of friction stir welding using machine learning approaches,” MethodsX , vol. 5, pp. 1048–1058, 2018, doi: 10.1016/j.mex.2018.09.002. B. A. Tama and S. Lim, “Ensemble learning for intrusion detection systems: A systematic mapping study and cross-benchmark evaluation,” Comput. Sci. Rev. , vol. 39, p. 100357, 2021, doi: https://doi.org/10.1016/j.cosrev.2020.100357. A. Taha and O. Barukab, “Android Malware Classification Using Optimized Ensemble Learning Based on Genetic Algorithms,” Sustainability , vol. 14, no. 21, 2022, doi: 10.3390/su142114406. U. V Nikam and V. M. Deshmuh, “Performance Evaluation of Machine Learning Classifiers in Malware Detection,” in 2022 IEEE International Conference on Distributed Computing and Electrical Circuits and Electronics (ICDCECE) , 2022, pp. 1–5. doi: 10.1109/ICDCECE53908.2022.9793102. Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 21 Aug, 2024 Reviewers invited by journal 20 Aug, 2024 Editor assigned by journal 05 Aug, 2024 First submitted to journal 31 Jul, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4834721","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":342824298,"identity":"37511f37-745d-44b5-95c8-1289b4bc107c","order_by":0,"name":"Pragya Saxena","email":"","orcid":"","institution":"Symbiosis International University Symbiosis Institute of Technology","correspondingAuthor":false,"prefix":"","firstName":"Pragya","middleName":"","lastName":"Saxena","suffix":""},{"id":342824299,"identity":"9d05c338-c2a2-4b2a-ab64-594bd7ebb730","order_by":1,"name":"Arun Bongale","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABDElEQVRIiWNgGAWjYJACZjDJzsBw4EOFhByIfeABUVqYGRgPzjhjYwzWkkCkFubDvG1piQ0gDj4t/LPPmD0uqLhj19/MfOAwz5nD6fPDDj8E2mInp9uAXYvEuRxz4xlnniXPOMyWcHBOxeHcjbfTDIBako3NDuCw5gyPmTRv2+FkhsM8BgfenAFqmZ0A0nIgcRsOLfIwLfKH+T8cADLSDWenf8CrxQCqxc7gMA/DQaD3E+Slc/DbYniGrUx6xpnDCYaH2QxAgWy4QTqn4ECCAW6/yJ1h3iZdUHHYXu548+MPwKiUl5+dvhnIsJPD6X0ogEQH2KlglQb4lYOAPZwl34Bb1SgYBaNgFIxMAABLemjVa/AWggAAAABJRU5ErkJggg==","orcid":"","institution":"Symbiosis International University Symbiosis Institute of Technology","correspondingAuthor":true,"prefix":"","firstName":"Arun","middleName":"","lastName":"Bongale","suffix":""},{"id":342824300,"identity":"d0875ad7-1ec6-453c-a4b1-165e37bd8ac8","order_by":2,"name":"Satish Kumar","email":"","orcid":"","institution":"Symbiosis International University Symbiosis Institute of Technology","correspondingAuthor":false,"prefix":"","firstName":"Satish","middleName":"","lastName":"Kumar","suffix":""},{"id":342824301,"identity":"bafdb233-9368-46e2-9a33-ff82a016e4b8","order_by":3,"name":"Rajesh Kodbal","email":"","orcid":"","institution":"Symbiosis International University Symbiosis Institute of Technology","correspondingAuthor":false,"prefix":"","firstName":"Rajesh","middleName":"","lastName":"Kodbal","suffix":""}],"badges":[],"createdAt":"2024-07-31 10:33:49","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4834721/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4834721/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":64787494,"identity":"8324b041-347b-45e0-85d2-d54775f660aa","added_by":"auto","created_at":"2024-09-18 20:23:35","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":670359,"visible":true,"origin":"","legend":"\u003cp\u003eFramework for classification of faults in composites fabricated by FSP\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-4834721/v1/6c7c399d8a0cf2b97b259aec.png"},{"id":64787384,"identity":"4bf1729a-b8e1-4678-b596-a63b0ee7f466","added_by":"auto","created_at":"2024-09-18 20:15:35","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":684442,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental setup with sensors\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-4834721/v1/cb7dbdd4624cbfcaaa6eeb03.png"},{"id":64787888,"identity":"26be916e-8b1b-4be6-a13c-91fa27a82c29","added_by":"auto","created_at":"2024-09-18 20:31:35","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":529708,"visible":true,"origin":"","legend":"\u003cp\u003eSensor data signals for vibration (table) sensor for all five labels of composites\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-4834721/v1/6f2f48e27ef08b02d7920300.png"},{"id":64787889,"identity":"20a9cbd4-9061-4272-81e1-29fc1deaed9a","added_by":"auto","created_at":"2024-09-18 20:31:35","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1314298,"visible":true,"origin":"","legend":"\u003cp\u003eDetailed Methodology adopted for fault classification of surface composites prepared by FSP\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-4834721/v1/2a46df98ea923b4c4bd3e0d5.png"},{"id":64787391,"identity":"f15580d3-09bf-4ba0-9422-a77766b2add8","added_by":"auto","created_at":"2024-09-18 20:15:35","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":238122,"visible":true,"origin":"","legend":"\u003cp\u003eMachine learning process for classification of defects in composites.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-4834721/v1/f429bcd3c26c2ea5c562dcbf.png"},{"id":64787492,"identity":"961971f5-8558-4df5-9780-becdabc6da7b","added_by":"auto","created_at":"2024-09-18 20:23:35","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":278908,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Bar plot for comparison of accuracy scores of all ML classifiers; (b) ROC curve for classification of sensor data; for features selected by Information gain method\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-4834721/v1/fc5f248ec3ac07d61d776662.png"},{"id":64787491,"identity":"58ad2edb-967f-488f-92ae-c751fcc1225e","added_by":"auto","created_at":"2024-09-18 20:23:35","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":270017,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Bar plot for comparison of accuracy scores of all ML classifiers; (b) ROC curve for classification of sensor data; for features selected by FFS method\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-4834721/v1/637b5fce1ba511bdc7e9bd17.png"},{"id":64787388,"identity":"401d6f0d-cb68-415f-881d-6e714826f398","added_by":"auto","created_at":"2024-09-18 20:15:35","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":309528,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Bar plot for comparison of accuracy scores of all ML classifiers; (b) ROC curve for classification of sensor data; for features selected by RF method\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-4834721/v1/4b549f25ae579d93c65e234a.png"},{"id":64788083,"identity":"59dcb76f-9ab6-4adf-b61b-8927f433b47a","added_by":"auto","created_at":"2024-09-18 20:39:40","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6238016,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4834721/v1/bb26d7a9-6253-4b8c-8a73-9ebb1dba08b2.pdf"}],"financialInterests":"","formattedTitle":"Classification of faults in friction stir processed composites using a machine learning and ensemble learning approach","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eFriction stir processing (FSP) has recently been established as the advanced method not only to alter the exterior microstructure and characteristics of substance but also to create high-quality surface composites through localized plastic deformation without undergoing melting of the material. Thus, it is an environmentally friendly technique for manufacturing surface composites. FSP has recently been utilized to obtain refined grain microstructure in aluminium alloy-based composites by adding suitable reinforcements and other metal alloys. It is based upon the principles of Friction Stir Welding (FSW) developed by researchers at The Welding Institute, UK [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], for joining two metal pieces utilizing the heat caused by friction. The joining of dissimilar metals, ultrasonic-assisted, and underwater joinings is obtained much more efficiently with friction stir welding technologies. FSP is utilized significantly in various applications, such as aerospace [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e] and the automobile sector, requiring improved physical, chemical, and surface mechanical properties of metallic materials. Recent studies indicate that fundamental mechanisms drive significant microstructural and textural changes during FSW or FSP of different metals and alloys, resulting in a promising technique for fabricating surface hybrid composites [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. FSP consists of a specifically shaped tool with a shoulder integrated with a pin, which, in rotating condition, is moved in the direction of the surface of the metal workpiece as fixed on a machine bed. The pin, integrated into the shoulder, is pushed inside the workpiece so that tool shoulder comes in complete contact with the metal surface. The FSP tool is then passed over the specified lane over the workpiece surface to be processed. The thermal energy is produced due to the friction produced when the tool shoulder slides over the metal surface, ensures localized recrystallization on the workpiece surface [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. The recrystallization helps refine the microstructure of the grain near the matrix surface in the traversed region.\u003c/p\u003e \u003cp\u003eRecent research includes improving the mechanical, thermal, and microstructural properties of the composites manufactured by FSP, with a wide range of mechanical tests optimizing the process parameters [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. The influence of various factors on the surface characteristics is beneficial in determining the ideal range of values for achieving the desired results. However, today\u0026rsquo;s industry demands high production in less time with fewer defects, which leads to the application of Industry 4.0 in FSW and FSP of metal alloys or composites. The application of a machine learning approach for real-time monitoring surface composite manufacturing using the FSP technique with the help of multi-sensor data analysis is very rare. The recent research in this field includes the condition monitoring of the FSW and FSP of the composites utilizing machine learning techniques with the collected sensor data. Machine learning methods significantly regulate the process parameters and monitor friction stir welding and processing techniques. As surveyed by Ammar H. Elsheikh et al. [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], the utilization of machine learning techniques in FSW was examined. The study focused on predicting joint characteristics, online process csontrol, tool failure analysis, and the application of optimization techniques with machine learning methods. The study encompassed various methodologies, including linear regression, kNN, random forest, ANN, SVM, fuzzy system, adaptive neuro-fuzzy inference system, and random vector functional link [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Shivraman Thapliyal et al. [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] utilized a machine learning-based model to investigate how input factors influence the physical characteristics during the copper welding by FSW. It was observed that the neural network obtained the maximum accuracy using a deep leaning model in predicting the ultimate tensile strength over kNN, decision tree, and information gain methods. Researchers are now attempting to employ real-time monitoring by integrating multiple sensors throughout the FSP.\u003c/p\u003e \u003cp\u003eBipul Das et al. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] utilized the online torque sensor data recorded during the FSW process by using the discrete wavelet technique to identify the process faults. Also, the feature selection methods employ the chosen features in the support vector machine to predict the tensile strength of the joints formed by the FSW of AA1100. In another similar study by Sangyoung Yoon et al. [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], defects are identified in the laminated composites using deep learning by highly non-linear solitary waves. The several characteristics influence the classification efficiency of the deep learning algorithms which is investigated. It is observed that the proposed algorithm achieved high accuracies in recognizing the defects and locating them on the composite surfaces. These studies indicate the possibilities of utilizing machine learning and deep learning techniques to achieve online detection system for identification of defects in the composites. The applicability of machine learning (ML) classifiers, fitted and tested on different systems, is investigated by Yuxuan Chen et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] for predicting the faults in air conditioners using sensor data. After selecting suitable features with feature selection techniques, they investigated the significant improvement in the accuracies of machine learning-based classifiers. In another work, Hossein Moosaei et al. [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] attempted to improve the classification accuracy by simultaneously implementing the feature selection and machine learning classification by the twin k-class support vector machine (SVM) algorithm. It was found that the method improved the classification accuracy significantly. An investigation conducted by Ayat Naji Hussain et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] includes applying feature extraction techniques, feature selection techniques, and utilizing machine learning based classifiers to automatically classify the foot disorders to help in leg rehabilitation. It was obtained that the feature selection techniques help in modifying the classification accuracies of the machine learning models up to 99%.\u003c/p\u003e \u003cp\u003eWei Guan et al. [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] evaluated the distinctive patterns of change in the force signals as measured by force sensors for FSW joints, including metal joints that have defects and metal joints that do not have defects, using Fast Fourier transform to analyze the distortion in waveform in the traverse force helping in investigating the material flow and defect formation. Zhiwei Li et al. [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] introduced a technique for circular laser scanning by sensing many parameters in the 5-axis RFSW. The laser displacement sensor underwent rotational movement around the tool, generating a sequence of measured points to capture the three-dimensional form of the workpiece accurately. Using a laser displacement sensor decreases the required quantity of sensors and enables dependable process sensing. Sharda et al. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], 2013, proposed a hybrid filter-type technique for selecting features that combines two methods, Relief-F and Pearson Correlation. The technique aims to reduce the complexity of the classifier and achieve high accuracy in classification algorithms. The result is a feature set with improved performance. The increasing popularity of IIoT tools and applications has led to a growing interest in online condition monitoring systems that offer remote access capabilities. Utilizing these tools to monitor production operations can result in improved performance and heightened efficiency. A group of researchers working on tool wear prediction utilized a tool wear dataset [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Various feature selection techniques, including the PCC, random forest (RF), PCC with RF and PCA with PCC, were employed. The same group of authors also utilized the dataset to estimate the Remaining Useful Life (RUL) of the cutting tool in milling [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. The more the level of excellence in feature engineering conducted on the raw data, higher is the accuracy of the forecast. Adham E. Ragab et al. [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] presented the development of an online monitoring system utilizing Raspberry Pi and thermocouple and force sensors for the purpose of monitoring the peak temperature and axial force during spot welding using friction stir process (FSSW). They examined influence of tool rotation rpm, plunge rate, and time duration on maximum temperatures and axial forces during spot welding by FSW. The study by Santwana Gudadhe et al. [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] involved the classification of cerebral CT images by hemorrhage using various machine learning classifiers. A method is created by combining the features based on texture and transform to develop a combined feature-based approach. The proposed collaborative feature extraction technique enhances the accuracy of CT image classification compared to the features extraction based on texture and transform respectively. The transform-based feature approach utilizes the energy coefficients of CT images. A review work by Balachandar K et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] focuses on the present level of comprehension and advancement of the input factors during the friction based joining and processing techniques. They focused that to determine the most likely causes of failure, many techniques are employed to gather data, such as vibration monitoring. Thermal imaging refers to the process of using specialized equipment to capture and display the infrared radiation emitted by objects, allowing for the visualization of heat patterns and temperature variations. Some researchers examined the additive manufacturing of multilayered Al6061 alloy using the friction stir powder additive manufacturing [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. The goal was to modify physical and grain structural characteristics of the alloy by reducing the temperature variation in the direction of process. This was achieved by maintaining the substrate in close proximity to its designated temperature for artificial aging by utilizing an external heat source within a closed-loop system.\u003c/p\u003e \u003cp\u003eIn this paper, the machine learning approach is utilized for classification of surface defects occurring in composites manufactured by FSP. The multi-senosor data from vibration (tool, matrix), current (tool) and force sensors (matrix) mounted on the milling machine, is collected at a sampling rate of 1000 Hz with the help of a data acquisition system during the fabrication of surface composite samples by FSP. The raw sensor collected is subjected to preprocessing, i.e., data cleaning, eliminating noise, outliers, etc. The relevant features of the preprocessed data are extracted to characterize and reduce the dimension of data. Also, the ten most important features are selected using each of the feature selection techniques used. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the proposed framework for classifying faults in the composite samples prepared by FSP. The paper is categorized into five sections. Section 1 provides an introduction to the framework of the defect classification of composites with related literature. Section 2 represents the material and detailed methodology, viz., sample preparation, FSP experimentation, and, Physical identification of surface defects in fabricated composites utilized for the preparation and analysis of composites. Section 3 represents the sensor data analysis including data acquisition and preprocessing; extraction of relevant features, selection of important features; fault classification models; and, evaluation of models. Section 4 represents the discussion of results, including; extracted features, selected features, performance evaluation using classification models, and comparative result analysis. Section 5 throws light on an important conclusion regarding the selection of optimal methods for the effective classification of faults in surface composites.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"2 Equipment and Methodology","content":"\u003cp\u003e2.1 \u0026nbsp; \u0026nbsp; \u0026nbsp; Experimentation Detail\u003c/p\u003e\n\u003cp\u003eIn the study, the surface composites of Al6061 alloy matrix with copper and graphene reinforcements are prepared by FSP on the CNC milling machine. FSP parameters selected are: Tool rpm\u0026thinsp;=\u0026thinsp;1000 rpm, Tool feed\u0026thinsp;=\u0026thinsp;45 mm/min, depth of cut\u0026thinsp;=\u0026thinsp;pin length, i.e., 5 mm [\u003cspan\u003e23\u003c/span\u003e]. The experimentation involves advancing FSP tool into the Al6061 alloy matrix having holes filled with reinforcement with traverse direction motion. The rubbing action between the tool shoulder and the matrix filled with reinforcement causes friction and the heat generation which produces surface composite with uniform distribution of the particles [\u003cspan\u003e24\u003c/span\u003e]. Various sensors pre-installed on the milling machine, are utilized, during FSP experimentation, i.e., composite fabrication, for analyzing specific parameters. These include two vibration sensors (spindle and table), one current sensor (spindle), and a dynamometer (table) with loads in X, Y, Z axes. The FSP setup for multi-sensor data collection is shown in Fig. \u003cspan\u003e2\u003c/span\u003e.\u003c/p\u003e\n\u003cp\u003e2.2 \u0026nbsp; \u0026nbsp; \u0026nbsp; Signal Acquisition\u003c/p\u003e\n\u003cp\u003eThe composites prepared by FSP are likely to be subjected to various surface defects due to fabrication or machining errors. These defects are identified with the help of signals collected through various sensors, such as vibration sensor, current sensor, and dynamometer. Vibration sensors are sensitive to surface condition and any irregularities on surface may alter their amplitude resulting in detection of faults. The current sensor mounted on the machine spindle, determines the variation in the amount of current drawn by spindle providing insights into the stability of the operation and and indicate inconsistenies occurred due to defects. A table-based dynamometer, is placed between workpiece and machine bed, for collecting the signals for load in X, Y and Z directions to obtain slight variation in the axial and transverse loads [\u003cspan\u003e25\u003c/span\u003e] on the matrix plate during FSP which indicate presence of defects. When a defect occurs as the tool is passed over the matrix surface with reinforcement filled into the holes, the magnitudes of the cutting forces suddenly increase. Vibration levels might vary significantly during FSP. The vibration mainly affects the holding devices for tool or workpieces in multiple directions [\u003cspan\u003e26\u003c/span\u003e]. A current sensor is installed on the machine spindle and determines the current drawn by the tool during process. Also, the vibrations caused during the process are investigated in this study using accelerometers. Two vibration sensors, one on tool spindle, and another on the machine bed where matrix plate is fixed, are utilized for measuring vibration signals on the tool and on the composite being prepared respectively. The signal acquisition equipment are mentioned in Table \u003cspan\u003e1\u003c/span\u003e. A signal acquisition system by National Instruments, (NI USB-6210) has been used to collect the data for further processing. The data acquisition module acts as a medium that helps to digitize the analog signals collected from all the sensors so that they can be synchronized and analyzed simultaneously. One data file consists of the multisensory data during one pass (lap) of FSP tool travel on the matrix filled with reinforcement to form surface composites.\u003c/p\u003e\n\u003cdiv\u003e\n \u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 1\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eDetails of Signal Collection equipment and sampling frequency\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eS. No.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSignal\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMounting location\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSignal Collection Equipment\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVibration (Tool)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTool spindle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVBR1/D0-3V\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVibration (Matrix)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMachine bed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVBR1/D0-3V\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCurrent (Tool)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTool spindle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePico TA018\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDynamometer (Load cell X)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMachine bed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eXEEPL FHC 031\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDynamometer (Load cell Y)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMachine bed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eXEEPL FHC 031\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDynamometer (Load cell Z)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMachine bed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eXEEPL FHC 031\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e\n\u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e\n\u003cp\u003e2.3 \u0026nbsp; \u0026nbsp; \u0026nbsp; Sensor data processing\u003c/p\u003e\n\u003cp\u003eThe sensor data collected is preprocessed, i.e., cleaned and organized, to structure data and remove unnecessary data. This includes the data being transformed or encoded into a more understandable form so that the computer can easily comprehend the multi-sensor signals as recorded. The data labelling includes defining the target variables for each sensor data. According to the visual inspection of the composites prepared, the data is categorized in terms of their defect condition. The reason for the occurrence of these surface defects is the manual sample preparation and experimentation errors. The data is split into features (X) and labels (y). The features refer to the multi-sensor data. The five different labels (categories) as the target variable (y) are selected as normal composite, pin-break composite, rough surface composite, no composite, and brazing-break composite.\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan\u003e3\u003c/span\u003e shows the raw sensor data signals for vibration (table) sensor for all five labels of composites. Normal composite refers to those samples where moderate levels of vibration, current and cutting forces with no significant changes in sensor readings are observed. Pin break type composites are those in which, after the FSP tool traverses, the pin is broken from the shoulder and is stuck into the matrix surface. It is caused by excessively high pin temperature due to the lack of a sufficient amount of release of material chips during the tool traverse. The break point refers to the sudden increase in current drawn by motor due to excessive load and a moderate rise in vibration amplitudes during the end of tool travel. In the rough surface type composites, an uneven surface composite layer is formed which includes voids and clusters of particles. The reason is the uneven surface contact between the tool and matrix surface during FSP due to excessive traverse speed resulting in lack of axial pressure on the matrix surface and improper recrystallization. High cutting forces occur throughout the tool traverse in the rough surface composites. a moderate rise in vibration amplitudes is observed during the end of tool travel in pin break and brazing break composites. No composite type are those in which the tool is not properly in contact with the matrix surface and the holes filled with reinforcement during FSP. This results in no or negligible dispersion embedding the reinforcement into the matrix and hence no manufacturing of composite occurs. This type of composite fabrication shows a significant drop in vibration, current and force signals. The brazing break type composites refer to the composites during fabrication of whom, the brazed tool (FSP tool is a steel body brazed with tungsten carbide shoulder and pin) undergoes failure as the rise in temperature in the tool body reaches the brazing temperature (450֯C approx.) during the tool traverse. There is a gradual rise and then a sudden drop in the vibration signals during the brazing failure. However, the current and load in z direction suddenly rise and drop at the breaking point. Figure \u003cspan\u003e4\u003c/span\u003e shows the detailed methodology adopted for fault classification of surface composites prepared by FSP. Also, the detailed explanation on the steps followed in the methodology, such as sensor data processing, feature extraction, feature selection, and classification of faults in the methodology are discussed in the current and sections further.\u003c/p\u003e\n\u003cp\u003e2.4 \u0026nbsp; \u0026nbsp; \u0026nbsp; Feature Extraction\u003c/p\u003e\n\u003cp\u003eThe important characteristics of sensor data which reflect the distinct and significant characteristics of the process, are extracted for characterizing various aspects of data. The sensor data is acquired at each millisecond as the FSP tool is traversed on the matrix surface. Several significant characteristics of signals in the temporal and spectral domain are retrieved at every second from the preprocessed dataset for reducing the dimension of data and analyzing it effectively. The features are retrieved using standard python libraries including NumPy, SciPy, Pandas, scikit-learn, etc. The time domain and frequency domain features represent the variation in signals with time and frequency bands respectively.\u003c/p\u003e\n\u003cdiv id=\"Sec3\"\u003e\n \u003ch2\u003e\u003cstrong\u003e2.4.1 \u0026nbsp; \u0026nbsp;\u003c/strong\u003eTime domain features\u003c/h2\u003e\n \u003cp\u003eTime based or temporal characteristics are displayed with their functions in Table \u003cspan\u003e4\u003c/span\u003e. These features obtain the time-based characteristics of the signal such as its trend, variability, dynamics, etc. These features study the amplitude statistics, such as mean, median, variance, etc., which provides insights into the central tendency, spread and distribution, etc. The signal shape, magnitude, variation, frequency, stability, etc., are studied by these features.\u003c/p\u003e\n \u003cp\u003eTable 4: Methods utilized in various types of feature selection techniques\u003c/p\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"642\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eS.No.\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eTime-domain Features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFunction\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eECDF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eEmpirical cumulative distribution function concerning time\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eInterquartile range\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eDifference between the upper and lower quartile values of a time series signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eKurtosis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eMeasurement of tailedness of a signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eMaximum\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eCalculation of highest amplitude of a signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eMean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates the average amplitude of signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eMean absolute deviation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eMeasure of the variability of signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eMedian\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eMiddle value or mean of two middle values in a signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003emedian absolute deviation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eAverage distance of data from the median of the signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eMinimum\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eDetermines the smallest utility in signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eRMS value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eRoot mean square of the signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eSkewness\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eAmount of deviation of the data from the sample mean, indicating the extent of asymmetry.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eStandard deviation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eDeviation from the mean of a time series signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eVariance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates Signal variance.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eabsolute energy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eDetermines the true energy of a signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eDistance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eComputes signal travelled distance.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"8.722741433021806%\" valign=\"top\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.077881619937695%\" valign=\"top\"\u003e\n \u003cp\u003eEntropy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"66.19937694704049%\" valign=\"top\"\u003e\n \u003cp\u003eCalculate the Shannon entropy of the signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e\n \u003cdiv id=\"Sec4\"\u003e\n \u003ch2\u003e\u003cstrong\u003e2.4.2 \u0026nbsp; \u0026nbsp;\u003c/strong\u003eSpectral-domain features\u003c/h2\u003e\n \u003cp\u003eThe frequency domain or spectral features provide insight into the spectral characteristics of a signal. They reveal the energy distribution across various components related to frequency in the signal. These features obtain the significant frequency magnitude, dispersion, shape, variation, etc. These are displayed with their functions in Table \u003cspan\u003e5\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eTable 5: Methods used in various machine learning and ensemble learning models\u003c/p\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"633\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eS.No.\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFrequency-domain features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFunction\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eFFT Mean coefficient\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates the average amount of every frequency in the spectrogram.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eFundamental frequency\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eThe signal\u0026apos;s lowest periodic waveform frequency.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eLPCC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eCepstral coefficients for linear prediction.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eMFCC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates the Mel-frequency cepstral coefficients.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eMaximum power spectrum\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eHighest signal\u0026apos;s spectral density and power.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eMaximum frequency\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eHighest signal frequency\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eMedian frequency\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates the centre point of the power distribution.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003ePower_bandwidth\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eBandwidth of the signal\u0026apos;s power spectrum density.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eSpectral centroid\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eRepresents Barycenter of the spectrum.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eSpectral_decrease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eQuantifies reduction in amplitude of the spectra.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eSpectral distance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eComputes the signal spectral gap.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eSpectral entropy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates signal\u0026apos;s spectral entropy using the Fourier transform.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eSpectral kurtosis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eMeasures dispersion relative to the average value.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eSpectral positive turning\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates count of positive inflection points in signal by FFT\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eSpectral roll-off\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates frequency below to signal\u0026apos;s total power is contained.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e17\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eSpectral roll-on\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eComputes the signal spectral roll-on\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eSpectral_skewness\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eQuantifies the lack of symmetry relative to its average.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eSpectral slope\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eIt calculates the gradient of frequency data.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eSpectral spread\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eQuantifies the extent to which the values in the spectrum deviate from their average value.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eSpectral variation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates the degree of variability in the spectrum over time.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eWavelet absolute mean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates the wavelet scales\u0026apos; absolute mean value using the Continuous Wavelet Transform (CWT).\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e23\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eWavelet energy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates the CWT energy for each scale of the wavelet.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eWavelet entropy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eContinuous Wavelet Transform (CWT) entropy of the signal.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eWavelet standard\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates standard value of the CWT for each scale of wavelet.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"7.424960505529226%\" valign=\"top\"\u003e\n \u003cp\u003e26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.436018957345972%\" valign=\"top\"\u003e\n \u003cp\u003eWavelet variance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"64.13902053712481%\" valign=\"top\"\u003e\n \u003cp\u003eCalculates variance of each wavelet scale using CWT\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e\n \u003cp\u003eTable 6: Methods utilized in various types of feature selection techniques\u003c/p\u003e\n \u003cdiv\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"10.782241014799155%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eS.No.\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"37.2093023255814%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eFeature Selection Type\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"52.00845665961945%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMethods used\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"10.782241014799155%\" valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"37.2093023255814%\" rowspan=\"6\" valign=\"top\"\u003e\n \u003cp\u003eFilter Methods\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"52.00845665961945%\" valign=\"top\"\u003e\n \u003cp\u003eChi square method\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"17.171717171717173%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"82.82828282828282%\" valign=\"top\"\u003e\n \u003cp\u003eF_classif method\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"17.171717171717173%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"82.82828282828282%\" valign=\"top\"\u003e\n \u003cp\u003eInformation gain\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"17.171717171717173%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"82.82828282828282%\" valign=\"top\"\u003e\n \u003cp\u003eCorrelation\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"17.171717171717173%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"82.82828282828282%\" valign=\"top\"\u003e\n \u003cp\u003eMean absolute difference\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"17.171717171717173%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"82.82828282828282%\" valign=\"top\"\u003e\n \u003cp\u003eVariance threshold\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"10.782241014799155%\" valign=\"top\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"37.2093023255814%\" rowspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eWrapper Methods\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"52.00845665961945%\" valign=\"top\"\u003e\n \u003cp\u003eForward feature selection (FFS)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"17.171717171717173%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"82.82828282828282%\" valign=\"top\"\u003e\n \u003cp\u003eRecursive feature elimination (RFE)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"10.782241014799155%\" valign=\"top\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"37.2093023255814%\" rowspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003eEmbedded Methods\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"52.00845665961945%\" valign=\"top\"\u003e\n \u003cp\u003eLasso Regularization\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"17.171717171717173%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"82.82828282828282%\" valign=\"top\"\u003e\n \u003cp\u003eRandom Forest Importance\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003e2.5 \u0026nbsp; \u0026nbsp; \u0026nbsp; Feature selection\u003c/p\u003e\n \u003cp\u003eThe technique for selection of features is the act of choosing very important relevant features among the extracted features which promotes to develop simpler models for further analysis. The reduced dimension of data helps to improve the accuracies of the machine learning classifiers with a shorter training time and reduced overfitting. In this study, three different types of feature selection techniques are utilized: Filter type, wrapper type and embedded type methods.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec5\"\u003e\n \u003ch2\u003e\u003cstrong\u003e2.5.1 \u0026nbsp; \u0026nbsp;\u003c/strong\u003eFilter methods\u003c/h2\u003e\n \u003cp\u003eThese methods select the variables or features based on the attributes of the data and are independent of the machine learning approach. In these methods, the features are ranked based on their relevance and highest rank features are chosen to induce the classification models. The filter methods utilized in the study are: Chi square method, F_classif method, information gain, correlation, dispersion ratio, mean absolute difference and variance threshold methods. Among filter methods, Chi square method for selecting important characteristics is utilized to perform the chi square (\u0026chi;2) test which measures the autonomy among two parameters and identifies most informative ten characteristics according to their importance with respect to the target parameter. The F_Classif method for feature selection is used for calculating the ANOVA F-statistic values for all the features extracted and selecting the top ten features based on those values. The information gain method is utilized to obtain mutual information between the features and their target variables by mutual_info_classif function. The features with ten highest scores obtained are selected for further analysis. The correlation method quantifies the linear interconnection of the features with target variable. The correlation coefficients are calculated and most significant ten features are selected. The Mean Absolute Difference (MAD) is used to calculate the mean absolute deviation or variability of each feature from the feature\u0026apos;s mean value and features with ten highest variabilities are selected. Variance Threshold is used to eliminate features with low variance. In all of the filter methods utilized, the features are ranked based on the respective evaluation criteria (different for all methods) by eliminating the features having constant or quasi constant values. The filter methods provide simple and powerful methods to eliminate irrelevant, duplicated, correlated, and hence the redundant features in an easier and quicker way.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec6\"\u003e\n \u003ch2\u003e\u003cstrong\u003e2.5.2 \u0026nbsp; \u0026nbsp;\u003c/strong\u003eWrapper methods\u003c/h2\u003e\n \u003cp\u003eThese use the machine learning models for evaluating the quality for selecting the subset of features. It involves the training of new model on each feature subset and determines best performing feature subset based on machine learning algorithm. They are also capable to detect the interactions between the variables, so provide better prediction performance than that by filter methods. The wrapper methods used in this study are: forward feature selection (FFS) and recursive feature elimination (RFE) techniques. FFS begins with an empty model, adds features one by one, and evaluates the model, until the accuracy is increased. On the contrary, RFE begins with a model with all features, evaluates the model, eliminates features one by one based on least importance, until desired number of features is reached.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec7\"\u003e\n \u003ch2\u003e\u003cstrong\u003e2.5.3 \u0026nbsp; \u0026nbsp;\u003c/strong\u003eEmbedded techniques\u003c/h2\u003e\n \u003cp\u003eThe embedded techniques perform the feature selection during the training of data in machine learning classification algorithm. These methods are comparatively faster than the wrapper methods and achieve higher accuracies in comparison to filter methods with lower risk of overfitting of data. Embedded methods utilized in the study are: Lasso regularization and Random Forest Importance methods. Lasso regularization method can shrink some of the coefficients to zero so that some features can be eliminated. Random Forest Importance method involves building of random forest trees, calculating feature importance, and eliminating the unimportant features until the condition is achieved. Table \u003cspan\u003e6\u003c/span\u003e shows methods used for various the feature selection types discussed.\u003c/p\u003e\n \u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e\n \u003ch2\u003e2.6 \u0026nbsp; \u0026nbsp; Fault Classification\u003c/h2\u003e\n \u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e\n \u003cp\u003eFeature classification refers to the categorization of features using machine learning. The selected features of the sensor data are classified using different machine learning classification algorithms (classifiers). Figure \u003cspan\u003e5\u003c/span\u003e shows the workflow of the classification of defects in FSP prepared composites using machine learning approach [\u003cspan\u003e27\u003c/span\u003e]. The multi-sensor data selected using feature selection techniques, which is in the form of features and their labels, divides itself into a training and a testing dataset. The machine learning classifier is fit with the labelled training data and its efficacy is tested with the unlabelled test dataset by comparing the predicted labels with the true labels of the test dataset. Also, there is scope for adjustment of different parameters of the ML model using the hyperparameter tuning [\u003cspan\u003e28\u003c/span\u003e]. Eight classifiers are utilized to classify the features selected by each of the selection methods as discussed individually. The Different machine learning (LR, NB, kNN, SVM, DT) [\u003cspan\u003e29\u003c/span\u003e], [\u003cspan\u003e30\u003c/span\u003e] and ensemble learning classifiers (RF, AB, GB) [\u003cspan\u003e31\u003c/span\u003e]\u0026ndash;[\u003cspan\u003e33\u003c/span\u003e] categorize the selected feature data on dataset for training and validate it on the test dataset. Various feature selection methods and machine learning classifiers are utilized for identifying and categorizing the most important features and hence classifying the defects in fabricated composites.\u003c/p\u003e\n \u003cp\u003ei. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;k-Nearest Neighbors method (kNN)\u003c/p\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; The k-Nearest Neighbors is a non-parametric method which clusters the data and assigns a new label to the data on the basis of the closeness of the values according to the Euclidean distance between the data points to make classification of data. It calculates the distance to the training data using the selected distance metric.The hyperparameter tuning is used for adjustment of parameters, i.e., the number of neighbours and window size. kNN is a non-parametric method capable of handling multisensory data.\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eii. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; Support vector machine (SVM)\u003c/p\u003e\n \u003cp\u003eSupport Vector Machine method applied for multi-class classification, is capable to handle high dimensional data exhibiting non-linear complex relationships among data. Two approaches are mainly used: one versus one where all the pairs of classes are considered for training the model and one versus all where n number of output classes and n number of classification models are trained concurrently while accounting for the separation between the real class and the remaining classes.\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eiii. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Gradient boosting classifier (GB)\u003c/p\u003e\n \u003cp\u003eGradient boosting is an ensemble method that constructs models in a sequential manner. Every subsequent model is educated to rectify the mistakes committed by its previous models. The algorithm consists of initializing the model with a base estimator and then adding more models iteratively by computing errors. Then the model is trained to predict errors and the ensemble is updated with new model for classification and evaluation of the model.\u003c/p\u003e\n \u003cp\u003eiv. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Decision tree classifier (DT)\u003c/p\u003e\n \u003cp\u003eThe decision tree classifier is highly useful for classification of time series \u0026nbsp; sensor data since it is capable of handling non-linear relationships of data and feature interactions. It is a non-parametric type of machine learning technique that is easier to interpret, robust to noise, and provide a measure of feature importance. However, it may promote overfitting of data during training and data instability, i.e., small variation in data may lead to different splits. These limitations can be overcome by using ensemble methods such as, random forest classifiers, etc.\u003c/p\u003e\n \u003cp\u003ev. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; Random forest Classifier (RF)\u003c/p\u003e\n \u003cp\u003eRandom Forest is a type of ensemble method for classification that improves the accuracy of predictions by combining multiple decision trees trained on various groups of the data and using averaging. The random forest method combines predictions from many decision trees and generates the final output based on the majority vote among these predictions. It functions in two separate stages. The initial stage is creating a random forest by combining N decision trees. The second stage involves creating predictions for each independent tree that was constructed in the beginning stage. Random forest classifier is capable to effectively manage extensive datasets with a high number of dimensions. Also, it improves the precision of the model and mitigates the risk of overfitting the data.\u003c/p\u003e\n \u003cp\u003evi. \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Ada boost classifier (AB)\u003c/p\u003e\n \u003cp\u003eThe AdaBoost \u0026nbsp;or Adaptive Boosting, is a magnifying approach employed as an ensembling technique among different machine learning approaches. The technique includes reassigning weightage to each instance, with bigger weights given to examples that were categorized erroneously. While the data is trained, it produces a variable number of decision trees. When first model is built, the improperly categorized data in the first model gets a priority. This imposes these data records to be provided as intake to the second model and so on. This method continues til it is designated by a specific quantity of base learners needed to generate.\u003c/p\u003e\n \u003cp\u003evii. \u0026nbsp; \u0026nbsp; \u0026nbsp; Na\u0026iuml;ve bayes (NB)\u003c/p\u003e\n \u003cp\u003eIt is a widely used machine learning method employed for tasks involving classification, such as text classification. This algorithm is classified as a generative learning algorithm, as it models the input distribution for a certain class or category. This approach relies on the premise that the characteristics of the input data are statistically unrelated for the class, enabling the algorithm to produce fast and accurate predictions.\u003c/p\u003e\n \u003cp\u003eviii. \u0026nbsp; \u0026nbsp; \u0026nbsp;Logistic Regression (LR)\u003c/p\u003e\n \u003cp\u003eIt is a statistical technique and a machine learning method utilized for classification tasks, relying on the principle of probability. This method is employed when the dependent variable, also known as the target variable, is categorical in nature. Logistic Regression is a commonly employed method because to its great efficiency and little requirement for computer resources. It exhibits enhanced efficiency by eliminating variables that possess negligible or no correlation with the output variable. Table 7 displays the methods used for various types of classification models used.\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eTable 7: Selected list of features by wrapper methods used with accuracy scores\u003c/p\u003e\n \u003cdiv\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"14.470284237726098%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eS.No.\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"48.83720930232558%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eClassification Model\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"36.69250645994832%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMethods Used\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"14.470284237726098%\" valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"48.83720930232558%\" rowspan=\"5\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eMachine Learning Models\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"36.69250645994832%\" valign=\"top\"\u003e\n \u003cp\u003eLR\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"28.282828282828284%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"71.71717171717172%\" valign=\"top\"\u003e\n \u003cp\u003eNB\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"28.282828282828284%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"71.71717171717172%\" valign=\"top\"\u003e\n \u003cp\u003ekNN\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"28.282828282828284%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"71.71717171717172%\" valign=\"top\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"28.282828282828284%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"71.71717171717172%\" valign=\"top\"\u003e\n \u003cp\u003eDT\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"14.470284237726098%\" valign=\"top\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"48.83720930232558%\" rowspan=\"3\" valign=\"top\"\u003e\n \u003cp\u003eEnsemble Learning Models\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"36.69250645994832%\" valign=\"top\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"28.282828282828284%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"71.71717171717172%\" valign=\"top\"\u003e\n \u003cp\u003eAB\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"28.282828282828284%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"71.71717171717172%\" valign=\"top\"\u003e\n \u003cp\u003eGB\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e\n \u003ch2\u003e2.7 \u0026nbsp; \u0026nbsp; Evaluation of Models\u003c/h2\u003e\n \u003cp\u003eThe machine learning classifiers are utilized to categorize the training set of data and then classification is employed on the test data. The classification of the data as conducted by the model algorithm, is afterwards, validated from the true values with the help of evaluation of models\u0026nbsp;[34]\u0026nbsp;which suggests how the model has performed in terms of predicting classes. There are several measures which determine the model performance:\u003c/p\u003e\n \u003cp\u003ei) \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Accuracy\u003cstrong\u003e:\u0026nbsp;\u003c/strong\u003e This evaluates the model\u0026apos;s accuracy by calculating the accuracy as the percentage of correctly anticipated instances to the total number of occurrences. Also, it can be determined as the rate of accurately recognized affirmative instances and true negatives to the overall positive and negative values. \u0026nbsp;It is expressed as:\u003c/p\u003e\n \u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e\n \u003cdiv id=\"Equa\"\u003e\n \u003cdiv id=\"FileID_Equa\" name=\"EquationSource\"\u003e$$\\:Accuracy=\\:\\frac{Number\\:of\\:correct\\:instances}{Total\\:number\\:of\\:occurences}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003e\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;...(1)\u003c/p\u003e\n \u003c/div\u003e\n \u003cp\u003eii) \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; Precision\u003cstrong\u003e: \u0026nbsp;\u003c/strong\u003eIt quantifies the precision of optimistic forecasts. The term \u0026quot;precision\u0026quot; refers to the precise ratio of accuracy predicted the ratio of positive cases to the overall cases or instances predicted as positive. It helps in determining the frequency at which a positive test accurately identifies a true positive. The term \u0026quot;positive predictive value\u0026quot; refers to the proportion of real positive results compared to false positive results. It is expressed as:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e$$\\:Precision=\\frac{True\\:Positives\\:}{True\\:Positives+False\\:Positives}$$\u003c/p\u003e\n \u003c/span\u003e\n \u003cdiv id=\"Equb\"\u003e\u003cbr\u003e\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003e\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;.(2)\u003c/p\u003e\n \u003c/div\u003e\u003cspan\u003e\n \u003cp\u003eiii) \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Recall\u003cstrong\u003e: \u0026nbsp;\u003c/strong\u003eIt is also called as sensitivity or the real positive rate. It evaluates the model\u0026apos;s capacity to accurately identify and include all events that are classified as positive. The term \u0026quot;precision\u0026quot; refers to the ratio of correctly predicted positive instances to the total \u0026nbsp;confirmed positive instances. It evaluates the test\u0026apos;s ability to yield a positive result when the condition is present. It can be expressed as:\u003c/p\u003e\n \u003c/span\u003e\n \u003cdiv id=\"Equc\"\u003e\n \u003cdiv id=\"FileID_Equc\" name=\"EquationSource\"\u003e$$\\:Recall=\\frac{True\\:Positives\\:}{True\\:Positives+False\\:Negatives}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003e\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;.(3)\u003c/p\u003e\n \u003c/div\u003e\u003cspan\u003e\n \u003cp\u003eiv) \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;F1\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003escore\u003cstrong\u003e:\u0026nbsp;\u003c/strong\u003eIt integrates the measures of precision and recall into a unified metric. It is especially beneficial in cases where there is an imbalanced distribution of classes. Also, it is said to be the harmonic mean of the recall and accuracy. It can be expressed as:\u003c/p\u003e\n \u003c/span\u003e\n \u003cdiv id=\"Equd\"\u003e\n \u003cdiv id=\"FileID_Equd\" name=\"EquationSource\"\u003e$$\\:F1=2*\\frac{precision*recall}{precision+recall}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003e\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;\u0026hellip;.(4)\u003c/p\u003e\n \u003c/div\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"3 Results and Discussion","content":"\u003cp\u003eThe sensor data recorded during the manufacturing of surface composites by FSP, includes six incoming signals, i.e., three cutting forces (X, Y and Z directions on matrix), one current (tool spindle) and two vibration sensors (tool spindle and matrix). The multi sensor signals are digitized for synchronized analysis with the help of the data acquisition system module. The raw sensor data is preprocessed and labelled with their defect condition as target variables. Then important features are extracted, and most important fewer features are selected for further analysis. The fault identification and classification of the surface defects in the composites is performed using different machine learning classifiers with their performance evaluation. The classification algorithms utilized include kNN, SVM, GB, decision tree, and random forest, ada boost, naive bayes, and logistic regression. The performance of all the classifiers used are compared using different evaluation metrics. The python programming in Jupiter notebook was used for extracting, selecting and classifying the sensor data.\u003c/p\u003e\n\u003cp\u003e3.1 \u0026nbsp; \u0026nbsp; \u0026nbsp; Feature Extraction\u003c/p\u003e\n\u003cp\u003eThe features are extracted from the raw sensor data (six sensors) that can be processed while preserving the information in the original dataset using standard libraries in python. Different time (temporal) and feature (spectral) domain features are extracted at a sampling frequency of 1000 Hz and a window size of 1000 samples is selected for the extraction. Time domain features are extracted by \u0026lsquo;NumPy\u0026rsquo; (numerical python) and \u0026lsquo;pandas\u0026rsquo; libraries in python. Spectral domain features are extracted using \u0026lsquo;PyWavelets\u0026rsquo;, \u0026lsquo;scikit-learn\u0026rsquo;, \u0026lsquo;joblib\u0026rsquo;, \u0026lsquo;SciPy\u0026rsquo; (Scientific Python) libraries, etc. Time domain features extracted in the present research are 16 for each raw sensor data. The frequency domain features extracted are 26 for each sensor data. Overall, 252 features are extracted which includes 96 time domain and 156 sprectral features for all six sensor data. All the extracted features, alongwith the label (condition of defect) are stored in the form of csv files where the features are considered as the independent parameter and label is taken as the target parameter.\u003c/p\u003e\n\u003cp\u003e3.2 \u0026nbsp; \u0026nbsp; \u0026nbsp; Feature Selection\u003c/p\u003e\n\u003cp\u003eThe most important ten features relevant for defect classification are selected out of the extracted features dataset using the features selection techniques including various filter methods, viz., Chi Square method, F_classif method, information gain method, correlation method, dispersion ratio, mean absolute difference, and, variance threshold method. Also, various wrapper methods, such as, forward feature selection and recursive feature elimination methods, and embedded methods, such as lasso regularization, and random forest importance method are utilized for feature selection.\u003c/p\u003e\n\u003cdiv id=\"Sec9\"\u003e\n \u003ch2\u003e3.2.1 Feature Selection using Filter Methods\u003c/h2\u003e\n \u003cp\u003eAmong all the filter methods used for feature selection, i.e., Chi square, F_classif, information gain, correlation, dispersion ratio, mean absolute difference and variance threshold method, It is obtained that among these, the information gain filter method achieves high performance accuracies with the machine learning classification models. This approach quantifies the quantity of data acquired about one variable through another variable. The information gain is well-suited for capturing complex dependency of features on the target variable, rendering it a good option for selecting the sensor data exhibiting linear as well as non-linear relationships. From the extracted time domain and spectral domain features, ten most important features are selected by calculating the mutual information (MI) of features with respect to the target variables. This calculated value refers to the reduction in uncertainty of target variable, given the feature. Then, the features are ranked according to their mutual information scores as calculated and the features with top ten MI scores are selected for further classification. Table \u003cspan\u003e8\u003c/span\u003e displays the record of ten most significant characteristics selected by each of the six filter methods used.\u003c/p\u003e\n \u003cdiv\u003e\u0026nbsp;\u003ctable id=\"Tab12\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 8\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eSelected ten features by each of the Filter methods\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eS. No.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eChi Square Method\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eF_Classif\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eInformation gain\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eCorrelation\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMean Absolute Difference\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eVariance Threshold\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_ECDF Percentile_0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_ECDF Percentile_1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Median\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_LPCC_1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Spectral distance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_LPCC_1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_ECDF Percentile_1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Mean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Root mean square\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_LPCC_11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4_Spectral distance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_LPCC_10\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Max\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Median\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_ECDF Percentile_0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_MFCC_0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_Spectral distance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_LPCC_11\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Mean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Root mean square\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Mean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_Spectral slope\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5_Spectral distance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_LPCC_2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Median\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Wavelet absolute mean_1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_ECDF Percentile_1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_Spectral centroid\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Spectral distance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_MFCC_0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Root mean square\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Wavelet absolute mean_4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Min\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_MFCC_9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0_Spectral distance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_MFCC_1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_Histogram_5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Wavelet absolute mean_5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Wavelet absolute mean_8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_LPCC_2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_Power bandwidth\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_MFCC_10\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Wavelet energy_7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Wavelet absolute mean_6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Wavelet absolute mean_7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_LPCC_10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4_Histogram_5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_MFCC_11\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Wavelet standard deviation_7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Wavelet absolute mean_7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Wavelet absolute mean_6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_MFCC_1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4_Power bandwidth\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_MFCC_9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Wavelet variance_7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Wavelet absolute mean_8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Wavelet absolute mean_5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_MFCC_11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4_Histogram_4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_Spectral centroid\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e\n \u003cdiv id=\"Sec10\"\u003e\n \u003ch2\u003e3.2.2 Feature Selection using Wrapper Methods\u003c/h2\u003e\n \u003cp\u003eAmong the two wrapper methods used, i.e., forward feature selection (FFS) and recursive feature elimination (RFE) for selecting the significant features for classification of sensor data, it is obtained that the features selected by FFS are more efficiently classified by machine learning. In FFS, one by one the features, which improve the model performance the most (highest accuracy scores), are added until ten most significant features are achieved. It is advantageous for the large set of data, is less computationally expensive, promotes incremental improvement, and prevents multicollinearity which is common issue in sensor data. Table \u003cspan\u003e9\u003c/span\u003e shows the ten selected features by both of the wrapper type methods with individual acuuracy scores.\u003c/p\u003e\n \u003cdiv\u003e\u0026nbsp;\u003ctable id=\"Tab14\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 9\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eSelected list of features by wrapper methods used with accuracy scores\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eS.No.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eForward Feature Selection\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eRecursive Feature Elimination\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFeature\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAccuracy Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFeature\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAccuracy Score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_ECDF Percentile\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.996\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_ECDF Percentile_1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1067.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Mean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.996\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Mean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1107.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4_ECDF Percentile_1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.996\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Median\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1081.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4_ECDF_0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.996\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Root mean square\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1115.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_FFT mean coefficient_69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.996\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Wavelet absolute mean_1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1037.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_Wavelet variance_0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.996\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Wavelet absolute mean_4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1055.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4_Wavelet absolute mean_0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.996\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Wavelet absolute mean_5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1077.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4_Wavelet energy_4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.996\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Wavelet absolute mean_6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1091.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5_MFCC_0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.996\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Wavelet absolute mean_7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1099.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5_Spectral centroid\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.996\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Wavelet absolute mean_8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1103.7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec11\"\u003e\n \u003ch2\u003e3.2.3 Feature Selection using Embedded Methods\u003c/h2\u003e\n \u003cp\u003eIn this study, among embedded methods for selecting important characteristics, Lasso regularization and random forest importance methods are utilized to select ten very important characteristics of the dataset. It is obtained that the Random Forest Importance method is capable of handling large amount of data having non-linear relationships, prevents overfitting, and creates ranking of feature importances. Hence the features selected by this method achieve high accuracy with the machine learning classifiers. Table \u003cspan\u003e10\u003c/span\u003e shows the list of top ten features selected by both of the embedded methods alongwith the rank of importance for further classification.\u003c/p\u003e\n \u003cdiv\u003e\u0026nbsp;\u003ctable id=\"Tab16\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 10\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eSelected list of features by embedded methods used with accuracy scores\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eS.No.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eLASSO Regularization method\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eRandom Forest Feature Selection\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eFeature\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eImportance\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eFeature\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eImportance\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_Spectral decrease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.811674\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_Wavelet energy_0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.013291\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Histogram_6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.615873\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_Wavelet standard deviation_0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.011518\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0_Spectral kurtosis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.582926\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Root mean square\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.011256\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Spectral kurtosis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.535394\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_ECDF Percentile_0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.010952\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2_Histogram_9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.487217\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Mean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.010774\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5_Spectral centroid\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.392809\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Median\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.009951\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5_Spectral decrease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.374709\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3_Wavelet variance_0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.009383\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Fundamental frequency\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.259443\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_ECDF Percentile_1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.008665\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0_Maximum frequency\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.158904\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Wavelet absolute mean_4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.008509\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_MFCC_4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.079898\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1_Wavelet absolute mean_8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.008349\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cdiv\u003e\n \u003ch2\u003e3.3 \u0026nbsp; \u0026nbsp; Performance Evaluation using Classification Models\u003c/h2\u003e\n \u003c/div\u003e\n \u003cp\u003eThe selected features of multi-sensor data with all the techniques utilized for selecting features (filter, wrapper and embedded type methods), are provided to different machine learning and ensemble learning classifiers, i.e., KNN, SVM, GB, decision tree, random forest, Ada Boost, naive bayes, and logistic regression. For classification, the selected data, i.e., features (X) and labels (y) are divided into training (70%) and testing (30%) datasets. The labelled training dataset (X\u003csub\u003e1\u003c/sub\u003e, y\u003csub\u003e1\u003c/sub\u003e) data is provided towards the machine learning models for fitting or training purpose. However, the unlabelled testing set (X\u003csub\u003e2\u003c/sub\u003e) is provided for the performance evaluation, i.e., for validating the predicted label values with the true labels (y\u003csub\u003e2\u003c/sub\u003e).\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec12\"\u003e\n \u003ch2\u003e3.3.1 Model Performance by Filter methods of Feature Selection\u003c/h2\u003e\n \u003cp\u003eThe evaluation of the accuracies of classifiers for the selected features by all the filter methods used suggest that the highest performance is obtained with the features selected by Information gain method of feature selection. The metrics for performance of the features selected by this method are shown in Table \u003cspan\u003e11\u003c/span\u003e. It is obtained that among all the classifiers used, the Gradient Boosting classifier achieves highest accuracy i.e., 0.99 for classification of the data. Figure \u003cspan\u003e6\u003c/span\u003e (a) shows the bar plot for comparing the accuracies of all the machine learning based classifiers used. Figure \u003cspan\u003e6\u003c/span\u003e (b) shows that ROC curve for classification model of the selected features is nearly perfect in distinguishing between positive and negative classes with no false positives or false negatives and AUC\u0026thinsp;=\u0026thinsp;1. The reason for the excellent performance of GB classifier is that it builds the model sequentially by fitting new models to the residual errors made by previous models by improving mistakes of the other models, leading to high efficiency. Also, being ensemble of multiple trees, it captures non-linear relationships, can model complex patterns, handles overfitting, and involves feature interactions.\u003c/p\u003e\n \u003cdiv\u003e\u0026nbsp;\u003ctable id=\"Tab18\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 11\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003ePerformance of the classifiers of features selected by Information gain FS method\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eS.No.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMachine learning\u0026nbsp;Classifier\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eF1 score\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eKNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGradient Boosting Classifier\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDecision Tree\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAda Boost Classifier\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNaive Bayes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLogistic Regression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec13\"\u003e\n \u003ch2\u003e3.3.2 Model Performance by Wrapper methods of Feature Selection\u003c/h2\u003e\n \u003cp\u003eThe classification models are examined for their efficacies to classify the features selected by both the wrapper methods used, i.e., FFS and RFE, suggests that the highest performance is obtained with the variables chosen by FFS method. The metrics for the performance of the classifiers are shown in Table \u003cspan\u003e12\u003c/span\u003e. It is obtained that Random Forest classifier achieves maximum accuracy i.e., 0.99 for classification of the data. Figure \u003cspan\u003e7\u003c/span\u003e (a) shows the bar plot for comparison of accuracies by all the machine learning and ensemble learning classifiers used. Figure \u003cspan\u003e7\u003c/span\u003e (b) displays the ROC curve for classification model of the selected features which is good in distinguishing between positive and negative classes with no false positives or false negatives and AUC is 0.98 which is close to 1. The reason for the excellent performance of Random Forest classifier is that being an ensemble learning technique, it produces multiple decision trees and integrates the results obtained by all of them, perfroms averaging reducing the variance of the model leading to enhanced and precise forecasts in comparison to a single decision tree. Also, it employs bootstrap aggregating or bagging which includes training of every tree from the random data subset, while reducing overfitting and enhancing generalization.\u003c/p\u003e\n \u003cdiv\u003e\u0026nbsp;\u003ctable id=\"Tab20\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 12\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003ePerformance of the classifiers of features selected by FFS method\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eS.No.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMachine learning\u0026nbsp;Classifier\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eF1 score\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eKNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGradient Boosting Classifier\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDecision Tree\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAda Boost Classifier\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.66\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.64\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNaive Bayes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLogistic Regression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.86\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec14\"\u003e\n \u003ch2\u003e3.3.3 Model Performance by Embedded methods of Feature Selection\u003c/h2\u003e\n \u003cp\u003eThe multiple machine learning and ensemble learning classifiers are investigated of performance of the classification algorithms on the data features selected by both the embedded methods used suggest that the highest performance is obtained with the features selected by Random Forest Importance method of feature selection. The metrics to evaluate performance of the classifiers are shown in Table \u003cspan\u003e13\u003c/span\u003e. It is obtained that the classifiers used, the Random Forest classifier achieves maximum accuracy i.e., 0.96 for classification of the data. Figure \u003cspan\u003e8\u003c/span\u003e (a) shows the bar plot for comparison of accuracies of all the machine learning classifiers used. Figure \u003cspan\u003e8\u003c/span\u003e (b) shows that ROC curve for classification model of the selected features is good in distinguishing between positive and negative classes and AUC is 0.95 which is close to 1. The reason for the better performance of Random Forest classifier is that it is an ensembles different decision trees simultaneously and can model complex non-linear relationships and features interactions without requiring explicit specification. Random forest classification is especially useful for time series data, as acquired in the present study, where relationships can be highly non-linear and intricate.\u003c/p\u003e\n \u003cdiv\u003e\u0026nbsp;\u003ctable id=\"Tab22\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 13\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003ePerformance of the classifiers of features selected by RFI method\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eS.No.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMachine learning\u0026nbsp;Classifier\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eF1 score\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eKNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.81\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.83\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGradient Boosting Classifier\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDecision Tree\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAda Boost Classifier\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.42\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNaive Bayes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLogistic Regression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003e3.4 Comparative Result Analysis\u003c/p\u003e\n \u003cp\u003eThis subsection discusses the comparison and analysis of efficiency of various machine learning classifiers utilized for the selected features with different feature selection techniques. Table \u003cspan\u003e14\u003c/span\u003e displays the comparative analysis of performance of several feature selection techniques used with different classifiers for the detecting and classifying the surface defects on the composites prepared by FSP. This analysis verifies that the chi square method for feature selection method performs better among other filter methods. Similarly, the forward feature selection method performs well among wrapper methods, and, random forest feature selection performs well among embedded feature selection methods. It is obtained that the ensemble type machine learning classifiers perfrom better than other classifiers.\u003c/p\u003e\n \u003cdiv\u003e\u0026nbsp;\u003ctable id=\"Tab23\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 14\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eComparison of Model Accuracies of several feature selection methods for all classifiers.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eS. No.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eFeature Selection / Classifier\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eKNN\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eGB\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDT\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRF\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAB\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNB\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eLR\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eChi-Square\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.50\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eF_classif\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.32\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eInformation Gain\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCorrelation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.66\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.55\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMean Absolute Difference\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVariance Threshold\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.66\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.66\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.54\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eForward Feature Selection\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRecursive Feature elimination\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLASSO regularization\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.90\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRandom Forest importance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.66\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003c/div\u003e\n\u003c/div\u003e\t"},{"header":"4 Conclusion","content":"\u003cp\u003eInvestigation, includes fabricating Al6061 alloy composites with copper and graphene reinforcement via FSP and analyzing multi-sensor data to identify and classify surface defects. During each FSP process, data from vibration sensors, current sensors, and dynamometers are collected and processed. Features are extracted to reduce data size and highlight significant characteristics. These features are labeled based on five defect conditions\u0026mdash;normal composite, pin break, rough surface, no composite, and brazing break composite\u0026mdash;determined by visual inspection. Various feature selection methods and machine learning classifiers are employed to determine the key characteristics that have the highest significance. and classify the defects. The study focuses on the significance of techniques for selecting important features in enhancing accuracies achieved by classification models and aims to identify the optimal techniques, providing recommendations for researchers working with the dataset. The results reveal optimal combinations of machine learning classifiers and feature selection methods for classifying defects in FSP-prepared composites. The highest accuracies were achieved with the following combinations of feature selection and machine learning techniques:\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e1. Among all the Filter Methods used, Information gain feature selection in combination with Gradient Boosting ensemble technique achieves maximum accuracy.\u003cbr\u003e\u003c/span\u003e\u003cspan\u003e2. Among all the Wrapper Methods used, Forward Feature Selection in combination with Random Forest ensemble technique achieves maximum accuracy.\u003cbr\u003e\u003c/span\u003e\u003cspan\u003e3. Among all embedded methods used, Random Forest importance in combination with Random Forest ensemble technique achieves maximum accuracy.\u003c/span\u003e\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cul\u003e\n \u003cli\u003eFSP \u0026ndash; Friction Stir Processing\u003c/li\u003e\n \u003cli\u003eFFS - Forward feature selection\u003c/li\u003e\n \u003cli\u003eRFE - Recursive feature elimination\u003c/li\u003e\n \u003cli\u003eLR \u0026ndash; Logostic Regression\u003c/li\u003e\n \u003cli\u003eNB \u0026ndash; Naive Bayes\u003c/li\u003e\n \u003cli\u003ekNN \u0026ndash; k-Nearest Neighbours\u003c/li\u003e\n \u003cli\u003eSVM \u0026ndash; Support Vector Machines\u003c/li\u003e\n \u003cli\u003eDT \u0026ndash; Decision Trees\u003c/li\u003e\n \u003cli\u003eRF \u0026ndash; Random Forest\u003c/li\u003e\n \u003cli\u003eAB \u0026ndash; Ada Boosting\u003c/li\u003e\n \u003cli\u003eGB \u0026ndash; Gradient Boosting\u003c/li\u003e\n\u003c/ul\u003e"},{"header":"Declarations","content":"\u003cp\u003e5 \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Authors\u0026rsquo; declaration\u003c/p\u003e\n\u003cp\u003eThis manuscript is the authors\u0026rsquo; original work and has not been published elsewhere. All authors have checked the manuscript and have agreed to this submission.\u003c/p\u003e\n\u003cp\u003e6\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Author Contributions\u003c/p\u003e\n\u003cp\u003ePragya Saxena: Writing \u0026ndash; original draft, Methodology, Review and editing; Arunkumar Bongale: Review, Formal Analysis, Conceptualization; Satish Kumar: Project administration, Formal Analysis Conceptualization, Results and discussions.; Rajesh Kodbal: Review and Formal Analysis.\u003c/p\u003e\n\u003cp\u003e7\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Funding\u003c/p\u003e\n\u003cp\u003eNo funding sources were provided for publishing the manuscript.\u003c/p\u003e"},{"header":"References ","content":"\u003col\u003e\n\u003cli\u003eG. K. Padhy, C. S. Wu, and S. Gao, \u0026ldquo;Friction stir based welding and processing technologies - processes, parameters, microstructures and applications: A review,\u0026rdquo; \u003cem\u003eJ. Mater. Sci. Technol.\u003c/em\u003e, vol. 34, no. 1, pp. 1\u0026ndash;38, 2018, doi: 10.1016/j.jmst.2017.11.029.\u003c/li\u003e\n\u003cli\u003eV. C. Kale, \u0026ldquo;Aluminium Based Metal Matrix Composites for Aerospace Application: A Literature Review,\u0026rdquo; \u003cem\u003eIOSR J. Mech. Civ. Eng.\u003c/em\u003e, vol. 12, no. 6, pp. 2278\u0026ndash;1684, 2015, doi: 10.9790/1684-12653136.\u003c/li\u003e\n\u003cli\u003eA. Heidarzadeh \u003cem\u003eet al.\u003c/em\u003e, \u003cem\u003eFriction stir welding/processing of metals and alloys: A comprehensive review on microstructural evolution\u003c/em\u003e, vol. 117. Elsevier Ltd, 2021. doi: 10.1016/j.pmatsci.2020.100752.\u003c/li\u003e\n\u003cli\u003eS. A. Alidokht, A. Abdollah-zadeh, S. Soleymani, and H. Assadi, \u0026ldquo;Microstructure and tribological performance of an aluminium alloy based hybrid composite produced by friction stir processing,\u0026rdquo; \u003cem\u003eMater. Des.\u003c/em\u003e, vol. 32, no. 5, pp. 2727\u0026ndash;2733, 2011, doi: 10.1016/j.matdes.2011.01.021.\u003c/li\u003e\n\u003cli\u003eM. M. Tash and E. R. I. Mahmoud, \u0026ldquo;Development of in-Situ Al-Si/CuAl2 metal matrix composites: Microstructure, hardness, and wear behavior,\u0026rdquo; \u003cem\u003eMaterials (Basel).\u003c/em\u003e, vol. 9, no. 6, 2016, doi: 10.3390/ma9060442.\u003c/li\u003e\n\u003cli\u003eA. H. Elsheikh, \u0026ldquo;Applications of machine learning in friction stir welding: Prediction of joint properties, real-time control and tool failure diagnosis,\u0026rdquo; \u003cem\u003eEng. Appl. Artif. Intell.\u003c/em\u003e, vol. 121, no. June 2022, p. 105961, 2023, doi: 10.1016/j.engappai.2023.105961.\u003c/li\u003e\n\u003cli\u003eM. Sajid, A. K. Malik, M. Tanveer, and P. N. Suganthan, \u0026ldquo;Neuro-Fuzzy Random Vector Functional Link Neural Network for Classification and Regression Problems,\u0026rdquo; \u003cem\u003eIEEE Trans. Fuzzy Syst.\u003c/em\u003e, vol. 32, no. 5, pp. 2738\u0026ndash;2749, 2024, doi: 10.1109/TFUZZ.2024.3359652.\u003c/li\u003e\n\u003cli\u003eS. Thapliyal and A. Mishra, \u0026ldquo;Machine learning classification-based approach for mechanical properties of friction stir welding of copper,\u0026rdquo; \u003cem\u003eManuf. Lett.\u003c/em\u003e, vol. 29, pp. 52\u0026ndash;55, 2021, doi: 10.1016/j.mfglet.2021.05.010.\u003c/li\u003e\n\u003cli\u003eB. Das, S. Pal, and S. Bag, \u0026ldquo;Torque based defect detection and weld quality modelling in friction stir welding process,\u0026rdquo; \u003cem\u003eJ. Manuf. Process.\u003c/em\u003e, vol. 27, pp. 8\u0026ndash;17, 2017, doi: 10.1016/j.jmapro.2017.03.012.\u003c/li\u003e\n\u003cli\u003eS. Yoon \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Defect detection in composites by deep learning using solitary waves,\u0026rdquo; \u003cem\u003eInt. J. Mech. Sci.\u003c/em\u003e, vol. 239, no. October 2022, p. 107882, 2023, doi: 10.1016/j.ijmecsci.2022.107882.\u003c/li\u003e\n\u003cli\u003eY. Chen, A. Ebrahimifakhar, Y. Hu, and D. P. Yuill, \u0026ldquo;Generalizability of machine learning-based fault classification for residential air-conditioners,\u0026rdquo; \u003cem\u003eEnergy Build.\u003c/em\u003e, vol. 295, no. June, p. 113263, 2023, doi: 10.1016/j.enbuild.2023.113263.\u003c/li\u003e\n\u003cli\u003eH. Moosaei and M. Hlad\u0026iacute;k, \u0026ldquo;Sparse solution of least-squares twin multi-class support vector machine using ℓ0 and ℓp-norm for classification and feature selection,\u0026rdquo; \u003cem\u003eNeural Networks\u003c/em\u003e, vol. 166, pp. 471\u0026ndash;486, 2023, doi: 10.1016/j.neunet.2023.07.039.\u003c/li\u003e\n\u003cli\u003eA. Naji Hussain, S. A. Abboud, B. A. baki Jumaa, and M. N. Abdullah, \u0026ldquo;Impact of feature reduction techniques on classification accuracy of machine learning techniques in leg rehabilitation,\u0026rdquo; \u003cem\u003eMeas. Sensors\u003c/em\u003e, vol. 25, no. September 2022, p. 100544, 2023, doi: 10.1016/j.measen.2022.100544.\u003c/li\u003e\n\u003cli\u003eW. Guan \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Detection of tunnel defects in friction stir welded aluminum alloy joints based on the in-situ force signal,\u0026rdquo; \u003cem\u003eJ. Manuf. Process.\u003c/em\u003e, vol. 71, no. July, pp. 1\u0026ndash;11, 2021, doi: 10.1016/j.jmapro.2021.09.014.\u003c/li\u003e\n\u003cli\u003eZ. Li, H. Zhao, X. Zhang, J. Dong, L. Hu, and H. Gao, \u0026ldquo;Multi-parameter sensing of robotic friction stir welding based on laser circular scanning,\u0026rdquo; \u003cem\u003eJ. Manuf. Process.\u003c/em\u003e, vol. 89, no. 92, pp. 92\u0026ndash;104, 2023, doi: 10.1016/j.jmapro.2023.01.060.\u003c/li\u003e\n\u003cli\u003eS. Sharda \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;A hybrid machine learning technique for feature optimization in object-based classification of debris-covered glaciers,\u0026rdquo; \u003cem\u003eAin Shams Eng. J.\u003c/em\u003e, vol. 13, no. 6, p. 101809, 2022, doi: 10.1016/j.asej.2022.101809.\u003c/li\u003e\n\u003cli\u003eS. Sayyad, S. Kumar, A. Bongale, K. Kotecha, G. Selvachandran, and P. N. Suganthan, \u0026ldquo;Tool wear prediction using long short-term memory variants and hybrid feature selection techniques,\u0026rdquo; \u003cem\u003eInt. J. Adv. Manuf. Technol.\u003c/em\u003e, vol. 121, no. 9\u0026ndash;10, pp. 6611\u0026ndash;6633, 2022, doi: 10.1007/s00170-022-09784-y.\u003c/li\u003e\n\u003cli\u003eS. Sayyad, S. Kumar, A. Bongale, K. Kotecha, and A. Abraham, \u0026ldquo;Remaining Useful-Life Prediction of the Milling Cutting Tool Using Time\u0026ndash;Frequency-Based Features and Deep Learning Models,\u0026rdquo; \u003cem\u003eSensors\u003c/em\u003e, vol. 23, no. 12, 2023, doi: 10.3390/s23125659.\u003c/li\u003e\n\u003cli\u003eA. E. Ragab \u003cem\u003eet al.\u003c/em\u003e, \u0026ldquo;Open-source real-time monitoring system of temperature and force during friction stir spot welding,\u0026rdquo; \u003cem\u003eJ. Eng. Res.\u003c/em\u003e, no. July, 2023, doi: 10.1016/j.jer.2023.08.020.\u003c/li\u003e\n\u003cli\u003eS. Gudadhe, A. Thakare, and A. M. Anter, \u0026ldquo;A novel machine learning-based feature extraction method for classifying intracranial hemorrhage computed tomography images,\u0026rdquo; \u003cem\u003eHealthc. Anal.\u003c/em\u003e, vol. 3, no. April, p. 100196, 2023, doi: 10.1016/j.health.2023.100196.\u003c/li\u003e\n\u003cli\u003eK. Balachandar, R. Jegadeeshwaran, J. Lakshmipathi, and D. Saravanakumar, \u0026ldquo;Friction Stir Welding Tool Condition Prediction Using Vibrational Analysis through Machine Learning - A Review,\u0026rdquo; \u003cem\u003eJ. Phys. Conf. Ser.\u003c/em\u003e, vol. 1969, no. 1, 2021, doi: 10.1088/1742-6596/1969/1/012051.\u003c/li\u003e\n\u003cli\u003eB. Chaudhary, N. K. Jain, J. Murugesan, and V. Patel, \u0026ldquo;Friction stir powder additive manufacturing of Al 6061 alloy: Enhancing microstructure and mechanical properties by reducing thermal gradient,\u0026rdquo; \u003cem\u003eJ. Mater. Res. Technol.\u003c/em\u003e, vol. 26, pp. 1168\u0026ndash;1184, 2023, doi: 10.1016/j.jmrt.2023.07.270.\u003c/li\u003e\n\u003cli\u003eP. Saxena, A. Bongale, S. Kumar, and P. Jadhav, \u0026ldquo;Investigation of microstructural and wear behavior of Al6061 surface composites fabricated by friction stir process using Taguchi approach,\u0026rdquo; \u003cem\u003eMater. Res. Express\u003c/em\u003e, vol. 9, no. 1, 2022, doi: 10.1088/2053-1591/ac4a2d.\u003c/li\u003e\n\u003cli\u003eP. Saxena, A. Bongale, S. Kumar, and P. Jadhav, \u0026ldquo;Microstructural and sensor data analysis of friction stir processing in fabricating Al6061 surface composites,\u0026rdquo; \u003cem\u003eEng. Res. Express\u003c/em\u003e, vol. 5, no. 1, p. 015065, 2023, doi: 10.1088/2631-8695/acc158.\u003c/li\u003e\n\u003cli\u003eY. Y. Lim and C. K. Soh, \u0026ldquo;Effect of varying axial load under fixed boundary condition on admittance signatures of electromechanical impedance technique,\u0026rdquo; \u003cem\u003eJ. Intell. Mater. Syst. Struct.\u003c/em\u003e, vol. 23, no. 7, pp. 815\u0026ndash;826, 2012, doi: 10.1177/1045389X12437888.\u003c/li\u003e\n\u003cli\u003eM. Shah, V. Vakharia, R. Chaudhari, J. Vora, D. Y. Pimenov, and K. Giasin, \u0026ldquo;Tool wear prediction in face milling of stainless steel using singular generative adversarial network and LSTM deep learning models,\u0026rdquo; \u003cem\u003eInt. J. Adv. Manuf. Technol.\u003c/em\u003e, vol. 121, no. 1\u0026ndash;2, pp. 723\u0026ndash;736, 2022, doi: 10.1007/s00170-022-09356-0.\u003c/li\u003e\n\u003cli\u003eM. Liu, H. Li, H. Zhou, H. Zhang, and G. Huang, \u0026ldquo;Development of machine learning methods for mechanical problems associated with fibre composite materials: A review,\u0026rdquo; \u003cem\u003eCompos. Commun.\u003c/em\u003e, vol. 49, p. 101988, 2024, doi: https://doi.org/10.1016/j.coco.2024.101988.\u003c/li\u003e\n\u003cli\u003eL. Yang and A. Shami, \u0026ldquo;On hyperparameter optimization of machine learning algorithms: Theory and practice,\u0026rdquo; \u003cem\u003eNeurocomputing\u003c/em\u003e, vol. 415, pp. 295\u0026ndash;316, 2020, doi: https://doi.org/10.1016/j.neucom.2020.07.061.\u003c/li\u003e\n\u003cli\u003eS. R. Anvari, F. Karimzadeh, and M. H. Enayati, \u0026ldquo;Wear characteristics of Al-Cr-O surface nano-composite layer fabricated on Al6061 plate by friction stir processing,\u0026rdquo; \u003cem\u003eWear\u003c/em\u003e, vol. 304, no. 1\u0026ndash;2, pp. 144\u0026ndash;151, 2013, doi: 10.1016/j.wear.2013.03.014.\u003c/li\u003e\n\u003cli\u003eO. Iparraguirre-Villanueva, K. Espinola-Linares, R. O. Flores Casta\u0026ntilde;eda, and M. Cabanillas-Carbonell, \u0026ldquo;Application of Machine Learning Models for Early Detection and Accurate Classification of Type 2 Diabetes,\u0026rdquo; \u003cem\u003eDiagnostics\u003c/em\u003e, vol. 13, no. 14, 2023, doi: 10.3390/diagnostics13142383.\u003c/li\u003e\n\u003cli\u003eS. Verma, M. Gupta, and J. P. Misra, \u0026ldquo;Performance evaluation of friction stir welding using machine learning approaches,\u0026rdquo; \u003cem\u003eMethodsX\u003c/em\u003e, vol. 5, pp. 1048\u0026ndash;1058, 2018, doi: 10.1016/j.mex.2018.09.002.\u003c/li\u003e\n\u003cli\u003eB. A. Tama and S. Lim, \u0026ldquo;Ensemble learning for intrusion detection systems: A systematic mapping study and cross-benchmark evaluation,\u0026rdquo; \u003cem\u003eComput. Sci. Rev.\u003c/em\u003e, vol. 39, p. 100357, 2021, doi: https://doi.org/10.1016/j.cosrev.2020.100357.\u003c/li\u003e\n\u003cli\u003eA. Taha and O. Barukab, \u0026ldquo;Android Malware Classification Using Optimized Ensemble Learning Based on Genetic Algorithms,\u0026rdquo; \u003cem\u003eSustainability\u003c/em\u003e, vol. 14, no. 21, 2022, doi: 10.3390/su142114406.\u003c/li\u003e\n\u003cli\u003eU. V Nikam and V. M. Deshmuh, \u0026ldquo;Performance Evaluation of Machine Learning Classifiers in Malware Detection,\u0026rdquo; in \u003cem\u003e2022 IEEE International Conference on Distributed Computing and Electrical Circuits and Electronics (ICDCECE)\u003c/em\u003e, 2022, pp. 1\u0026ndash;5. doi: 10.1109/ICDCECE53908.2022.9793102.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"journal-of-the-brazilian-society-of-mechanical-sciences-and-engineering","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmse","sideBox":"Learn more about [Journal of the Brazilian Society of Mechanical Sciences and Engineering](http://link.springer.com/journal/40430)","snPcode":"40430","submissionUrl":"https://www.editorialmanager.com/bmse/default2.aspx","title":"Journal of the Brazilian Society of Mechanical Sciences and Engineering","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"composites, defects, feature selection, sensor data, machine learning classification","lastPublishedDoi":"10.21203/rs.3.rs-4834721/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4834721/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAluminium alloy based surface composites with hard reinforcement particles have wide scope in aerospace and automobile manufacturing industries. In this paper, the aluminium composites, manufactured by friction stir processing (FSP) with varying parameters are investigated for the faults occurred during fabrication process. It explores a machine-learning approach to detect defects of surface hybrid composites with an Al6061 alloy matrix, reinforced with copper and graphene nano-powders, using friction stir processing and a tungsten carbide tool on a milling machine. Multi-sensor time series data (vibration, force, and current) collected during fabrication, is preprocessed and labelled with normal and defective categories (e.g., pin break, brazing break, rough surface, no composite) using visual inspection. The important time domain and frequency domain features are extracted using different libraries in python. Thenafter, various types of feature selection techniques, viz filter, wrapper and embedded methods are implemented to select most relevant features. The selected subset of features from all selection methods used, are applied to different machine learning and ensemble learning classifiers and their performances are evaluated. The optimal combinations of the type of feature selection method and classifier used, are obtained for efficient classification of surface defects in composited formed by FSP. The real time monitoring and defect detection system can be developed in future for the composites developed by FSP using the developed models.\u003c/p\u003e","manuscriptTitle":"Classification of faults in friction stir processed composites using a machine learning and ensemble learning approach","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-09-18 20:15:30","doi":"10.21203/rs.3.rs-4834721/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"","date":"2024-08-21T04:20:47+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-08-20T18:04:52+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-08-05T06:47:48+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of the Brazilian Society of Mechanical Sciences and Engineering","date":"2024-07-31T06:33:15+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-the-brazilian-society-of-mechanical-sciences-and-engineering","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmse","sideBox":"Learn more about [Journal of the Brazilian Society of Mechanical Sciences and Engineering](http://link.springer.com/journal/40430)","snPcode":"40430","submissionUrl":"https://www.editorialmanager.com/bmse/default2.aspx","title":"Journal of the Brazilian Society of Mechanical Sciences and Engineering","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"31c3692b-602e-4a8c-98dc-3731bc1e809a","owner":[],"postedDate":"September 18th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2024-09-18T20:15:30+00:00","versionOfRecord":[],"versionCreatedAt":"2024-09-18 20:15:30","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4834721","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4834721","identity":"rs-4834721","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}