Deep Transfer Learning and Data Augmentation for Automatic Flank Wear Detection and Classification

Deep Transfer Learning and Data Augmentation for Automatic Flank Wear Detection and Classification | 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 Deep Transfer Learning and Data Augmentation for Automatic Flank Wear Detection and Classification Mohamed ELBAH, Oussama DAHMOUNE, Ikhlas MEDDOUR This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7189226/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Flank wear in cutting tools remains a critical issue in the metal cutting industry due to its direct impact on dimensional precision, surface integrity, and overall manufacturing efficiency. Failure to detect or accurately classify wear may result in premature tool replacement or extended use of worn tools, leading to increased operational costs, heat accumulation, machining vibrations, and potential damage to workpieces and equipment. This study provides a comparative analysis of four state-of-the-art deep convolutional neural network (CNN) architectures VGG16, GoogleNet, DenseNet121, and MobileNetV2 for the automatic classification of flank wear across four predefined severity levels. To overcome dataset limitations and improve model adaptability, transfer learning was employed by fine-tuning pre-trained models on a domain-specific dataset of tool wear images captured using a DM9 industrial digital microscope under dry turning conditions. A comprehensive data augmentation strategy was also applied to enhance model generalization and mitigate overfitting. Experimental results show that GoogleNet and DenseNet121 achieved the highest classification accuracy of 96.80%, with GoogleNet being favored for its shorter training time and computational efficiency. These findings underscore the effectiveness of deep learning-based approaches for reliable, scalable, and real-time tool condition monitoring in advanced manufacturing environments Flank Wear Turning process Deeplearning Transfer learning GoogleNet DenseNet12 MobileNetV2 Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Highlights A CNN-based system was developed to classify flank wear into four severity levels. Transfer learning improved model accuracy using pre-trained architectures. Data augmentation reduced overfitting and enhanced generalization performance. GoogleNet and DenseNet121 achieved the best accuracy, reaching 96.80%. GoogleNet was preferred due to faster training time and stable performance. I. INTRODUCTION In machining processes, achieving a superior surface finish is critical and is largely influenced by the condition of the cutting tool insert [ 1 ]. However, various factors such as wear, thermal fatigue, and corrosion [ 2 , 3 ] can adversely in machining processes, achieving a superior surface finish and high dimensional accuracy is a critical performance criterion, which depends largely on the condition of the cutting tool insert [ 1 ]. However, this condition is subject to degradation due to various mechanisms such as tool wear, thermal fatigue, and corrosion [ 2 , 3 ]. Among these, tool wear has the most pronounced impact, as it directly influences tool life, machining quality, productivity, and overall operational costs. Consequently, real-time monitoring of tool wear is essential to maintain manufacturing efficiency and ensure product quality. Tool wear is typically assessed using either direct methods (such as visual inspection or image acquisition) or indirect methods (such as acoustic emission, vibration, and current signal analysis) [ 4 ]. In recent years, Artificial Intelligence (AI) technologies—particularly Convolutional Neural Networks (CNNs)—have demonstrated strong potential for automating the classification of different wear types. CNNs have proven especially effective in recognizing common tool wear forms such as flank wear, chipping, spalling, and built-up edge, which primarily occur at the flank and tip regions of the cutting insert [ 5 – 8 ]. Notable contributions in the literature reinforce the relevance of AI in tool wear monitoring. For example, Khoury Junior et al. [ 9 ] developed an image-based classification system using Bayesian discriminant functions combined with the Hough Transform, achieving 94.3% accuracy in detecting flank wear (VB). Similarly, Miao et al. [ 10 ] proposed a U-Net-based CNN model to classify VB wear in CNC cutting tools. By augmenting their dataset from 186 to 1,100 images, they significantly improved the robustness and accuracy of their wear detection system. Complementary studies further support these findings. Panda et al. [ 11 ] compared the performance of Artificial Neural Networks (ANNs) and Response Surface Methodology (RSM) for predicting flank wear (VBc) during the turning of AISI 4340 steel. Their experiments, designed using a Taguchi L16 orthogonal array, revealed that ANN outperformed regression models, achieving a remarkably low average error of 2.02%. Additionally, Agrawal et al. [ 12 ] conducted a comprehensive review of AI-based approaches for tool wear monitoring in turning operations. Their findings emphasized the effectiveness of CNNs for real-time and precise detection, while also highlighting persistent challenges such as the variability in wear progression and the limited integration of CNC machine data in existing models. Brili et al. [ 13 ] proposed a thermal imaging approach combined with CNNs to classify tool wear states into “GO” or “NO-GO” conditions, achieving 99.55% accuracy in turning processes. Sun et al [ 14 ] applied a PCNN-based monitoring system to classify various tool defects—flank wear, built-up edge, chipping, and cracks—using CNC turning data. Similarly, Lim et al. [ 15 ] analyzed machined surface profiles using deep neural models, demonstrating the superiority of CNNs over DNNs, achieving 98.9% accuracy while minimizing RMSE.In another study, Kumar et al. [ 16 ] developed a CNN-based system for flank wear classification using raw and minimally preprocessed images, reaching 96.3% and 99.9% accuracy, respectively. Ambadekar et al. [ 17 ] employed a ResNet-50 CNN to estimate the remaining life of cutting tools by categorizing VB wear into severity levels (A, B, and C). In a more adaptive approach, Dahmoune et al. [ 18 ] combined CNNs with Case-Based Reasoning (CBR) to develop a hybrid Tool Condition Monitoring (TCM) system, achieving 98% training accuracy and 96% validation accuracy. Despite notable advancements in deep learning-based tool condition monitoring systems, several critical limitations persist. Many existing approaches exhibit poor generalization due to limited dataset sizes, class imbalance, and insufficient adaptation to real-world industrial environments. Conventional CNN-based frameworks are frequently applied in a generic manner, lacking customization for the specific wear patterns and operational constraints encountered in machining processes. Furthermore, the integration of human validation or incremental learning strategies remains rare, which restricts the system’s adaptability in ambiguous or transitional wear scenarios. Recent studies, including the one cited in [ 19 ], have highlighted these shortcomings, emphasizing the urgent need for more robust, interpretable, and industry-ready AI solutions. These works underscore the effectiveness of advanced image processing and deep convolutional neural networks (CNNs) in improving classification accuracy, while simultaneously pointing out challenges related to data variability and limited integration within CNC environments ,In response to these challenges, the present study proposes a comprehensive system for the automatic classification of flank wear using four well-established CNN architectures: VGG16, GoogleNet (Inception), DenseNet121, and MobileNetV2. Transfer learning is employed to leverage the representational power of pre-trained models, thereby reducing the dependence on large annotated datasets. In addition, custom-designed data augmentation techniques are applied to enhance generalization and mitigate overfitting. The key scientific contributions of this work include the development of a domain-specific CNN-based classification pipeline and a thorough performance evaluation using metrics such as accuracy, precision, recall, and F1-score, all validated under realistic dry turning industrial conditions. This paper is structured as follows. The next section reviews the relevant CNN architectures and related literature. This is followed by a detailed description of the experimental setup, dataset composition, and the implemented methodology. The results are then presented and analyzed, concluding with a discussion of future research directions. II. EXPERIMENTS AND METHODOLOGY This section presents the detailed methodologies employed to evaluate the effectiveness of the proposed automated system for cutting tool flank wear classification and detection. A carefully curated dataset, specifically designed for this purpose, is utilized. The classification models are trained on a designated training subset of this dataset and subsequently evaluated using unseen test images. The research methodology comprises two key phases: Experimental phase: This phase focuses on conducting turning operations on a lathe machine to acquire a comprehensive dataset of tool wear images. Data processing and model training/testing: This phase encompasses processing the acquired data, training the deep learning models using the training dataset, and evaluating their performance on the unseen test images. Ⅱ.1 Experimental set-up Turning operations were carried out on a conventional SN40 lathe (TOSTRENCIN), as illustrated in Fig. 1 a. The workpiece material, shown in Fig. 1 c, consisted of AISI 1045 steel a widely utilized medium carbon steel in industrial machining applications. Its detailed chemical composition, presented in Table 1 , was analyzed in the laboratory of the national company ORSIM using specialized, high-precision analytical equipment to ensure compositional accuracy and process reliability. The cylindrical workpiece initially measured 77 mm in diameter and 115 mm in machining length. Tool wear monitoring was performed using a DM9 industrial-grade digital microscope (Fig. 1 b), equipped with high-resolution optical lenses and a calibrated LED illumination system. This setup allowed for consistent, glare-free visualization of the cutting tool’s flank surface. The microscope captured images at a resolution of 1280 × 720 pixels and transmitted them to a connected computer, enabling high-fidelity data acquisition suitable for automated wear analysis. Based on manufacturer specifications and experimental calibration, the measurement uncertainty for flank wear width (VB) using the DM9 system was estimated at ± 10 µm well within the precision range recomended by the ISO 3685 standard for tool wear assessment in turning processes. A standard PSBNR 2525 M12 tool holder was used to secure an SNMG 120408 carbide insert. This insert geometry, featuring eight cutting edges, was selected based on its suitability for turning AISI1045 steel in its normalized condition. Figure 1 (d) highlights the insert's mounting within the tool holder via a central hole to ensure stability during turning. Table 1 Chemical composition of the workpiece material. Material Chemical composition % C S Mn P Si AISI1045 0.51 0.031 0.68 0.029 0.36 Turning experiments were structured using an L4(2³) orthogonal array to methodically vary three critical input parameters: cutting speed, feed rate, and depth of cut, each at two discrete levels, in accordance with the robust design principles established by Taguchi et al. [ 20 ]. This orthogonal approach allows for efficient evaluation of the main effects of each factor while significantly reducing the number of experimental trials required. For Table 2 , the selection of factors and their corresponding levels was based on prior authoritative studies (such as Sandvik tooling guidelines and ISO 3685 standards) and was further validated through feasibility tests conducted in the laboratory using an SN40 lathe and SNMG insert. These choices are fully justified in the experimental methodology section. Table 2 presents the final levels selected for each machining parameter, and Table 3 details the specific combinations applied in each experimental run. The parameters were ultimately aligned with the tool manufacturer’s recommended conditions for machining AISI 1045 steel, ensuring both experimental relevance and industrial applicability. Table 2 Assignment of the levels to the factors. No Parameter Unit Level -1 + 1 1 Cutting Speed m/min 140 180 2 Feed Rate mm/rev 0.08 0.16 3 Depth of Cut mm 0.2 0.4 Table 3 Design matrix as per L4 orthogonal array. Experiment Number Machining parameters Cutting speed (m/min) Feed Rate (mm/rev) Depth of cut (mm) 1 140 0.08 0.2 2 140 0.12 0.4 3 180 0.08 0.2 4 180 0.12 0.4 Ⅱ.2 Flank wear and tool life There is a direct correlation between machining time and the rate of flank wear progression [ 21 ]. This relationship is a key factor in determining tool life, as excessive wear can lead to poor surface finish, dimensional inaccuracies, and ultimately, tool failure. ASTM Standard guidelines provide a framework for evaluating tool longevity, often assessed by the extent of flank wear, referred to as VB, or the width of the worn edge. As illustrated in Fig. 2 , VB is typically measured in zone B of the cutting edge, which extends from the corner radius to the point where the width of the undeformed chip is b/4. When wear progresses uniformly across the tool, VB typically reaches 0.3 mm as a critical threshold. This critical value is often used as a criterion for tool replacement, ensuring optimal machining performance and minimizing downtime. III. DATASET AND MODEL ARCHITECTURES The implementation was carried out in the Google Colab runtime environment, utilizing libraries available within this platform and employing Python 3 as operating environment. To accelerate the computationally intensive tasks of machine learning and deep learning, the research utilized T4 GPUs, which provide significantly faster processing speeds compared to CPUs. This section describes the dataset used for training and evaluating deep learning models for classifying cutting tool flank wear, including its creation process, categories, and employed data augmentation techniques. It also introduces different deep convolutional neural network architectures used to monitor flank wear by applying transfer learning techniques to extract visual features for classification using a SoftMax layer. Ⅲ.1 Data Set The dataset comprises images of cutting tools classified into four distinct flank wear categories, as illustrated in Table 4 , corresponding to progressive stages of tool degradation. This classification framework is grounded in the conventional tool wear progression curve, which includes the initial wear phase, the steady-state phase, and the accelerated wear phase leading to tool failure. To enhance diagnostic granularity and support proactive maintenance strategies, an intermediate wear class was deliberately introduced between the steady and critical stages. This additional category enables more accurate detection of wear evolution and provides operators with a critical window for intervention before catastrophic failure occurs. The proposed four-class system significantly improves diagnostic precision, supports real-time decision-making, and aligns with the stringent requirements of modern industrial monitoring and predictive maintenance systems. Table 4 Flank wear classification based on VB measurement. Categories Limit values of the flank wear level Class 1 0 µm < VB ≤ 100 µm Class 2 100 µm < VB ≤ 180 µm Class 3 180 µm < VB ≤ 270 µm Class 4 270 µm < VB Figure 3 illustrates the composition of the image dataset before augmentation. It consists of 480 images categorized into four distinct classes, with each class represented by a specific VB range. The figure also showcases sample images from each class, highlighting the visual differences in flank wear progression. To overcome the limited size of the image dataset, data augmentation techniques were employed to synthetically expand it. Data augmentation not only increases the quantity of images but also enhances the generalization ability of the model, helping to reduce overfitting [ 23 ]. This approach, illustrated in Fig. 4 , involves applying transformations such as brightness adjustment, contrast modification, shearing, and zooming to the original images. By generating diverse variations of the original images, a richer dataset for training the models is created. This method ensures that the models are robust and better equipped to generalize to new and unseen examples of flank wear [ 24 ]. Ⅲ.2 Flank Wear Classification and Detection After extracting visual features from the target dataset through transfer learning, classification and detection tasks were conducted. The dataset is divided into training and testing sets, with each set containing images from all four wear intensity categories. Specifically, 80% of all images were allocated to the training set, and 20% were assigned to the test set, as illustrated in Table 5. Flank wear classification of the cutting tool was performed using a Softmax layer as the final activation function within the transfer learning model. TABLE.5 Distribution of training and testing images in each flank wear category. Light wear (Class1) Moderate wear (Class 2) Heavy wear (Class 3) Extreme wear (Class 4) Training set 277 273 294 270 Test set 69 68 74 67 Total 346 341 368 337 Ⅲ.3 CNN-based architectures CNN or ConvNet stands for convolutional neural network, designed to automatically extract significant visual patterns from raw image pixels with minimal pre-processing. To leverage the power of these networks for flank wear prediction, a transfer learning approach was adopted, utilizing pre-trained models that have already learned general image features. This study investigates four popular convolutional neural network (CNN) architectures: VGG16, GoogleNet (Inception), DenseNet121, and MobileNetV2. Ⅲ.3.1VGG16 VGG16 [ 25 ] is a convolutional neural network architecture developed by the Visual Geometry Group (VGG) at the University of Oxford in 2014. It features a deep structure comprising 16 layers, including 13 convolutional layers and 3 fully connected layers. VGG16 has demonstrated significant success in various computer vision tasks, including image classification and object detection. Trained on the ImageNet dataset, VGG16 serves as a widely adopted benchmark model in the deep learning domain, typically operating on input sizes of 224x224 pixels. Ⅲ.3.2 GoogleNet (Inception) GoogleNet [ 26 ], also referred to as Inception, is a CNN architecture devised by Google researchers, acclaimed as the winner of the 2014 ImageNet (sitiha) Large-Scale Visual Recognition Challenge (ILSVRC). Characterized by its deep structure and efficient utilization of computational resources, GoogleNet introduces inception modules enabling parallel processing at multiple scales. This architecture effectively captures both local features and global information, rendering it highly accurate and efficient for various computer vision tasks. Ⅲ.3.3 DenseNet121 DenseNet121 [ 27 ], proposed by researchers at the Computer Vision Center (CVC) and the Autonomous University of Barcelona in 2016, is part of the DenseNet family of models. DenseNet, denoting Dense Convolutional Network, features densely connected layers, with each layer linked to every other layer in a feed-forward manner. This architecture promotes feature reuse and facilitates gradient flow, thereby enhancing learning efficiency and model performance. DenseNet121 finds widespread application in image classification, object detection, and semantic segmentation tasks. Ⅲ.3.4 MobileNetV2 MobileNetV2 [ 28 ] is a convolutional neural network architecture introduced by Google in 2018, specifically designed for mobile and embedded vision applications. It builds upon the original MobileNet architecture with improvements aimed at efficiency and performance. MobileNetV2 features inverted residual blocks with linear bottleneck layers, enabling deeper networks while maintaining low computational cost and memory footprint. This architecture leverages depth wise separable convolutions to reduce computation and parameter size, making it suitable for resource-constrained environments such as smartphones and IoT devices. Trained on large-scale datasets like ImageNet, MobileNetV2 achieves competitive accuracy in tasks like image classification, object detection, and semantic segmentation, making it a popular choice for real-world mobile vision applications. Ⅲ.4 Evaluation metrics The performance of the VGG16, GoogleNet, DenseNet121, and MobileNetV2 algorithms was analyzed using common metrics: precision, recall, F1-score, and accuracy. These metrics are computed using equations (1) to (4) [ 29 ], where TP denotes true positives, TN represents true negatives, FP signifies false positives, and FN stands for false negatives. An effective classifier aims to optimize both precision and recall concurrently, ensuring precise classification into the correct categories. The F1-score directly reflects the balance between precision and recall, while accuracy offers a straightforward interpretation of classifier performance by indicating the ratio of correctly classified images to the total number of images evaluated. \(\:\mathbf{P}\mathbf{r}\mathbf{e}\mathbf{c}\mathbf{i}\mathbf{s}\mathbf{i}\mathbf{o}\mathbf{n}=\frac{\:\:\:\:\text{N}\text{o}.\:\text{o}\text{f}\:\text{c}\text{o}\text{r}\text{r}\text{e}\text{c}\text{t}\:\text{c}\text{l}\text{a}\text{s}\text{s}\text{i}\text{f}\text{i}\text{c}\text{a}\text{t}\text{i}\text{o}\text{n}\text{s}\:\text{i}\text{n}\text{t}\text{o}\:\text{t}\text{h}\text{e}\:\text{j}\:\text{c}\text{l}\text{a}\text{s}\text{s}}{\text{N}\text{o}.\:\text{o}\text{f}\:\text{a}\text{l}\text{l}\:\text{c}\text{l}\text{a}\text{s}\text{s}\text{i}\text{f}\text{i}\text{c}\text{a}\text{t}\text{i}\text{o}\text{n}\text{s}\:\text{i}\text{n}\text{t}\text{o}\:\text{t}\text{h}\text{e}\:\text{j}\:\text{c}\text{l}\text{a}\text{s}\text{s}}\:=\frac{\mathbf{T}\mathbf{P}}{\mathbf{T}\mathbf{P}\:+\:\mathbf{F}\mathbf{P}}\) (1) \(\:\mathbf{R}\mathbf{e}\mathbf{c}\mathbf{a}\mathbf{l}\mathbf{l}=\frac{\text{N}\text{o}.\:\text{o}\text{f}\:\text{c}\text{o}\text{r}\text{r}\text{e}\text{c}\text{t}\:\text{c}\text{l}\text{a}\text{s}\text{s}\text{i}\text{f}\text{i}\text{c}\text{a}\text{t}\text{i}\text{o}\text{n}\text{s}\:\text{i}\text{n}\text{t}\text{o}\:\text{t}\text{h}\text{e}\:\text{j}\:\text{c}\text{l}\text{a}\text{s}\text{s}}{\text{N}\text{o}.\:\text{o}\text{f}\:\text{a}\text{c}\text{t}\text{u}\text{a}\text{l}\:\text{i}\text{n}\text{s}\text{t}\text{a}\text{n}\text{c}\text{e}\text{s}\:\text{i}\text{n}\text{t}\text{o}\:\text{t}\text{h}\text{e}\:\text{j}\:\text{c}\text{l}\text{a}\text{s}\text{s}}=\frac{\mathbf{T}\mathbf{P}}{\mathbf{T}\mathbf{P}\:+\:\mathbf{F}\mathbf{N}}\) (2) \(\:\mathbf{F}{1}_{-}\mathbf{S}\mathbf{c}\mathbf{o}\mathbf{r}\mathbf{e}=2\times\:\frac{\mathbf{P}\mathbf{r}\mathbf{e}\mathbf{c}\mathbf{i}\mathbf{s}\mathbf{i}\mathbf{o}\mathbf{n}\times\:\mathbf{R}\mathbf{e}\mathbf{c}\mathbf{a}\mathbf{l}\mathbf{l}}{\mathbf{P}\mathbf{r}\mathbf{e}\mathbf{c}\mathbf{i}\mathbf{s}\mathbf{i}\mathbf{o}\mathbf{n}+\mathbf{R}\mathbf{e}\mathbf{c}\mathbf{a}\mathbf{l}}\) (3) \(\:\mathbf{A}\mathbf{c}\mathbf{c}\mathbf{u}\mathbf{r}\mathbf{a}\mathbf{c}\mathbf{y}=\frac{\text{N}\text{o}.\:\text{o}\text{f}\:\text{a}\text{l}\text{l}\:\text{c}\text{o}\text{r}\text{r}\text{e}\text{c}\text{t}\:\text{c}\text{l}\text{a}\text{s}\text{s}\text{i}\text{f}\text{i}\text{c}\text{a}\text{t}\text{i}\text{o}\text{n}\text{s}}{\text{N}\text{o}.\:\text{o}\text{f}\:\text{a}\text{l}\text{l}\:\text{c}\text{l}\text{a}\text{s}\text{s}\text{i}\text{f}\text{i}\text{c}\text{a}\text{t}\text{i}\text{o}\text{n}\text{s}}=\frac{\sum\:_{\mathbf{i}=1}^{\mathbf{j}}{\mathbf{T}\mathbf{P}}_{\mathbf{i}}}{\mathbf{n}\mathbf{o}.\:\mathbf{o}\:\mathbf{f}\:\mathbf{a}\mathbf{l}\mathbf{l}\:\mathbf{c}\mathbf{l}\mathbf{a}\mathbf{s}\mathbf{s}\mathbf{i}\mathbf{f}\mathbf{i}\mathbf{c}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}\mathbf{s}}\) (4) IV. COMPARATIVE ANALYSIS, RESULTS AND DISCUSSION Ⅳ.1 Evaluating CNN Performance with Different Optimizers To optimize the performance of the CNN models (VGG16, GoogleNet, DenseNet121, and MobileNetV2), a grid search strategy was employed to tune hyperparameters. This involved systematically exploring different combinations. The optimizers evaluated were Adam, SGD, and RMSprop, chosen for their common use in deep learning and distinct optimization strategies. A learning rate and 10 epochs were identified as yielding the best results across the models. This combination struck a balance between achieving high accuracy and minimizing training time. The resulting performance metrics for each model and optimizer are presented in detail in Tables 6, 7, 8, and 9, allowing for a comprehensive comparison and analysis of the models' effectiveness. TABLE 6. Comparative Performance of CNN Architectures for Image (for VGG16) Model Optimizer Accuracy F1_Score Precision Recall Avg Epoch Time (s) Initial Epoch Time (s) VGG16 Adam 0.9537 0.9528 0.9542 0.9526 3 6 SGD 0.5487 0.4595 0.5234 0.5587 271 257 RMSprop 0.8773 0.8773 0.8934 0.8795 283 296 Table 6 presents a comparative performance analysis of three optimizers – Adam, SGD, and RMSprop – applied to the VGG16 model. Adam showcases superior performance across all metrics, achieving the highest accuracy (0.9537), F1 score (0.9528), precision (0.9542), and recall (0.9526). While RMSprop demonstrates moderate performance with an accuracy of 0.8773, SGD significantly lags behind with an accuracy of 0.5487. Notably, Adam also exhibits significantly faster training times, averaging 3 seconds per epoch, compared to 271 seconds for SGD and 283 seconds for RMSprop. Therefore, Adam emerges as the clear winner for VGG16 in this context, achieving high accuracy and requiring substantially less training time. TABLE 7. Comparative Performance of CNN Architectures for Image (for GoogleLeNet (Inception )) Model Optimizer Accuracy F1_Score Precision Recall Avg Epoch Time (s) Initial Epoch Time (s) GoogleNet (Inception) Adam 0.9680 0.9670 0.9673 0.9672 1 10 SGD 0.9675 0.9673 0.9684 0.9666 44 53 RMSprop 0.8267 0.7989 0.8774 0.8137 43 52 In Table7, a comparative evaluation is provided for three optimizers – Adam, SGD, and RMSprop – concerning their performance on the GoogLeNet (Inception) model. Adam and SGD showcase notably high accuracy, achieving 0.9680 and 0.9675, respectively. The F1 scores, precision, and recall values for both optimizers are strikingly similar, indicating strong overall performance. However, Adam demonstrates a significant advantage in training speed, averaging only 1 second per epoch compared to SGD's 44 seconds. RMSprop attains a respectable accuracy of 0.8267, but it is outperformed by both Adam and SGD, other performance metrics, and training speed (43 seconds per epoch). Therefore, for the GoogLeNet model, Adam emerges as the most efficient optimizer, providing comparable accuracy to SGD but with significantly faster training times. TABLE 8. Comparative Performance of CNN Architectures for Image (for DenseNet121 ) Model Optimizer Accuracy F1_Score Precision Recall Avg Epoch Time (s) Initial Epoch Time (s) DenseNet121 Adam 0.9680 0.9670 0.9680 0.9665 2 14 SGD 0.9242 0.9222 0.9231 0.9235 76 82 RMSprop 0.6679 0.6470 0.7940 0.6866 75 87 Table 8 provides a comparative assessment of three optimization algorithms — Adam, SGD, and RMSprop — applied to the DenseNet121 model. Adam achieves the highest accuracy (0.9680), F1 score (0.9670), precision (0.9680), and recall (0.9665), indicating superior overall performance. SGD follows with an accuracy of 0.9242, demonstrating commendable performance but trailing behind Adam. Conversely, RMSprop exhibits notably lower performance with an accuracy of 0.6679, suggesting potential inadequacy for this model. Regarding training time, Adam shows the shortest average epoch time (2 seconds) compared to SGD (76 seconds) and RMSprop (75 seconds). This analysis confirms Adam as the most efficient optimizer for DenseNet121, achieving the highest accuracy with the shortest training duration. TABLE 9. Comparative Performance of CNN Architectures for Image (for MobileNetV2 ) Model Optimizer Accuracy F1_Score Precision Recall Avg Epoch Time (s) Initial Epoch Time (s) MobileNetV2 Adam 0.9573 0.9565 0.9590 0.9581 1 7 SGD 0.2310 0.1014 0.1823 0.2542 18 22 RMSprop 0.7906 0.7744 0.8558 0.7834 19 23 Table 9 provides a comparative analysis of three optimization algorithms (Adam, SGD, and RMSprop) applied to the MobileNetV2 model. Adam emerges as the top performer among the optimizers, achieving a high accuracy of 0.9573, an F1 Score of 0.9565, a precision of 0.9590, and a recall of 0.9581. Notably, Adam also demonstrates the shortest training time, averaging just 1 second per epoch. In contrast, RMSprop exhibits moderate performance with an accuracy of 0.7906, while SGD lags significantly behind with an accuracy of only 0.2310, suggesting its inadequacy for this model. These results underscore Adam's effectiveness as the optimal optimizer for MobileNetV2, delivering superior accuracy and markedly faster training times compared to SGD and RMSprop. This comparison highlights the importance of empirically evaluating various optimization algorithms for CNN architectures and datasets. Across all four models (VGG16, GoogLeNet, DenseNet121, and MobileNetV2), the Adam optimizer consistently achieved the highest accuracy and fastest training times. Based on these results, the remaining evaluations of all four models were conducted using the Adam optimizer. This choice enables a consistent and efficient assessment of the models' performance on the flank wear classification task. Ⅳ.2 Analysing Confusion Matrices To gain a deeper understanding of each CNN model's performance, the confusion matrices generated during testing were examined. These matrices reveal both strengths and weaknesses by illustrating the model's ability to correctly classify instances within each category. All models exhibit strong overall performance, as indicated by the dominant diagonals in their confusion matrices. However, each model demonstrates a tendency to misclassify some instances between Class 2 and Class 3, suggesting a potential limitation in distinguishing these specific categories. This observation stems from the close proximity of these wear types in their VB (flank wear) value ranges. Specifically, moderate wear is characterized by VB values within the interval [100 µm – 180 µm], while severe wear falls within [180 µm – 270 µm]. As a result, when VB values approach the 180 µm threshold, it becomes a potential source of confusion between these two wear types. This phenomenon is illustrated in Figure 8. Ⅳ.3 Training and validation accuracy analysis Figure 9 displays the training and validation accuracy curves of our four convolutional neural network models across training epochs, offering valuable insights into their learning dynamics and generalization performance. While VGG16 achieves a respectable 0.9537 validation accuracy, its fluctuating curves hint indicating some level of overfitting, suggesting the model is learning the training data well, but it's not generalizing well to unseen data. In contrast, GoogLeNet (Inception) exhibits more stable performance, with both training and validation accuracy steadily increasing, indicating strong generalization ability. MobileNetV2, reaching a validation accuracy of 0.9573, demonstrates a similar learning pattern to DenseNet121, although with slightly lower overall performance. Both models show steady increases in training and validation accuracy, but a slight dip in validation accuracy near the end in both models suggest a degree of overfitting. Ⅳ.4 Model loss analysis during training Figure 10 visualizes the training and validation loss curves of four CNN models—VGG16, GoogleNet (Inception), DenseNet121, and MobileNetV2—over the training epochs, illustrating each model's training progression and performance. VGG16 exhibits a steady decrease in both training and validation loss across the epochs. However, a slight gap between the training and validation loss emerges towards the end, hinting at potential overfitting. GoogleNet shows a rapid initial decrease in training loss followed by stabilization around a relatively low value, The small gap between the training and validation loss indicates that the model is generalizing well to unseen data. DenseNet121 demonstrates a decreasing trend in both losses initially, but the validation loss fluctuates after a few epochs, indicating effective learning on training data but struggles with generalization. The increasing gap between training and validation loss after the initial drop supports this notion, suggesting overfitting. Lastly, MobileNetV2 exhibits rapid initial decrease in both losses, followed by stabilization at low values, a slight uptick in validation loss towards the end of training might signal the beginning of overfitting, potentially leading to decreased performance on unseen data. IV.5 HIL-CNN Integration with DenseNet121 To overcome the limitations of fully automated classification in cases of uncertain or transitional tool wear conditions—particularly between neighboring wear levels such as moderate and heavy we introduce a Human-in-the-Loop (HIL) collaborative model integrated into the DenseNet121-based convolutional framework. This hybrid approach enables a dynamic interplay between machine intelligence and human expertise, aiming to improve the reliability of decisions in practical industrial settings. Initially, each image is evaluated by a fine-tuned DenseNet121 model, which outputs both a predicted class and a confidence score from the softmax layer. If this score falls below a pre-established threshold (e.g., 80%) or if the predicted wear severity falls within a critical transition zone (e.g., between 180 µm and 200 µm VB), the system flags the case for human review. A Gradio-powered interface facilitates this process by allowing operators to visualize the image and either validate or correct the model’s prediction in real-time. Corrections made by the expert are systematically stored and can later be reintegrated into the dataset to support incremental or continual training, enhancing the model’s adaptability. This HIL mechanism minimizes high-impact classification errors, reinforces the interpretability of AI outputs, and fosters confidence in automated decisions particularly under the variable and high-stakes conditions characteristic of Industry 4.0 environments. Ultimately, the approach supports intelligent tool condition monitoring and predictive maintenance strategies, where precision, accountability, and adaptability are essential. IV.6 Identified Research Gaps and Future Perspectives Despite the promising results achieved in this study, several critical research gaps remain to be addressed to advance the practical deployment and scientific depth of tool wear classification systems. First, the current CNN-based models operate as black boxes, offering limited interpretability of their predictions. This hinders the ability of end-users, such as process engineers, to trust or validate the system’s decisions. Second, while classification performance was evaluated on static image datasets, the model’s applicability in real-time industrial settings remains untested. The impact of latency, sensor integration, and real-time decision-making under actual machining conditions requires further investigation. Additionally, the confusion matrix reveals significant ambiguity between adjacent wear classes particularly moderate and heavy wear highlighting the need for probabilistic modeling or soft classification strategies to handle transitional states more effectively. Furthermore, although inference time was assessed, the suitability of the models for deployment on resource-constrained edge devices (e.g., Raspberry Pi or smart industrial cameras) was not fully explored. Lastly, the temporal nature of wear progression was not captured, as the study focused on isolated images. Future research could benefit from incorporating sequential learning architectures, such as CNNs combined with LSTM or Transformer networks, to track and predict wear evolution over time in a more robust and informed manner. V. CONCLUSION This study presented an approach for monitoring flank wear on cutting tools using deep convolutional neural networks (CNNs) and transfer learning techniques. By leveraging pre-trained CNN architectures (VGG16, GoogleNet, DenseNet121, and MobileNetV2) originally trained on the ImageNet dataset and subsequently fine-tuned on the tool wear image dataset, promising results were achieved in classifying flank wear into four distinct categories based on wear severity. The findings indicate that: This study effectively utilized four CNN architectures (VGGNet, GoogleNet, DenseNet121, and MobileNetV2) with transfer learning to classify cutting tool flank wear into four categories. Data augmentation techniques proved crucial in enhancing model robustness and mitigating overfitting. Adam optimizer consistently outperformed SGD and RMSprop, achieving higher accuracy and faster training for all tested CNN models. GoogleNet and DenseNet121 achieved the highest classification accuracy (96.80%), outperforming VGG16 and MobileNetV2. GoogleNet is preferred over DenseNet121 due to its significantly better training performance. Future research could explore Expanding the dataset to address misclassification between specific wear categories, further improving the system's accuracy. Investigating other CNN architectures and transfer learning techniques for potential performance improvements. Finally, Implementing the system in real-time machining environments for continuous tool wear monitoring. Declarations corresponding author : ELBAH Mohamed E-mail address: [email protected] Acknowledgements: This work was achieved in the Mechanics and Structures Laboratory (LMS) (Guelma University, Algeria). The authors would like to thank the Algerian Ministry of Higher Education and Scientific Research (MESRS). Funding: This study was funded by the Algerian Ministry of Higher Education and Scientific Research Financial interests: The authors declare they have no financial interests. Competing interests: The authors have no competing interests to declare that are relevant to the content of this article. Author contributions: The study was carried out with the contribution of all the authors. Material preparation and data collection were performed by Oussama Dahmoune, Ikhlas Meddour, Mohamed Atmane Yalles, and Mohamed Elbah. Data analysis was performed by Oussama Dahmoune, Mohamed Elbah and Ikhlas Meddour. Python programming was performed by Oussama Dahmoune and Mohamed Elbah. The first draft of the manuscript was prepared by Oussama Dahmoune. Mohamed Elbah and Ikhlas Meddour reviewed it, and wrote the final manuscript. All authors agreed the final manuscript. References Vetrichelvan G, Sundaran S, Kumaran SS, Velmurugan P (2014) An investigation of tool wear using acoustic emission and genetic algorithm. J Vib Control 21(15) : 3061–3066 Ezugwu, E.O., Wang, Z.M., Machado, A.R., 1999. The machinability of nickel-based alloys: a review. Journal of Materials Processing Technology 86 (1-3), 1–16, https://doi.org/10.1016/S0924-0136(98)00314-8. Wang, B., Liu, Z., 2018. Influences of tool structure, tool material and tool wear on machined surface integrity during turning and milling of titanium and nickel alloys: a review. Int J Adv Manuf Technol 98 (5-8), 1925–1975. Jeon, J.U., Kim, S.W., 1988. Optical flank wear monitoring of cutting tools by image processing. Wear 127 (2), 207–217 Y.Q. Zhou, B.T. Sun, W.F. Sun, Z. Lei, Tool wear condition monitoring based on a two-layer angle kernel extreme learning machine using sound sensor for milling process, J. Intell. Manuf. (2020), https://doi.org/10.1007/s10845-020-01663-1 Avinash, C.; Raguraman, S.; Ramaswamy, S.; Muthukrishnan, N. An Investigation on Effect of Workpiece Reinforcement Percentage on Tool Wear in Cutting Al-SiC Metal Matrix Composites. In Proceedings of the ASME International Mechanical Engineering Congress and Exposition, Seattle, WA, USA, 11–15 November 2008; pp. 561–566. Ee, K.C.; Balaji, A.K.; Jawahir, I.S. Progressive tool-wear mechanisms and their effects on chip-curl/chip-form in machining with grooved tools: An extended application of the equivalent toolface (et) model. Wear 2003, 255, 1404–1413. Nordgren, A.; Melander, A. Tool wear and inclusion behaviour during turning of a calciumtreated quenched and tempered steel using coated cemented carbide tools. Wear 1990, 139, 209–223. Junior, Joseph KalilKhoury, et al. "Monitoring of flank wear and damage on turning cutting tools by image processing." The Journal of Engineering and Exact Sciences 6.2 (2020): 0098-0106. Miao, Huihui, et al. "A U-Net-based approach for tool wear area detection and identification." IEEE Transactions on Instrumentation and Measurement 70 (2020): 1-10. Panda, A., & Kumar, R. (2022, May). On Tool Wear Prediction Using Artificial Neural Network and Regression Methodology During. In Proceedings of 2nd International Conference on Smart Computing and Cyber Security: Strategic Foresight, Security Challenges and Innovation (SMARTCYBER 2021) (Vol. 395, p. 315). Springer Nature. Agrawal, K., Panda, A., & Sahoo, A. K. (2024, August). Advancements in tool wear monitoring in turning operations: digital image processing and AI techniques. In Journal of Physics: Conference Series (Vol. 2818, No. 1, p. 012040). IOP Publishing. Brili, Nika, MirkoFicko, and Simon Klančnik. "Automatic identification of tool wear based on thermography and a convolutional neural network during the turning process." Sensors 21.5 (2021): 1917. Sun, Wei-Heng, and Syh-ShiuhYeh. "Using the machine vision method to develop an on-machine insert condition monitoring system for computer numerical control turning machine tools." Materials 11.10 (2018): 1977 Lim, Meng Lip, et al. "Tool wear prediction in turning using workpiece surface profile images and deep learning neural networks." The International Journal of Advanced Manufacturing Technology 120.11-12 (2022): 8045-8062. Kumar, M. Phani, Samik Dutta, and N. C. Murmu. "Tool wear classification based on machined surface images using convolution neural networks." Sādhanā 46 (2021): 1-12. Ambadekar, P. K., and C. M. Choudhari. "CNN based tool monitoring system to predict life of cutting tool." SN Applied Sciences 2 (2020): 1-11. Dahmoune, O. Meddour, I. Elbah, M. Atmane Yallese, M. & Belhadi, S. (2025). Development of an adaptive tool condition monitoring system: integration of case-based reasoning with CNN. Journal of Intelligent Manufacturing, 1-14. A. Rehman, S. Naz, M.I. Razzak, H.A. Ibrahim, Automatic visual features for writer identification: a deep learning approach. IEEE Access 7, 17149–17157 (2019) G. Taguchi, S. Chowdhury, and S. Taguchi, Robust Engineering: Learn How to Boost Quality While Reducing Costs & Time to Market. New York, NY, USA: McGraw-Hill, 2000. Mikołajczyk, T., et al. "Predicting tool life in turning operations using neural networks and image processing." Mechanical systems and signal processing 104 (2018): 503-513 ISO 3685 Tool Life Testing with Single-Point Turning Tools. 1993. Available online: https://www.iso.org/fr/standard/9151.html (accessed on 7 January 2024). L. Perez and J. Wang, “The effectiveness of data augmentation in image classification using deep learning,” 2017, arXiv:1712.04621. [Online]. Available: https://arxiv.org/abs/1712.04621 A. Hernández-García and P. König, “Data augmentation instead of explicit regularization,” Jun. 2018, arXiv:1806.03852. Accessed: Sep. 4, 2020. [Online]. Available: http://arxiv.org/abs/1806.03852 K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition. ArXiv preprint: arXiv:1409.1556 (2014) C. Szegedy,W.Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, A. Rabinovich, Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1–9 (2015) Nandhini, S., and K. Ashokkumar. "An automatic plant leaf disease identification using DenseNet-121 architecture with a mutation-based henry gas solubility optimization algorithm." Neural Computing and Applications 34.7 (2022): 5513-5534. Dong, Ke, et al. "MobileNetV2 model for image classification." 2020 2nd International Conference on Information Technology and Computer Application (ITCA). IEEE, 2020. Miao, Huihui, et al. "A U-Net-based approach for tool wear area detection and identification." IEEE Transactions on Instrumentation and Measurement 70 (2020): 1-10. Cite Share Download PDF Status: Posted Version 1 posted 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-7189226","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":490345141,"identity":"201478d6-93f2-469c-aff2-235e801299ab","order_by":0,"name":"Mohamed ELBAH","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABE0lEQVRIiWNgGAWjYFACHhDBzMDAzgMkK2CibEB8gJAWZpCWM1DVxGthbCNCi25778FPNyqsGfiZeQ9+Lpxnkyc/v/kBw4eywwx857HrMTtzLlk650w6g2QzX7L0zG1pxQbH2AwYZ5w7zCB5AIeWGzkG0rlthxkMDvMYSPNuO5y4gY3BgJkXJHKwAZcW49+5/w4z2B/mMf7NO+d/4vw29g/Mf8GGYPcLUIuZdG4DUAEzj5k0b8OBxIZjPAbAcACKHMOh5cwZM+ucY+k8Eof50qx5jiUnbjiWU3Cw51w6j+QZHFqO9xjfzqmxluNv7z18m6fGLnF+8/GND36UWcvhCjEY4EHhHcAQGQWjYBSMglFAEgAAKWZdOwpBW84AAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0002-7194-2789","institution":"National School of Advanced Technology: Ecole Nationale Superieure de Techniques Avancees","correspondingAuthor":true,"prefix":"","firstName":"Mohamed","middleName":"","lastName":"ELBAH","suffix":""},{"id":490345142,"identity":"4b063482-60eb-480d-a8be-359734062cc0","order_by":1,"name":"Oussama DAHMOUNE","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Oussama","middleName":"","lastName":"DAHMOUNE","suffix":""},{"id":490345143,"identity":"d6a2eb21-7b0e-499d-9ffe-3b14777e935c","order_by":2,"name":"Ikhlas MEDDOUR","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Ikhlas","middleName":"","lastName":"MEDDOUR","suffix":""}],"badges":[],"createdAt":"2025-07-22 16:44:51","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7189226/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7189226/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":87801681,"identity":"bf22a526-8fb6-491a-a960-7d68a474878d","added_by":"auto","created_at":"2025-07-29 07:54:19","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":128931,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of the experimental and numerical setup\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7189226/v1/e437666ab8e69888525d47ce.jpg"},{"id":87802394,"identity":"2d588910-f1d8-464d-9e54-24eeb882af18","added_by":"auto","created_at":"2025-07-29 08:02:24","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":46466,"visible":true,"origin":"","legend":"\u003cp\u003eThe standardized measurement of tool wear follows the guidelines defined in ISO 3685:1993. [22]\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7189226/v1/0f467dfd8896ba42fdc0411e.jpg"},{"id":87801727,"identity":"4432b6b3-32e1-44d9-849e-af899ffed870","added_by":"auto","created_at":"2025-07-29 07:54:25","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":161815,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e(a) \u003c/strong\u003eThe categorization of the acquired images into wear classes, \u003cstrong\u003e(b)\u003c/strong\u003e sample images for flank wear dataset, \u003cstrong\u003e(c) \u003c/strong\u003eMicroscopic visualization of cutting tool flank wear progression, calibrated with a 100 µm real scale.\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7189226/v1/170d90437522d74e277bd507.jpg"},{"id":87801620,"identity":"2eb34634-2dee-4d0d-bb44-2f979a75c4e6","added_by":"auto","created_at":"2025-07-29 07:54:17","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":49543,"visible":true,"origin":"","legend":"\u003cp\u003edata augmentation techniques: brightness, contrast, shearing, and zooming.\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7189226/v1/ee86cfc7a212dc6ffe31be87.jpg"},{"id":87801704,"identity":"2f81cf00-28f3-42b6-9438-92ac5086df08","added_by":"auto","created_at":"2025-07-29 07:54:22","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":68692,"visible":true,"origin":"","legend":"\u003cp\u003eblocks diagram of the cutting tool flank wear classification system.\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7189226/v1/f5ee6d6ea0a6d0f596d605b4.jpg"},{"id":87801609,"identity":"1fca1277-2ec9-4cc8-abfa-23814e2602d8","added_by":"auto","created_at":"2025-07-29 07:54:15","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":168622,"visible":true,"origin":"","legend":"\u003cp\u003eoverarching framework of the proposed automated classification system for cutting tool flank wear\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7189226/v1/f769c63d7da171a5df0eb175.jpg"},{"id":87801742,"identity":"df81f976-dbb0-40f1-b8dd-3a200e623a79","added_by":"auto","created_at":"2025-07-29 07:54:27","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":96191,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion Matrix of (a) VGGNet-16 Model, (b) GoogLeNet, (c) DenseNet121, and (d) MobileNetV2\u003c/p\u003e","description":"","filename":"7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7189226/v1/a9c6de29dde7b0dbcb313cb1.jpg"},{"id":87801730,"identity":"9698ab8f-1458-41e1-99a8-7cec008ca2cb","added_by":"auto","created_at":"2025-07-29 07:54:26","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":33145,"visible":true,"origin":"","legend":"\u003cp\u003eMisclassification between two categories of flank wear\u003c/p\u003e","description":"","filename":"8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7189226/v1/0d05ba8d7a92b324711fdb95.jpg"},{"id":87802392,"identity":"32800d7f-3491-47b9-a1cb-d6e47cce95b2","added_by":"auto","created_at":"2025-07-29 08:02:19","extension":"jpg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":54657,"visible":true,"origin":"","legend":"\u003cp\u003etraining and validation accuracy of three different convolutional neural network (CNN) models - (a) VGG16, (b) GoogLeNet (Inception), (c) DenseNet121, and (d) MobileNetV2 over the course of training epochs.\u003c/p\u003e","description":"","filename":"9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7189226/v1/d88b727ae7d744453cd0247e.jpg"},{"id":87801737,"identity":"a6674ae1-cc93-43cd-8f6e-01723825d592","added_by":"auto","created_at":"2025-07-29 07:54:26","extension":"jpg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":41943,"visible":true,"origin":"","legend":"\u003cp\u003eTraining and Validation Loss for (a) VGG16, (b) GoogLeNet (Inception), and (c) DenseNet121, and (d) MobileNetV2\u003c/p\u003e","description":"","filename":"10.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7189226/v1/50e66c23c2c6250f2285c5c1.jpg"},{"id":87802393,"identity":"4e413be6-f463-4c06-b3e6-2834f6cbaf15","added_by":"auto","created_at":"2025-07-29 08:02:24","extension":"jpg","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":56935,"visible":true,"origin":"","legend":"\u003cp\u003eUnnumbered Image in the iv. Comparative Analysis, Results and Discussion Section.\u003c/p\u003e","description":"","filename":"Unnumber.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7189226/v1/3d3adcca607b8c9cea3b1d09.jpg"},{"id":94021687,"identity":"1ca30dfe-7eef-4196-b87a-42e31fb64bc9","added_by":"auto","created_at":"2025-10-21 12:29:42","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1897780,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7189226/v1/bd077e87-f539-4d77-a53a-c496c2385c40.pdf"}],"financialInterests":"","formattedTitle":"Deep Transfer Learning and Data Augmentation for Automatic Flank Wear Detection and Classification","fulltext":[{"header":"Highlights","content":"\u003cul type=\"disc\"\u003e\n \u003cli\u003eA CNN-based system was developed to classify flank wear into four severity levels.\u003c/li\u003e\n \u003cli\u003eTransfer learning improved model accuracy using pre-trained architectures.\u003c/li\u003e\n \u003cli\u003eData augmentation reduced overfitting and enhanced generalization performance.\u003c/li\u003e\n \u003cli\u003eGoogleNet and DenseNet121 achieved the best accuracy, reaching 96.80%.\u003c/li\u003e\n \u003cli\u003eGoogleNet was preferred due to faster training time and stable performance.\u003c/li\u003e\n\u003c/ul\u003e"},{"header":"I. INTRODUCTION","content":"\u003cp\u003eIn machining processes, achieving a superior surface finish is critical and is largely influenced by the condition of the cutting tool insert [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. However, various factors such as wear, thermal fatigue, and corrosion [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e] can adversely in machining processes, achieving a superior surface finish and high dimensional accuracy is a critical performance criterion, which depends largely on the condition of the cutting tool insert [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. However, this condition is subject to degradation due to various mechanisms such as tool wear, thermal fatigue, and corrosion [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Among these, tool wear has the most pronounced impact, as it directly influences tool life, machining quality, productivity, and overall operational costs. Consequently, real-time monitoring of tool wear is essential to maintain manufacturing efficiency and ensure product quality.\u003c/p\u003e\u003cp\u003eTool wear is typically assessed using either direct methods (such as visual inspection or image acquisition) or indirect methods (such as acoustic emission, vibration, and current signal analysis) [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. In recent years, Artificial Intelligence (AI) technologies\u0026mdash;particularly Convolutional Neural Networks (CNNs)\u0026mdash;have demonstrated strong potential for automating the classification of different wear types. CNNs have proven especially effective in recognizing common tool wear forms such as flank wear, chipping, spalling, and built-up edge, which primarily occur at the flank and tip regions of the cutting insert [\u003cspan additionalcitationids=\"CR6 CR7\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eNotable contributions in the literature reinforce the relevance of AI in tool wear monitoring. For example, Khoury Junior et al. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] developed an image-based classification system using Bayesian discriminant functions combined with the Hough Transform, achieving 94.3% accuracy in detecting flank wear (VB). Similarly, Miao et al. [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] proposed a U-Net-based CNN model to classify VB wear in CNC cutting tools. By augmenting their dataset from 186 to 1,100 images, they significantly improved the robustness and accuracy of their wear detection system.\u003c/p\u003e\u003cp\u003eComplementary studies further support these findings. Panda et al. [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] compared the performance of Artificial Neural Networks (ANNs) and Response Surface Methodology (RSM) for predicting flank wear (VBc) during the turning of AISI 4340 steel. Their experiments, designed using a Taguchi L16 orthogonal array, revealed that ANN outperformed regression models, achieving a remarkably low average error of 2.02%. Additionally, Agrawal et al. [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] conducted a comprehensive review of AI-based approaches for tool wear monitoring in turning operations. Their findings emphasized the effectiveness of CNNs for real-time and precise detection, while also highlighting persistent challenges such as the variability in wear progression and the limited integration of CNC machine data in existing models.\u003c/p\u003e\u003cp\u003eBrili et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] proposed a thermal imaging approach combined with CNNs to classify tool wear states into \u0026ldquo;GO\u0026rdquo; or \u0026ldquo;NO-GO\u0026rdquo; conditions, achieving 99.55% accuracy in turning processes. Sun et al [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] applied a PCNN-based monitoring system to classify various tool defects\u0026mdash;flank wear, built-up edge, chipping, and cracks\u0026mdash;using CNC turning data. Similarly, Lim et al. [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] analyzed machined surface profiles using deep neural models, demonstrating the superiority of CNNs over DNNs, achieving 98.9% accuracy while minimizing RMSE.In another study, Kumar et al. [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] developed a CNN-based system for flank wear classification using raw and minimally preprocessed images, reaching 96.3% and 99.9% accuracy, respectively. Ambadekar et al. [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] employed a ResNet-50 CNN to estimate the remaining life of cutting tools by categorizing VB wear into severity levels (A, B, and C). In a more adaptive approach, Dahmoune et al. [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e] combined CNNs with Case-Based Reasoning (CBR) to develop a hybrid Tool Condition Monitoring (TCM) system, achieving 98% training accuracy and 96% validation accuracy. Despite notable advancements in deep learning-based tool condition monitoring systems, several critical limitations persist. Many existing approaches exhibit poor generalization due to limited dataset sizes, class imbalance, and insufficient adaptation to real-world industrial environments. Conventional CNN-based frameworks are frequently applied in a generic manner, lacking customization for the specific wear patterns and operational constraints encountered in machining processes. Furthermore, the integration of human validation or incremental learning strategies remains rare, which restricts the system\u0026rsquo;s adaptability in ambiguous or transitional wear scenarios.\u003c/p\u003e\u003cp\u003eRecent studies, including the one cited in [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], have highlighted these shortcomings, emphasizing the urgent need for more robust, interpretable, and industry-ready AI solutions. These works underscore the effectiveness of advanced image processing and deep convolutional neural networks (CNNs) in improving classification accuracy, while simultaneously pointing out challenges related to data variability and limited integration within CNC environments ,In response to these challenges, the present study proposes a comprehensive system for the automatic classification of flank wear using four well-established CNN architectures: VGG16, GoogleNet (Inception), DenseNet121, and MobileNetV2. Transfer learning is employed to leverage the representational power of pre-trained models, thereby reducing the dependence on large annotated datasets. In addition, custom-designed data augmentation techniques are applied to enhance generalization and mitigate overfitting.\u003c/p\u003e\u003cp\u003eThe key scientific contributions of this work include the development of a domain-specific CNN-based classification pipeline and a thorough performance evaluation using metrics such as accuracy, precision, recall, and F1-score, all validated under realistic dry turning industrial conditions.\u003c/p\u003e\u003cp\u003eThis paper is structured as follows. The next section reviews the relevant CNN architectures and related literature. This is followed by a detailed description of the experimental setup, dataset composition, and the implemented methodology. The results are then presented and analyzed, concluding with a discussion of future research directions.\u003c/p\u003e"},{"header":"II. EXPERIMENTS AND METHODOLOGY","content":"\u003cp\u003eThis section presents the detailed methodologies employed to evaluate the effectiveness of the proposed automated system for cutting tool flank wear classification and detection. A carefully curated dataset, specifically designed for this purpose, is utilized. The classification models are trained on a designated training subset of this dataset and subsequently evaluated using unseen test images. The research methodology comprises two key phases:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eExperimental phase: This phase focuses on conducting turning operations on a lathe machine to acquire a comprehensive dataset of tool wear images.\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eData processing and model training/testing: This phase encompasses processing the acquired data, training the deep learning models using the training dataset, and evaluating their performance on the unseen test images.\u003c/p\u003e\n \u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u003cstrong\u003eⅡ.1 Experimental set-up\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTurning operations were carried out on a conventional SN40 lathe (TOSTRENCIN), as illustrated in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003ea. The workpiece material, shown in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003ec, consisted of AISI 1045 steel a widely utilized medium carbon steel in industrial machining applications. Its detailed chemical composition, presented in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, was analyzed in the laboratory of the national company ORSIM using specialized, high-precision analytical equipment to ensure compositional accuracy and process reliability.\u003c/p\u003e\n\u003cp\u003eThe cylindrical workpiece initially measured 77 mm in diameter and 115 mm in machining length. Tool wear monitoring was performed using a DM9 industrial-grade digital microscope (Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003eb), equipped with high-resolution optical lenses and a calibrated LED illumination system. This setup allowed for consistent, glare-free visualization of the cutting tool\u0026rsquo;s flank surface. The microscope captured images at a resolution of 1280 \u0026times; 720 pixels and transmitted them to a connected computer, enabling high-fidelity data acquisition suitable for automated wear analysis. Based on manufacturer specifications and experimental calibration, the measurement uncertainty for flank wear width (VB) using the DM9 system was estimated at \u0026plusmn;\u0026thinsp;10 \u0026micro;m well within the precision range recomended by the ISO 3685 standard for tool wear assessment in turning processes.\u003c/p\u003e\n\u003cp\u003eA standard PSBNR 2525 M12 tool holder was used to secure an SNMG 120408 carbide insert. This insert geometry, featuring eight cutting edges, was selected based on its suitability for turning AISI1045 steel in its normalized condition. Figure\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e(d) highlights the insert\u0026apos;s mounting within the tool holder via a central hole to ensure stability during turning.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eChemical composition of the workpiece material.\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\u003eMaterial\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"5\"\u003e\n \u003cp\u003eChemical composition %\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eC\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eS\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMn\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eP\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSi\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\u003eAISI1045\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.031\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\u003e0.029\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.36\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\u003eTurning experiments were structured using an L4(2\u0026sup3;) orthogonal array to methodically vary three critical input parameters: cutting speed, feed rate, and depth of cut, each at two discrete levels, in accordance with the robust design principles established by Taguchi et al. [\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e]. This orthogonal approach allows for efficient evaluation of the main effects of each factor while significantly reducing the number of experimental trials required.\u003c/p\u003e\n\u003cp\u003eFor Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, the selection of factors and their corresponding levels was based on prior authoritative studies (such as Sandvik tooling guidelines and ISO 3685 standards) and was further validated through feasibility tests conducted in the laboratory using an SN40 lathe and SNMG insert. These choices are fully justified in the experimental methodology section.\u003c/p\u003e\n\u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e presents the final levels selected for each machining parameter, and Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e details the specific combinations applied in each experimental run. The parameters were ultimately aligned with the tool manufacturer\u0026rsquo;s recommended conditions for machining AISI 1045 steel, ensuring both experimental relevance and industrial applicability.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eAssignment of the levels to the factors.\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\u003eNo\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eParameter\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eUnit\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eLevel\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e-1\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e+\u0026thinsp;1\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\u003eCutting Speed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003em/min\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e140\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e180\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\u003eFeed Rate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003emm/rev\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.16\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\u003eDepth of Cut\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003emm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.4\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 class=\"gridtable\"\u003e\n \u003cdiv align=\"char\" class=\"colspec\"\u003e\u003cbr\u003e\u003c/div\u003e\n \u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eDesign matrix as per L4 orthogonal array.\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\u003eExperiment\u003c/p\u003e\n \u003cp\u003eNumber\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003eMachining parameters\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eCutting speed (m/min)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eFeed Rate (mm/rev)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDepth of cut (mm)\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=\"char\"\u003e\n \u003cp\u003e140\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.2\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=\"char\"\u003e\n \u003cp\u003e140\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.4\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=\"char\"\u003e\n \u003cp\u003e180\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.2\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=\"char\"\u003e\n \u003cp\u003e180\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.4\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\u003e\u003cstrong\u003eⅡ.2 Flank wear and tool life\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThere is a direct correlation between machining time and the rate of flank wear progression [\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e]. This relationship is a key factor in determining tool life, as excessive wear can lead to poor surface finish, dimensional inaccuracies, and ultimately, tool failure. ASTM Standard guidelines provide a framework for evaluating tool longevity, often assessed by the extent of flank wear, referred to as VB, or the width of the worn edge. As illustrated in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, VB is typically measured in zone B of the cutting edge, which extends from the corner radius to the point where the width of the undeformed chip is b/4. When wear progresses uniformly across the tool, VB typically reaches 0.3 mm as a critical threshold. This critical value is often used as a criterion for tool replacement, ensuring optimal machining performance and minimizing downtime.\u003c/p\u003e"},{"header":"III.\tDATASET AND MODEL ARCHITECTURES","content":"\u003cp\u003eThe implementation was carried out in the Google Colab runtime environment, utilizing libraries available within this platform and employing Python 3 as operating environment. To accelerate the computationally intensive tasks of machine learning and deep learning, the research utilized T4 GPUs, which provide significantly faster processing speeds compared to CPUs.\u003c/p\u003e\u003cp\u003eThis section describes the dataset used for training and evaluating deep learning models for classifying cutting tool flank wear, including its creation process, categories, and employed data augmentation techniques. It also introduces different deep convolutional neural network architectures used to monitor flank wear by applying transfer learning techniques to extract visual features for classification using a SoftMax layer.\u003c/p\u003e\u003cp\u003e\u003cb\u003eⅢ.1 Data Set\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe dataset comprises images of cutting tools classified into four distinct flank wear categories, as illustrated in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, corresponding to progressive stages of tool degradation. This classification framework is grounded in the conventional tool wear progression curve, which includes the initial wear phase, the steady-state phase, and the accelerated wear phase leading to tool failure. To enhance diagnostic granularity and support proactive maintenance strategies, an intermediate wear class was deliberately introduced between the steady and critical stages. This additional category enables more accurate detection of wear evolution and provides operators with a critical window for intervention before catastrophic failure occurs. The proposed four-class system significantly improves diagnostic precision, supports real-time decision-making, and aligns with the stringent requirements of modern industrial monitoring and predictive maintenance systems.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eFlank wear classification based on VB measurement.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCategories\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLimit values of the flank wear level\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eClass 1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0 \u0026micro;m\u0026thinsp;\u0026lt;\u0026thinsp;VB\u0026thinsp;\u0026le;\u0026thinsp;100 \u0026micro;m\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eClass 2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e100 \u0026micro;m\u0026thinsp;\u0026lt;\u0026thinsp;VB\u0026thinsp;\u0026le;\u0026thinsp;180 \u0026micro;m\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eClass 3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e180 \u0026micro;m\u0026thinsp;\u0026lt;\u0026thinsp;VB\u0026thinsp;\u0026le;\u0026thinsp;270 \u0026micro;m\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eClass 4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e270 \u0026micro;m\u0026thinsp;\u0026lt;\u0026thinsp;VB\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e illustrates the composition of the image dataset before augmentation. It consists of 480 images categorized into four distinct classes, with each class represented by a specific VB range. The figure also showcases sample images from each class, highlighting the visual differences in flank wear progression.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo overcome the limited size of the image dataset, data augmentation techniques were employed to synthetically expand it. Data augmentation not only increases the quantity of images but also enhances the generalization ability of the model, helping to reduce overfitting [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. This approach, illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, involves applying transformations such as brightness adjustment, contrast modification, shearing, and zooming to the original images. By generating diverse variations of the original images, a richer dataset for training the models is created. This method ensures that the models are robust and better equipped to generalize to new and unseen examples of flank wear [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eⅢ.2 Flank Wear Classification and Detection\u003c/b\u003e\u003c/p\u003e\u003cp\u003eAfter extracting visual features from the target dataset through transfer learning, classification and detection tasks were conducted. The dataset is divided into training and testing sets, with each set containing images from all four wear intensity categories. Specifically, 80% of all images were allocated to the training set, and 20% were assigned to the test set, as illustrated in Table\u0026nbsp;5. Flank wear classification of the cutting tool was performed using a Softmax layer as the final activation function within the transfer learning model.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eTABLE.5\u003c/b\u003e Distribution of training and testing images in each flank wear category.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLight wear\u003c/p\u003e\u003cp\u003e\u003cem\u003e(Class1)\u003c/em\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eModerate wear\u003c/p\u003e\u003cp\u003e\u003cem\u003e(Class 2)\u003c/em\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eHeavy wear\u003c/p\u003e\u003cp\u003e\u003cem\u003e(Class 3)\u003c/em\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eExtreme wear\u003c/p\u003e\u003cp\u003e\u003cem\u003e(Class 4)\u003c/em\u003e\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTraining set\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e277\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e273\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e294\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e270\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTest set\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e69\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e68\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e74\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e67\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTotal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e346\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e341\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e368\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e337\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eⅢ.3 CNN-based architectures\u003c/b\u003e\u003c/p\u003e\u003cp\u003eCNN or ConvNet stands for convolutional neural network, designed to automatically extract significant visual patterns from raw image pixels with minimal pre-processing. To leverage the power of these networks for flank wear prediction, a transfer learning approach was adopted, utilizing pre-trained models that have already learned general image features. This study investigates four popular convolutional neural network (CNN) architectures: VGG16, GoogleNet (Inception), DenseNet121, and MobileNetV2.\u003c/p\u003e\u003cp\u003e\u003cb\u003eⅢ.3.1VGG16\u003c/b\u003e\u003c/p\u003e\u003cp\u003eVGG16 [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] is a convolutional neural network architecture developed by the Visual Geometry Group (VGG) at the University of Oxford in 2014. It features a deep structure comprising 16 layers, including 13 convolutional layers and 3 fully connected layers. VGG16 has demonstrated significant success in various computer vision tasks, including image classification and object detection. Trained on the ImageNet dataset, VGG16 serves as a widely adopted benchmark model in the deep learning domain, typically operating on input sizes of 224x224 pixels.\u003c/p\u003e\u003cp\u003e\u003cb\u003eⅢ.3.2 GoogleNet (Inception)\u003c/b\u003e\u003c/p\u003e\u003cp\u003eGoogleNet [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], also referred to as Inception, is a CNN architecture devised by Google researchers, acclaimed as the winner of the 2014 ImageNet (sitiha) Large-Scale Visual Recognition Challenge (ILSVRC). Characterized by its deep structure and efficient utilization of computational resources, GoogleNet introduces inception modules enabling parallel processing at multiple scales. This architecture effectively captures both local features and global information, rendering it highly accurate and efficient for various computer vision tasks.\u003c/p\u003e\u003cp\u003e\u003cb\u003eⅢ.3.3 DenseNet121\u003c/b\u003e\u003c/p\u003e\u003cp\u003eDenseNet121 [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], proposed by researchers at the Computer Vision Center (CVC) and the Autonomous University of Barcelona in 2016, is part of the DenseNet family of models. DenseNet, denoting Dense Convolutional Network, features densely connected layers, with each layer linked to every other layer in a feed-forward manner. This architecture promotes feature reuse and facilitates gradient flow, thereby enhancing learning efficiency and model performance. DenseNet121 finds widespread application in image classification, object detection, and semantic segmentation tasks.\u003c/p\u003e\u003cp\u003e\u003cb\u003eⅢ.3.4 MobileNetV2\u003c/b\u003e\u003c/p\u003e\u003cp\u003eMobileNetV2 [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] is a convolutional neural network architecture introduced by Google in 2018, specifically designed for mobile and embedded vision applications. It builds upon the original MobileNet architecture with improvements aimed at efficiency and performance. MobileNetV2 features inverted residual blocks with linear bottleneck layers, enabling deeper networks while maintaining low computational cost and memory footprint. This architecture leverages depth wise separable convolutions to reduce computation and parameter size, making it suitable for resource-constrained environments such as smartphones and IoT devices. Trained on large-scale datasets like ImageNet, MobileNetV2 achieves competitive accuracy in tasks like image classification, object detection, and semantic segmentation, making it a popular choice for real-world mobile vision applications.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eⅢ.4 Evaluation metrics\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe performance of the VGG16, GoogleNet, DenseNet121, and MobileNetV2 algorithms was analyzed using common metrics: precision, recall, F1-score, and accuracy. These metrics are computed using equations (1) to (4) [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e], where TP denotes true positives, TN represents true negatives, FP signifies false positives, and FN stands for false negatives. An effective classifier aims to optimize both precision and recall concurrently, ensuring precise classification into the correct categories. The F1-score directly reflects the balance between precision and recall, while accuracy offers a straightforward interpretation of classifier performance by indicating the ratio of correctly classified images to the total number of images evaluated.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathbf{P}\\mathbf{r}\\mathbf{e}\\mathbf{c}\\mathbf{i}\\mathbf{s}\\mathbf{i}\\mathbf{o}\\mathbf{n}=\\frac{\\:\\:\\:\\:\\text{N}\\text{o}.\\:\\text{o}\\text{f}\\:\\text{c}\\text{o}\\text{r}\\text{r}\\text{e}\\text{c}\\text{t}\\:\\text{c}\\text{l}\\text{a}\\text{s}\\text{s}\\text{i}\\text{f}\\text{i}\\text{c}\\text{a}\\text{t}\\text{i}\\text{o}\\text{n}\\text{s}\\:\\text{i}\\text{n}\\text{t}\\text{o}\\:\\text{t}\\text{h}\\text{e}\\:\\text{j}\\:\\text{c}\\text{l}\\text{a}\\text{s}\\text{s}}{\\text{N}\\text{o}.\\:\\text{o}\\text{f}\\:\\text{a}\\text{l}\\text{l}\\:\\text{c}\\text{l}\\text{a}\\text{s}\\text{s}\\text{i}\\text{f}\\text{i}\\text{c}\\text{a}\\text{t}\\text{i}\\text{o}\\text{n}\\text{s}\\:\\text{i}\\text{n}\\text{t}\\text{o}\\:\\text{t}\\text{h}\\text{e}\\:\\text{j}\\:\\text{c}\\text{l}\\text{a}\\text{s}\\text{s}}\\:=\\frac{\\mathbf{T}\\mathbf{P}}{\\mathbf{T}\\mathbf{P}\\:+\\:\\mathbf{F}\\mathbf{P}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(1)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathbf{R}\\mathbf{e}\\mathbf{c}\\mathbf{a}\\mathbf{l}\\mathbf{l}=\\frac{\\text{N}\\text{o}.\\:\\text{o}\\text{f}\\:\\text{c}\\text{o}\\text{r}\\text{r}\\text{e}\\text{c}\\text{t}\\:\\text{c}\\text{l}\\text{a}\\text{s}\\text{s}\\text{i}\\text{f}\\text{i}\\text{c}\\text{a}\\text{t}\\text{i}\\text{o}\\text{n}\\text{s}\\:\\text{i}\\text{n}\\text{t}\\text{o}\\:\\text{t}\\text{h}\\text{e}\\:\\text{j}\\:\\text{c}\\text{l}\\text{a}\\text{s}\\text{s}}{\\text{N}\\text{o}.\\:\\text{o}\\text{f}\\:\\text{a}\\text{c}\\text{t}\\text{u}\\text{a}\\text{l}\\:\\text{i}\\text{n}\\text{s}\\text{t}\\text{a}\\text{n}\\text{c}\\text{e}\\text{s}\\:\\text{i}\\text{n}\\text{t}\\text{o}\\:\\text{t}\\text{h}\\text{e}\\:\\text{j}\\:\\text{c}\\text{l}\\text{a}\\text{s}\\text{s}}=\\frac{\\mathbf{T}\\mathbf{P}}{\\mathbf{T}\\mathbf{P}\\:+\\:\\mathbf{F}\\mathbf{N}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(2)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathbf{F}{1}_{-}\\mathbf{S}\\mathbf{c}\\mathbf{o}\\mathbf{r}\\mathbf{e}=2\\times\\:\\frac{\\mathbf{P}\\mathbf{r}\\mathbf{e}\\mathbf{c}\\mathbf{i}\\mathbf{s}\\mathbf{i}\\mathbf{o}\\mathbf{n}\\times\\:\\mathbf{R}\\mathbf{e}\\mathbf{c}\\mathbf{a}\\mathbf{l}\\mathbf{l}}{\\mathbf{P}\\mathbf{r}\\mathbf{e}\\mathbf{c}\\mathbf{i}\\mathbf{s}\\mathbf{i}\\mathbf{o}\\mathbf{n}+\\mathbf{R}\\mathbf{e}\\mathbf{c}\\mathbf{a}\\mathbf{l}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(3)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\mathbf{A}\\mathbf{c}\\mathbf{c}\\mathbf{u}\\mathbf{r}\\mathbf{a}\\mathbf{c}\\mathbf{y}=\\frac{\\text{N}\\text{o}.\\:\\text{o}\\text{f}\\:\\text{a}\\text{l}\\text{l}\\:\\text{c}\\text{o}\\text{r}\\text{r}\\text{e}\\text{c}\\text{t}\\:\\text{c}\\text{l}\\text{a}\\text{s}\\text{s}\\text{i}\\text{f}\\text{i}\\text{c}\\text{a}\\text{t}\\text{i}\\text{o}\\text{n}\\text{s}}{\\text{N}\\text{o}.\\:\\text{o}\\text{f}\\:\\text{a}\\text{l}\\text{l}\\:\\text{c}\\text{l}\\text{a}\\text{s}\\text{s}\\text{i}\\text{f}\\text{i}\\text{c}\\text{a}\\text{t}\\text{i}\\text{o}\\text{n}\\text{s}}=\\frac{\\sum\\:_{\\mathbf{i}=1}^{\\mathbf{j}}{\\mathbf{T}\\mathbf{P}}_{\\mathbf{i}}}{\\mathbf{n}\\mathbf{o}.\\:\\mathbf{o}\\:\\mathbf{f}\\:\\mathbf{a}\\mathbf{l}\\mathbf{l}\\:\\mathbf{c}\\mathbf{l}\\mathbf{a}\\mathbf{s}\\mathbf{s}\\mathbf{i}\\mathbf{f}\\mathbf{i}\\mathbf{c}\\mathbf{a}\\mathbf{t}\\mathbf{i}\\mathbf{o}\\mathbf{n}\\mathbf{s}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e(4)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"IV.\tCOMPARATIVE ANALYSIS, RESULTS AND DISCUSSION","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eⅣ.1 Evaluating CNN Performance with Different Optimizers\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo optimize the performance of the CNN models (VGG16, GoogleNet, DenseNet121, and MobileNetV2), a grid search strategy was employed to tune hyperparameters. This involved systematically exploring different combinations. The optimizers evaluated were Adam, SGD, and RMSprop, chosen for their common use in deep learning and distinct optimization strategies. A learning rate and 10 epochs were identified as yielding the best results across the models. This combination struck a balance between achieving high accuracy and minimizing training time. The resulting performance metrics for each model and optimizer are presented in detail in Tables 6, 7, 8, and 9, allowing for a comprehensive comparison and analysis of the models\u0026apos; effectiveness.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTABLE 6.\u0026nbsp;\u003c/strong\u003eComparative Performance of CNN Architectures for Image (for \u003cstrong\u003eVGG16)\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"551\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eOptimizer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eF1_Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAvg Epoch\u003c/p\u003e\n \u003cp\u003eTime (s)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eInitial Epoch\u003c/p\u003e\n \u003cp\u003eTime (s)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eVGG16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAdam\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9537\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9528\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9542\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9526\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSGD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.5487\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.4595\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e0.5234\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.5587\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 86px;\"\u003e\n \u003cp\u003e271\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e257\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 68px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRMSprop\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.8773\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.8773\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e0.8934\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 51px;\"\u003e\n \u003cp\u003e0.8795\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 86px;\"\u003e\n \u003cp\u003e283\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e296\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable 6 presents a comparative performance analysis of three optimizers \u0026ndash; Adam, SGD, and RMSprop \u0026ndash; applied to the VGG16 model. Adam showcases superior performance across all metrics, achieving the highest accuracy (0.9537), F1 score (0.9528), precision (0.9542), and recall (0.9526). While RMSprop demonstrates moderate performance with an accuracy of 0.8773, SGD significantly lags behind with an accuracy of 0.5487. Notably, Adam also exhibits significantly faster training times, averaging 3 seconds per epoch, compared to 271 seconds for SGD and 283 seconds for RMSprop. Therefore, Adam emerges as the clear winner for VGG16 in this context, achieving high accuracy and requiring substantially less training time.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTABLE 7.\u003c/strong\u003e\u0026nbsp; Comparative Performance of CNN Architectures for Image (for GoogleLeNet\u003cstrong\u003e\u0026nbsp;(Inception\u003c/strong\u003e))\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"575\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eOptimizer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eF1_Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAvg Epoch\u003c/p\u003e\n \u003cp\u003eTime (s)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eInitial Epoch\u003c/p\u003e\n \u003cp\u003eTime (s)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" valign=\"top\"\u003e\n \u003cp\u003eGoogleNet (Inception)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAdam\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9680\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9670\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9673\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9672\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 77px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSGD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.9675\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.9673\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e0.9684\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.9666\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 83px;\"\u003e\n \u003cp\u003e44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e53\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 77px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRMSprop\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.8267\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.7989\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e0.8774\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.8137\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 83px;\"\u003e\n \u003cp\u003e43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 84px;\"\u003e\n \u003cp\u003e52\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eIn Table7, a comparative evaluation is provided for three optimizers \u0026ndash; Adam, SGD, and RMSprop \u0026ndash; concerning their performance on the GoogLeNet (Inception) model. Adam and SGD showcase notably high accuracy, achieving 0.9680 and 0.9675, respectively. The F1 scores, precision, and recall values for both optimizers are strikingly similar, indicating strong overall performance. However, Adam demonstrates a significant advantage in training speed, averaging only 1 second per epoch compared to SGD\u0026apos;s 44 seconds. RMSprop attains a respectable accuracy of 0.8267, but it is outperformed by both Adam and SGD, other performance metrics, and training speed (43 seconds per epoch). Therefore, for the GoogLeNet model, Adam emerges as the most efficient optimizer, providing comparable accuracy to SGD but with significantly faster training times.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTABLE 8.\u003c/strong\u003e\u0026nbsp; \u0026nbsp;Comparative Performance of CNN Architectures for Image (for\u0026nbsp;\u003cstrong\u003eDenseNet121\u003c/strong\u003e)\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"590\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eOptimizer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eF1_Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAvg Epoch\u003c/p\u003e\n \u003cp\u003eTime (s)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eInitial Epoch\u003c/p\u003e\n \u003cp\u003eTime (s)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" valign=\"top\"\u003e\n \u003cp\u003eDenseNet121\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAdam\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9680\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9670\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9680\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9665\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 77px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSGD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.9242\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.9222\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e0.9231\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.9235\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 76px;\"\u003e\n \u003cp\u003e76\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e82\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 77px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRMSprop\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.6679\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.6470\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e0.7940\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.6866\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 76px;\"\u003e\n \u003cp\u003e75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 92px;\"\u003e\n \u003cp\u003e87\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable 8 provides a comparative assessment of three optimization algorithms \u0026mdash; Adam, SGD, and RMSprop \u0026mdash; applied to the DenseNet121 model. Adam achieves the highest accuracy (0.9680), F1 score (0.9670), precision (0.9680), and recall (0.9665), indicating superior overall performance. SGD follows with an accuracy of 0.9242, demonstrating commendable performance but trailing behind Adam. Conversely, RMSprop exhibits notably lower performance with an accuracy of 0.6679, suggesting potential inadequacy for this model. Regarding training time, Adam shows the shortest average epoch time (2 seconds) compared to SGD (76 seconds) and RMSprop (75 seconds). This analysis confirms Adam as the most efficient optimizer for DenseNet121, achieving the highest accuracy with the shortest training duration.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTABLE 9.\u003c/strong\u003e\u0026nbsp; Comparative Performance of CNN Architectures for Image (for \u003cstrong\u003eMobileNetV2\u003c/strong\u003e)\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"590\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eOptimizer\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eF1_Score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAvg Epoch\u003c/p\u003e\n \u003cp\u003eTime (s)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eInitial Epoch\u003c/p\u003e\n \u003cp\u003eTime (s)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eMobileNetV2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAdam\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9573\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9565\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9590\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.9581\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 77px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSGD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.2310\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 66px;\"\u003e\n \u003cp\u003e0.1014\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 64px;\"\u003e\n \u003cp\u003e0.1823\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.2542\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 76px;\"\u003e\n \u003cp\u003e18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eRMSprop\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.7906\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.7744\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.8558\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e0.7834\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e23\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable 9 provides a comparative analysis of three optimization algorithms (Adam, SGD, and RMSprop) applied to the MobileNetV2 model. Adam emerges as the top performer among the optimizers, achieving a high accuracy of 0.9573, an F1 Score of 0.9565, a precision of 0.9590, and a recall of 0.9581. Notably, Adam also demonstrates the shortest training time, averaging just 1 second per epoch. In contrast, RMSprop exhibits moderate performance with an accuracy of 0.7906, while SGD lags significantly behind with an accuracy of only 0.2310, suggesting its inadequacy for this model. These results underscore Adam\u0026apos;s effectiveness as the optimal optimizer for MobileNetV2, delivering superior accuracy and markedly faster training times compared to SGD and RMSprop.\u003c/p\u003e\n\u003cp\u003eThis comparison highlights the importance of empirically evaluating various optimization algorithms for CNN architectures and datasets. Across all four models (VGG16, GoogLeNet, DenseNet121, and MobileNetV2), the Adam optimizer consistently achieved the highest accuracy and fastest training times.\u003c/p\u003e\n\u003cp\u003eBased on these results, the remaining evaluations of all four models were conducted using the Adam optimizer. This choice enables a consistent and efficient assessment of the models\u0026apos; performance on the flank wear classification task.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eⅣ.2 Analysing Confusion Matrices\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo gain a deeper understanding of each CNN model\u0026apos;s performance, the confusion matrices generated during testing were examined. These matrices reveal both strengths and weaknesses by illustrating the model\u0026apos;s ability to correctly classify instances within each category.\u003c/p\u003e\n\u003cp\u003eAll models exhibit strong overall performance, as indicated by the dominant diagonals in their confusion matrices. However, each model demonstrates a tendency to misclassify some instances between Class 2 and Class 3, suggesting a potential limitation in distinguishing these specific categories.\u003c/p\u003e\n\u003cp\u003eThis observation stems from the close proximity of these wear types in their VB (flank wear) value ranges. Specifically, moderate wear is characterized by VB values within the interval [100 \u0026micro;m \u0026ndash; 180 \u0026micro;m], while severe wear falls within [180 \u0026micro;m \u0026ndash; 270 \u0026micro;m]. As a result, when VB values approach the 180 \u0026micro;m threshold, it becomes a potential source of confusion between these two wear types. This phenomenon is illustrated in Figure 8.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eⅣ.3 Training and validation accuracy analysis\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 9 displays the training and validation accuracy curves of our four convolutional neural network models across training epochs, offering valuable insights into their learning dynamics and generalization performance.\u003c/p\u003e\n\u003cp\u003eWhile VGG16 achieves a respectable 0.9537 validation accuracy, its fluctuating curves hint indicating some level of overfitting, suggesting the model is learning the training data well, but it\u0026apos;s not generalizing well to unseen data. In contrast, GoogLeNet (Inception) exhibits more stable performance, with both training and validation accuracy steadily increasing, indicating strong generalization ability. MobileNetV2, reaching a validation accuracy of 0.9573, demonstrates a similar learning pattern to DenseNet121, although with slightly lower overall performance. Both models show steady increases in training and validation accuracy, but a slight dip in validation accuracy near the end in both models suggest a degree of overfitting.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eⅣ.4 Model loss analysis during training\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 10 visualizes the training and validation loss curves of four CNN models\u0026mdash;VGG16, GoogleNet (Inception), DenseNet121, and MobileNetV2\u0026mdash;over the training epochs, illustrating each model\u0026apos;s training progression and performance.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eVGG16 exhibits a steady decrease in both training and validation loss across the epochs. However, a slight gap between the training and validation loss emerges towards the end, hinting at potential overfitting. GoogleNet shows a rapid initial decrease in training loss followed by stabilization around a relatively low value, The small gap between the training and validation loss indicates that the model is generalizing well to unseen data. DenseNet121 demonstrates a decreasing trend in both losses initially, but the validation loss fluctuates after a few epochs, indicating effective learning on training data but struggles with generalization. The increasing gap between training and validation loss after the initial drop supports this notion, suggesting overfitting. Lastly, MobileNetV2 exhibits rapid initial decrease in both losses, followed by stabilization at low values, a slight uptick in validation loss towards the end of training might signal the beginning of overfitting, potentially leading to decreased performance on unseen data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eIV.5 HIL-CNN Integration with DenseNet121\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo overcome the limitations of fully automated classification in cases of uncertain or transitional tool wear conditions\u0026mdash;particularly between neighboring wear levels such as moderate and heavy we introduce a Human-in-the-Loop (HIL) collaborative model integrated into the DenseNet121-based convolutional framework. This hybrid approach enables a dynamic interplay between machine intelligence and human expertise, aiming to improve the reliability of decisions in practical industrial settings. Initially, each image is evaluated by a fine-tuned DenseNet121 model, which outputs both a predicted class and a confidence score from the softmax layer. If this score falls below a pre-established threshold (e.g., 80%) or if the predicted wear severity falls within a critical transition zone (e.g., between 180 \u0026micro;m and 200 \u0026micro;m VB), the system flags the case for human review. A Gradio-powered interface facilitates this process by allowing operators to visualize the image and either validate or correct the model\u0026rsquo;s prediction in real-time. Corrections made by the expert are systematically stored and can later be reintegrated into the dataset to support incremental or continual training, enhancing the model\u0026rsquo;s adaptability. This HIL mechanism minimizes high-impact classification errors, reinforces the interpretability of AI outputs, and fosters confidence in automated decisions particularly under the variable and high-stakes conditions characteristic of Industry 4.0 environments. Ultimately, the approach supports intelligent tool condition monitoring and predictive maintenance strategies, where precision, accountability, and adaptability are essential.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eIV.6 Identified Research Gaps and Future Perspectives\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDespite the promising results achieved in this study, several critical research gaps remain to be addressed to advance the practical deployment and scientific depth of tool wear classification systems. First, the current CNN-based models operate as black boxes, offering limited interpretability of their predictions. This hinders the ability of end-users, such as process engineers, to trust or validate the system\u0026rsquo;s decisions. Second, while classification performance was evaluated on static image datasets, the model\u0026rsquo;s applicability in real-time industrial settings remains untested. The impact of latency, sensor integration, and real-time decision-making under actual machining conditions requires further investigation. Additionally, the confusion matrix reveals significant ambiguity between adjacent wear classes particularly moderate and heavy wear highlighting the need for probabilistic modeling or soft classification strategies to handle transitional states more effectively. Furthermore, although inference time was assessed, the suitability of the models for deployment on resource-constrained edge devices (e.g., Raspberry Pi or smart industrial cameras) was not fully explored. Lastly, the temporal nature of wear progression was not captured, as the study focused on isolated images. Future research could benefit from incorporating sequential learning architectures, such as CNNs combined with LSTM or Transformer networks, to track and predict wear evolution over time in a more robust and informed manner.\u003c/p\u003e"},{"header":"V. CONCLUSION","content":"\u003cp\u003eThis study presented an approach for monitoring flank wear on cutting tools using deep convolutional neural networks (CNNs) and transfer learning techniques. By leveraging pre-trained CNN architectures (VGG16, GoogleNet, DenseNet121, and MobileNetV2) originally trained on the ImageNet dataset and subsequently fine-tuned on the tool wear image dataset, promising results were achieved in classifying flank wear into four distinct categories based on wear severity. The findings indicate that:\u003c/p\u003e\n\u003cul class=\"decimal_type\"\u003e\n \u003cli\u003eThis study effectively utilized four CNN architectures (VGGNet, GoogleNet, DenseNet121, and MobileNetV2) with transfer learning to classify cutting tool flank wear into four categories.\u003c/li\u003e\n \u003cli\u003eData augmentation techniques proved crucial in enhancing model robustness and mitigating overfitting.\u003c/li\u003e\n \u003cli\u003eAdam optimizer consistently outperformed SGD and RMSprop, achieving higher accuracy and faster training for all tested CNN models.\u003c/li\u003e\n \u003cli\u003eGoogleNet and DenseNet121 achieved the highest classification accuracy (96.80%), outperforming VGG16 and MobileNetV2.\u003c/li\u003e\n \u003cli\u003eGoogleNet is preferred over DenseNet121 due to its significantly better training performance.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eFuture research could explore Expanding the dataset to address misclassification between specific wear categories, further improving the system\u0026apos;s accuracy. Investigating other CNN architectures and transfer learning techniques for potential performance improvements. Finally, Implementing the system in real-time machining environments for continuous tool wear monitoring.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003ecorresponding author\u003c/strong\u003e: ELBAH Mohamed \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;E-mail address: [email protected]\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements:\u003c/strong\u003e This work was achieved in the Mechanics and Structures Laboratory (LMS) (Guelma University, Algeria). The authors would like to thank the Algerian Ministry of Higher Education and Scientific Research (MESRS).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e This study was funded by the Algerian Ministry of Higher Education and Scientific Research\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFinancial interests:\u003c/strong\u003e The authors declare they have no financial interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests:\u003c/strong\u003e The authors have no competing interests to declare that are relevant to the content of this article.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions:\u003c/strong\u003e The study was carried out with the contribution of all the authors. Material preparation and data collection were performed by Oussama Dahmoune, Ikhlas Meddour, Mohamed Atmane Yalles, and Mohamed Elbah. Data analysis was performed by Oussama Dahmoune, Mohamed Elbah and Ikhlas Meddour. Python programming was performed by Oussama Dahmoune and Mohamed Elbah. The first draft of the manuscript was prepared by Oussama Dahmoune. Mohamed Elbah and Ikhlas Meddour reviewed it, and wrote the final manuscript. All authors agreed the final manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eVetrichelvan G, Sundaran S, Kumaran SS, Velmurugan P (2014) An investigation of tool wear using acoustic emission and genetic algorithm. J Vib Control 21(15) : 3061\u0026ndash;3066\u003c/li\u003e\n\u003cli\u003eEzugwu, E.O., Wang, Z.M., Machado, A.R., 1999. The machinability of nickel-based alloys: a review. Journal of Materials Processing Technology 86 (1-3), 1\u0026ndash;16, https://doi.org/10.1016/S0924-0136(98)00314-8. \u003c/li\u003e\n\u003cli\u003eWang, B., Liu, Z., 2018. Influences of tool structure, tool material and tool wear on machined surface integrity during turning and milling of titanium and nickel alloys: a review. Int J Adv Manuf Technol 98 (5-8), 1925\u0026ndash;1975.\u003c/li\u003e\n\u003cli\u003eJeon, J.U., Kim, S.W., 1988. Optical flank wear monitoring of cutting tools by image processing. Wear 127 (2), 207\u0026ndash;217\u003c/li\u003e\n\u003cli\u003eY.Q. Zhou, B.T. Sun, W.F. Sun, Z. Lei, Tool wear condition monitoring based on a two-layer angle kernel extreme learning machine using sound sensor for milling process, J. Intell. Manuf. (2020), https://doi.org/10.1007/s10845-020-01663-1\u003c/li\u003e\n\u003cli\u003eAvinash, C.; Raguraman, S.; Ramaswamy, S.; Muthukrishnan, N. An Investigation on Effect of Workpiece Reinforcement Percentage on Tool Wear in Cutting Al-SiC Metal Matrix Composites. In Proceedings of the ASME International Mechanical Engineering Congress and Exposition, Seattle, WA, USA, 11\u0026ndash;15 November 2008; pp. 561\u0026ndash;566.\u003c/li\u003e\n\u003cli\u003eEe, K.C.; Balaji, A.K.; Jawahir, I.S. Progressive tool-wear mechanisms and their effects on chip-curl/chip-form in machining with grooved tools: An extended application of the equivalent toolface (et) model. Wear 2003, 255, 1404\u0026ndash;1413. \u003c/li\u003e\n\u003cli\u003eNordgren, A.; Melander, A. Tool wear and inclusion behaviour during turning of a calciumtreated quenched and tempered steel using coated cemented carbide tools. Wear 1990, 139, 209\u0026ndash;223.\u003c/li\u003e\n\u003cli\u003eJunior, Joseph KalilKhoury, et al. \u0026quot;Monitoring of flank wear and damage on turning cutting tools by image processing.\u0026quot; The Journal of Engineering and Exact Sciences 6.2 (2020): 0098-0106.\u003c/li\u003e\n\u003cli\u003eMiao, Huihui, et al. \u0026quot;A U-Net-based approach for tool wear area detection and identification.\u0026quot; IEEE Transactions on Instrumentation and Measurement 70 (2020): 1-10.\u003c/li\u003e\n\u003cli\u003ePanda, A., \u0026amp; Kumar, R. (2022, May). On Tool Wear Prediction Using Artificial Neural Network and Regression Methodology During. In Proceedings of 2nd International Conference on Smart Computing and Cyber Security: Strategic Foresight, Security Challenges and Innovation (SMARTCYBER 2021) (Vol. 395, p. 315). Springer Nature.\u003c/li\u003e\n\u003cli\u003eAgrawal, K., Panda, A., \u0026amp; Sahoo, A. K. (2024, August). Advancements in tool wear monitoring in turning operations: digital image processing and AI techniques. In Journal of Physics: Conference Series (Vol. 2818, No. 1, p. 012040). IOP Publishing.\u003c/li\u003e\n\u003cli\u003eBrili, Nika, MirkoFicko, and Simon Klančnik. \u0026quot;Automatic identification of tool wear based on thermography and a convolutional neural network during the turning process.\u0026quot; Sensors 21.5 (2021): 1917. \u003c/li\u003e\n\u003cli\u003eSun, Wei-Heng, and Syh-ShiuhYeh. \u0026quot;Using the machine vision method to develop an on-machine insert condition monitoring system for computer numerical control turning machine tools.\u0026quot; Materials 11.10 (2018): 1977\u003c/li\u003e\n\u003cli\u003eLim, Meng Lip, et al. \u0026quot;Tool wear prediction in turning using workpiece surface profile images and deep learning neural networks.\u0026quot; The International Journal of Advanced Manufacturing Technology 120.11-12 (2022): 8045-8062.\u003c/li\u003e\n\u003cli\u003eKumar, M. Phani, Samik Dutta, and N. C. Murmu. \u0026quot;Tool wear classification based on machined surface images using convolution neural networks.\u0026quot; Sādhanā 46 (2021): 1-12.\u003c/li\u003e\n\u003cli\u003eAmbadekar, P. K., and C. M. Choudhari. \u0026quot;CNN based tool monitoring system to predict life of cutting tool.\u0026quot; SN Applied Sciences 2 (2020): 1-11.\u003c/li\u003e\n\u003cli\u003eDahmoune, O. Meddour, I. Elbah, M. Atmane Yallese, M. \u0026amp; Belhadi, S. (2025). Development of an adaptive tool condition monitoring system: integration of case-based reasoning with CNN. Journal of Intelligent Manufacturing, 1-14.\u003c/li\u003e\n\u003cli\u003eA. Rehman, S. Naz, M.I. Razzak, H.A. Ibrahim, Automatic visual features for writer identification: a deep learning approach. IEEE Access 7, 17149\u0026ndash;17157 (2019)\u003c/li\u003e\n\u003cli\u003eG. Taguchi, S. Chowdhury, and S. Taguchi, Robust Engineering: Learn How to Boost Quality While Reducing Costs \u0026amp; Time to Market. New York, NY, USA: McGraw-Hill, 2000.\u003c/li\u003e\n\u003cli\u003eMikołajczyk, T., et al. \u0026quot;Predicting tool life in turning operations using neural networks and image processing.\u0026quot; Mechanical systems and signal processing 104 (2018): 503-513\u003c/li\u003e\n\u003cli\u003eISO 3685 Tool Life Testing with Single-Point Turning Tools. 1993. Available online: https://www.iso.org/fr/standard/9151.html (accessed on 7 January 2024).\u003c/li\u003e\n\u003cli\u003eL. Perez and J. Wang, \u0026ldquo;The effectiveness of data augmentation in image classification using deep learning,\u0026rdquo; 2017, arXiv:1712.04621. [Online]. Available: https://arxiv.org/abs/1712.04621\u003c/li\u003e\n\u003cli\u003eA. Hern\u0026aacute;ndez-Garc\u0026iacute;a and P. K\u0026ouml;nig, \u0026ldquo;Data augmentation instead of explicit regularization,\u0026rdquo; Jun. 2018, arXiv:1806.03852. Accessed: Sep. 4, 2020. [Online]. Available: http://arxiv.org/abs/1806.03852\u003c/li\u003e\n\u003cli\u003eK. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition. ArXiv preprint: arXiv:1409.1556 (2014) \u003c/li\u003e\n\u003cli\u003eC. Szegedy,W.Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, A. Rabinovich, Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1\u0026ndash;9 (2015)\u003c/li\u003e\n\u003cli\u003eNandhini, S., and K. Ashokkumar. \u0026quot;An automatic plant leaf disease identification using DenseNet-121 architecture with a mutation-based henry gas solubility optimization algorithm.\u0026quot; Neural Computing and Applications 34.7 (2022): 5513-5534.\u003c/li\u003e\n\u003cli\u003eDong, Ke, et al. \u0026quot;MobileNetV2 model for image classification.\u0026quot; 2020 2nd International Conference on Information Technology and Computer Application (ITCA). IEEE, 2020.\u003c/li\u003e\n\u003cli\u003eMiao, Huihui, et al. \u0026quot;A U-Net-based approach for tool wear area detection and identification.\u0026quot; IEEE Transactions on Instrumentation and Measurement 70 (2020): 1-10.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"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":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Flank Wear, Turning process, Deeplearning, Transfer learning, GoogleNet, DenseNet12, MobileNetV2","lastPublishedDoi":"10.21203/rs.3.rs-7189226/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7189226/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eFlank wear in cutting tools remains a critical issue in the metal cutting industry due to its direct impact on dimensional precision, surface integrity, and overall manufacturing efficiency. Failure to detect or accurately classify wear may result in premature tool replacement or extended use of worn tools, leading to increased operational costs, heat accumulation, machining vibrations, and potential damage to workpieces and equipment. This study provides a comparative analysis of four state-of-the-art deep convolutional neural network (CNN) architectures VGG16, GoogleNet, DenseNet121, and MobileNetV2 for the automatic classification of flank wear across four predefined severity levels. To overcome dataset limitations and improve model adaptability, transfer learning was employed by fine-tuning pre-trained models on a domain-specific dataset of tool wear images captured using a DM9 industrial digital microscope under dry turning conditions. A comprehensive data augmentation strategy was also applied to enhance model generalization and mitigate overfitting. Experimental results show that GoogleNet and DenseNet121 achieved the highest classification accuracy of 96.80%, with GoogleNet being favored for its shorter training time and computational efficiency. These findings underscore the effectiveness of deep learning-based approaches for reliable, scalable, and real-time tool condition monitoring in advanced manufacturing environments\u003c/p\u003e","manuscriptTitle":"Deep Transfer Learning and Data Augmentation for Automatic Flank Wear Detection and Classification","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-29 07:53:21","doi":"10.21203/rs.3.rs-7189226/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"3555175b-c766-4c3a-8faf-a037c891b9e5","owner":[],"postedDate":"July 29th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-10-21T12:21:34+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-29 07:53:21","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7189226","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7189226","identity":"rs-7189226","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}