Predictive Maintenance for Offshore Wind Turbines through Deep Learning and Online Clustering of Unsupervised Subsystems: A Real-World Implementation

Predictive Maintenance for Offshore Wind Turbines through Deep Learning and Online Clustering of Unsupervised Subsystems: A Real-World Implementation | 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 Predictive Maintenance for Offshore Wind Turbines through Deep Learning and Online Clustering of Unsupervised Subsystems: A Real-World Implementation Uwe Lützen, Serdar Beji This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3906932/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 15 Jun, 2024 Read the published version in Journal of Ocean Engineering and Marine Energy → Version 1 posted 10 You are reading this latest preprint version Abstract Enterprises in increasing numbers allocate substantial expenses to offshore wind energy development as a pivotal component of the global energy transition from fossil fuels, hence the importance of ensuring the reliability of offshore wind technology becomes ever more significant. At the same time, operation and maintenance (O&M) of offshore wind farms are progressively focusing on the integration of artificial intelligence (AI) for enhancing the efficiency and performance of the wind energy facilities. Decision support strategies based on failure predictions are an important element in this trend. As a result, AI is more frequently used to create time-to-failure predictions based on large amount of data collected from sensors deployed to wind turbines. Nevertheless, unsupervised components or subsystems may occasionally lead to failures. This paper presents a real-life example that failures in unsupervised components can be reliably predicted by the use of AI. Two different methods, Support Vector Machine and Long Short Term Memory, are presented and their limitations and advantages discussed. Offshore wind energy Predictive maintenance Condition monitoring Prediction period unsupervised components asset degradation predictive asset degradation patterns Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 1. Introduction In view of escalating global pursuit of offshore wind energy resources as an important element in the transition from fossil fuels to renewable energy sources, ensuring the reliability of offshore wind technology becomes increasingly obvious. Consequently, entities responsible for the operation and maintenance (O&M) of offshore wind farms are progressively raising their standards by employing artificial intelligence (AI) solutions for improving performance and productivity of these vital installations. Accordingly, increased interest in AI integration has prompted a surge in corporate transactions, notably marked by acquisitions of machine learning and AI sectors as O&M companies strategically position themselves to harness the potential of AI to enhance output, sustainability and feasibility of offshore wind farms. In that context, predictive maintenance is quite an important tool for increasing operational availability of offshore wind turbines. Availability of a typical offshore wind turbine can be significantly lower than its onshore counterpart which has a typical availability rate of 95–97% (Cevasco et al. 2021 ). Lower availability rates necessitate new approaches to optimize O&M strategies for offshore wind farms to significantly increase the feasibility and grid stability. The need for new approaches to O&M strategies for offshore wind energy was recognized quite early about two decades ago, as Rademakers et al. ( 2003 ) drew attention to this point during a wind energy conference. In line with this concern Scheu et al. ( 2012 ) developed a simulation for the operational phase of offshore wind farms with a focus on modeling of failures and repair. However, until recently, optimization studies concerning the offshore wind O&M problems were mainly focused on specific components as the root cause of failures, as in Karyotakis ( 2011 ) who focused on wind turbine inverters as one of the main causes for the interruption of operation. Several factors, such as weather conditions, different component degradation and wear, influence the selection process for the optimal maintenance strategy. Therefore, decision support strategies (DSS) are important aids towards an informed decision. Kessler ( 2013 ) presented several DSS in detail; namely, data-driven, communication-driven, document-driven, knowledge-driven, and model-driven DSS. More specific works concerning decision support models and their development for offshore wind energy O&M strategies are also available. A notable work is due to Hofmann ( 2011 ) who review 49 different commercial as well as non-commercial decision support models focusing mainly on O&M but also including logistics, power production and overall project costs. Papatzimos (2020) gives a decision support framework for maintenance planning of a single wind turbine based on data management that builds a very good basis for extension to wind farms. Endrerud et al. ( 2014 ) develop a marine logistic simulation model for the O&M lifecycle phase. Dinwoodie et al. ( 2013 , 2015 ) present the development of a strategic decision support model for operation of wind farms and simulation models for O&M. A comprehensive review on the state of art in offshore wind farm O&M with a focus on decision support models for the scheduling of maintenance is given by Seyr and Muskulus ( 2019 ). Assessment of the input factors and their impact on offshore wind O&M are an important subjects of interest. In this rather broad area, Feuchtwang and Infield ( 2013 ) consider the delays for repair and maintenance caused by the sea, Douard et al. ( 2012 ) and Ioannou et al. ( 2019 ) treat the risk measurement indicators while Yu ( 2016 ) shows the benefits of condition monitoring. Pandit et al. ( 2023 ) present a review of predictive techniques to support decision making for O&M of wind turbines focusing on condition monitoring for main components. Optimization of a fleet of vessels is assessed by Halvorsen et al. ( 2013 ). Stalhane et al. (2015) and Lazakis and Kahn ( 2021 ) look into optimal routes and schedules. Sperstad et al. ( 2017 ) consider the vessel fleet optimization in relation to O&M cost, electricity price and vessel transit speed. Machine learning (ML) plays a pivotal role for handling large amount of data collected by the wind turbine controller and/or additional systems like condition monitoring system for supporting predictive maintenance. Perez Granados et al. ( 2023 ) present a methodology for predictive maintenance of wind turbines by using vibration monitoring as input. Tang et al. ( 2020 ) employ an improved lightGBM algorithm for online fault detection in gearboxes of wind turbines. Farrar et al. ( 2023 ) give an overview of the use of artificial intelligence and ML in grid connected wind turbine control systems. An important aspect of the application of ML is the feature selection, which is studied by Marti-Puig et al. ( 2019 ) for finding the optimal predictors via different feature selection algorithms. Quite recently, Masoumi ( 2023 ) has presented a comprehensive review of machine learning applications for offshore wind farms. Influence of different O&M strategies on the sustainability of a wind farm is treated in Nivet and Muk-Pavik ( 2015 ). Overall situation of offshore wind energy is laid out by the International Energy Agency ( 2019 ) as well as the GWEC ( 2020 ) by describing an outlook for the coming years with an emphasis on that offshore wind energy will become more and more important part of the energy mix. Rusu and Onea ( 2023 ) gives a review of the state of offshore wind energy sector in Europe in connection with the climate change. The above review of the relevant literature reveals that the significance of predictive maintenance for wind turbine systems, particularly in the context of offshore wind energy generation, has grown considerably. The present work endeavors to make a contribution to the field by introducing and simulating an innovative approach for the predictive maintenance of wind turbines. The central focus of this work lies in the design, development, and practical application of the prediction of failures in unsupervised sub-systems in a real-life wind turbine system, specifically parts of the yaw system, utilizing data-driven approaches to validate the proposed methodology. In particular, the present methodology leverages deep learning models in conjunction with online clustering techniques, thus establishing a robust foundation for predictive maintenance in offshore wind energy systems. Accordingly, a comprehensive framework, methodology, experimental setup, and results are presented here for giving new contributions and suggesting insights into the evolving landscape of predictive maintenance of offshore wind turbines. Real data of over six months have been taken from a 3MW direct-drive wind turbine for this work. The yaw system of this wind turbine experiences an increasing number of unintended slipping events where the yaw system is not able to station keep the wind turbine due to unsupervised wear out of a component. Section 2 is dedicated to elucidating the categorization of the data collected from turbines. Also, a rigorous evaluation of the selected methodology, which places significant reliance on the utilization of deep learning models, is provided. This analysis serves to underscore the rationale behind the methodological choices. Section 3 describes the formulation of the problem and elucidation of underlying assumptions pivotal to predictive maintenance strategies, specifically the inclusion of unsupervised sub-systems and components into predictive maintenance strategies. A detailed description of these considerations serves as the bedrock for the subsequent development of prediction strategies. Section 4 presents the results obtained through the application of the selected predictive strategy for unsupervised sub-systems. These results are scrutinized in detail to offer a comprehensive assessment of their implications and significance within the broader context of predictive maintenance for offshore wind turbines. Section 5 , the final section, encapsulates the key findings stemming from the application of the proposed predictive maintenance methodology. Additionally, insightful recommendations for prospective areas of research and development are provided thereby delineating a roadmap for future endeavors in this particular research field. 2. Applied methodology As already indicated, offshore wind turbine maintenance operations reveal a pressing need for more accurate maintenance strategies to optimize operational efficiency and reduce downtime. These strategies are highly dependent on accurate data and precise forecast of expected time to failure (TTF) since a fair forecasting period; namely, a sufficient duration is needed in order to plan maintenance operations early enough to prevent downtimes. As offshore wind farms are only accessible at certain weather conditions, the forecasting period and warnings of upcoming expected problems must be long enough to take adverse weather conditions into account. Achieving a fair or long-enough forecasting period requires advanced engineering approaches. With this viewpoint in mind this section describes the integration of data clustering techniques to Support Vector Machine (SVM) and Long-Short-Term Memory (LSTM) networks for forecasting. This approach offers a novel perspective in the context of offshore wind turbines by emphasizing the critical role of feature engineering in harnessing the full potential of deep learning models. By effectively grouping similar data points through clustering, it is possible to capture the underlying patterns and dependencies within the intricate operational data of offshore wind turbines. This not only improves the model's ability to learn from the data but also enhances its interpretability, enabling better informed maintenance decisions. What sets the proposed contribution apart is the emphasis on leveraging experimental data for training and testing the predictive maintenance model. In the offshore wind energy sector, where real-world data is often characterized by high-dimensional and noisy observations, the utilization of experimental data becomes paramount. Experimental data provides a controlled environment to validate the effectiveness of the clustering framework, ensuring its applicability and reliability in the challenging offshore conditions. 2.1 Data Preparation Data processing plays a crucial role in machine learning (ML) for the maintenance of wind turbines. Basic aspects of data processing can be listed as follows (Kumar 2022). Data Collection: Wind turbines gather continuous data on parameters like the wind speed, temperature, vibration, and power output through sensors and monitoring systems. This data is collected and stored for analysis. Data Cleaning: Raw data often contains errors, missing values, or outliers. Data cleaning involves identifying and addressing these issues to ensure the data accuracy and reliability. Techniques such as imputation, outlier detection, and error correction are applied. Feature Extraction: Relevant features or variables are extracted from the raw data to capture essential information for predictive modeling. These features could include wind speed patterns, temperature fluctuations, historical maintenance records, or other parameters crucial for identifying potential issues or predicting failures. Feature Engineering: Engineers create new features from the existing data to improve the predictive power of machine learning models. This step requires expertise and domain knowledge, resulting in features like wind turbulence intensity, power output degradation rates, or health indicators based on vibration patterns. Feature Scaling and Normalization: Features are scaled to ensure they are on a similar scale. Scaling methods like normalization or standardization are used to prevent certain features from dominating the learning process and aid in model convergence. Data Integration: Data from multiple turbines may be integrated to create a comprehensive dataset. This allows for a broader analysis of patterns and correlations across turbines, leading to more accurate predictions and maintenance strategies. Data Splitting: The processed dataset is divided into training and test sets. The training set is used to train machine learning models, while the test set is used to evaluate their performance. This ensures that the models generalize well to unseen data and can effectively predict maintenance needs. The above steps are essential for preparing data for predictive maintenance in wind turbines, ensuring that machine learning models can provide accurate and reliable maintenance recommendations. By leveraging these techniques, machine learning models can be trained on high-quality, informative datasets, enabling the prediction of potential failures, identification of maintenance requirements, and optimization of maintenance schedules for wind turbines. 2.1.1 Data preprocessing and balancing of the dataset Data preprocessing is a crucial step in data preparation, involving various techniques to ready raw data for further processing. It is used not only in data mining but also in training machine learning and AI models to ensure accuracy. Data balancing, on the other hand, addresses class imbalance issues within a dataset. Class imbalance occurs when one class significantly outnumbers another, which can lead to challenges in model training, especially for the minority class. Data balancing methods include over-sampling (increasing minority class instances), under-sampling (decreasing majority class instances), or hybrid approaches that combine both. The choice of technique depends on the dataset and problem requirements, but caution is needed to prevent overfitting or bias. Cross-validation and independent testing help assess the impact of data balancing on model performance. The choice of data balancing technique depends on the specifics of the dataset, the degree of class imbalance, and the requirements of the problem at hand. It's important to note that data balancing should be performed carefully to avoid overfitting, loss of important information, or introducing bias into the model. Cross-validation and evaluation on independent test sets are essential to assess the impact of data balancing on model performance (He and Garcia 2009). 2.1.2 Present application: Dataset for Wind Turbine Maintenance The dataset is divided into two classes: those with detected slipping, and those without detected slipping of the yaw system. The total data exhibits a significant class imbalance with class 0 (non-slipping) comprising 968,567 instances and class 1 (slipping) consisting of only 17,832 instances. This severe imbalance poses a challenge for the model's ability to accurately detect the class 1, which represents the deteriorating fault condition leading to maintenance requirements once a predefined threshold is reached. To address this issue and ensure proper identification of the smaller data set class 1, an under-sampling technique is employed. By selectively reducing the number of instances from the bigger data set class 0, it is aimed to rebalance the dataset and improve the model's performance in detecting the crucial fault condition that will trigger the maintenance actions. 2.1.3 Feature engineering Feature engineering is a crucial phase in the data processing workflow that focuses on transforming raw data into meaningful features, enabling more efficient training of data models and facilitating accurate inferences. It encompasses a range of techniques aimed at extracting, selecting, and transforming data attributes to enhance the performance and interpretability of machine learning algorithms. During feature engineering, domain knowledge and statistical analysis are employed to derive new features that capture relevant information and patterns in the data. This may involve techniques such as (Khalid et al. 2014) Feature extraction: Converting raw data into more informative representations. For example, extracting statistical measures (e.g., mean, variance) from time-series data or deriving textual features (e.g., word frequency, TF-IDF [1] ) from unstructured text. Feature selection: Identifying the most relevant features that have the greatest impact on the target variable. This can be done through methods like correlation analysis, statistical tests, or regularization techniques (e.g., L1 regularization). Feature transformation: Modifying the scale, distribution, or structure of features to meet modeling assumptions or improve performance. Common techniques include scaling features to a specific range (e.g., normalization, standardization), applying mathematical functions (e.g., logarithmic, polynomial transformations), or creating interaction terms. Feature encoding: Converting categorical variables into numerical representations suitable for modeling. This can involve techniques like one-hot encoding, label encoding, or ordinal encoding. Effective feature engineering enables models to capture relevant patterns and relationships in the data, improving their predictive power and interpretability. It also helps mitigate issues such as overfitting, and the curse of dimensionality. It is essential to approach feature engineering iteratively, evaluating the impact of engineered features on model performance using appropriate validation techniques. By iteratively refining the features, the data can be organized in ways that maximize the efficiency of data models and facilitate robust inferences (Guyon and Elisseeff 2003). For the present application the feature selection led to the following descriptions (variable names): WTURTurSt: Wind turbine status, with following values: 1 Idling, waiting for wind, ready for startup, no fault 4 Start up sequence started, self testing 6 start up sequence, connecting the inverter to the grid 8 Turbine in operation (no power limitation) 9 Turbine in operation (power limitation, value preset by operator) 21 Manual stop 23 Fault 24 Emergency stop WNACWdSpdFilVal: Wind speed measured on top of the wind turbine nacelle, filtered value averaged from two anemometers (m/s) WTURWActVal: Wind turbine actual generator power (kW) WNACDrillPosActVal: Absolute yaw position of the RNA (Rotor Nacelle Assembly) in ° WROTRotSpdActVal: The speed of the rotor (RPM) WYAWMot1On: Yaw motors activated (TRUE) or not activated (FALSE) 2.1.4 Feature scaling or normalization Frequently, it is observed that various variables undergo alterations across disparate scales, wherein one variable exhibits a linear progression while another variable demonstrates an exponential trend. To illustrate, remuneration may be quantified in terms of thousands of dollars, whereas age is typically denoted by two-digit figures. Employing scaling techniques facilitates the transformation of such data in a manner that simplifies the discernment of significant associations between variables by algorithms. The use of normalization is very influential in the application of SVM. Normalization can shorten the learning process and improve the performance; however, it needs to be applied correctly and therefore should be applied mutual to all features selected (Setiawan et al. 2019). The following scaling and normalizations have been applied: Normalization of variables to the interval [0,1] Transform ‘TRUE’ and ‘FALSE’ to 1 and 0 of 'WYAWMot1On' feature so as to be compatible with the data model Create events based on given conditions: WTURTurSt = 8 (or 9) WNACWdSpdFilVal > 7m/s WTURWActVal > 0 kW WYAWMot1On = False WNACDrillPosActVal change of value to the previous data set of typically 0.00675° In case all conditions are fulfilled, the dataset belongs to class 1, otherwise to class 0. Apply under-sampling to address class imbalance in the dataset. 2.2 Methodology for maintenance 2.2.1 Support Vector Machines (SVMs) Support vector machine (SVM) classifiers is an effective data analysis method, which is based on the structural risk minimization principle solving a quadratic programming problem (Cortes and Vapnik 1995). By applying the kernel function, the method can map the data into a higher dimensional input space and construct an optimal hyperplane (Kang et al. 2020). Support Vector Machines (SVMs) are primarily used for classification tasks, but they can also be applied to maintenance systems in various ways (Setiawan et al. 2019). Here are a few examples of how SVMs can be used in maintenance systems: 1. Fault detection and diagnosis: SVMs can be utilized to detect and diagnose faults in complex systems. By training an SVM with labeled data representing normal and faulty system behavior, the model can learn to classify new instances as either normal or faulty. This helps in identifying potential maintenance issues early on and triggering appropriate actions. 2. Anomaly detection: SVMs can be employed for anomaly detection in maintenance systems. By training the SVM on a dataset of normal system behavior, the model can learn to identify deviations from the norm. Any instances that significantly differ from the learned patterns can be flagged as anomalies, indicating potential maintenance needs. 3. Predictive maintenance: SVMs can be used to predict the remaining useful life (RUL) of equipment or components as done in this work. By training an SVM with historical data that includes information about the condition of the equipment and the time until failure, the model can learn to predict the RUL of new instances. This information helps maintenance teams plan proactive maintenance actions, minimizing downtime, maximizing equipment lifespan and increase the output. 4. Equipment health monitoring: SVMs can be applied to monitor the health of equipment by analyzing sensor data. For instance, in wind turbine components such as bearings or rotor blades, sensors might capture measurements such as temperature, vibration, or pressure that will be collected by a condition monitoring system (CMS). By training an SVM on this CMS data, it can learn to classify equipment conditions, such as normal, early degradation, or critical failure. This enables maintenance teams to take timely actions based on the SVM's predictions. It is important to note that SVMs are just one of many machine learning algorithms that can be used in maintenance systems. The choice of algorithm depends on the specific requirements of the problem at hand, the available data, and the desired outcome. Kernels In Support Vector Machines (SVMs), kernels play a crucial role by allowing the model to operate effectively in high-dimensional feature spaces without explicitly computing the coordinates of data points in those spaces (Schölkopf and Smola 2002). Kernels are functions that measure the similarity between two input vectors, working in the original input space but implicitly mapping data into a higher-dimensional feature space. This transformation can make it easier to separate classes using linear decision boundaries. SVMs can efficiently perform complex nonlinear classifications using kernel functions, enabling them to capture intricate data patterns and relationships (Cristianini and Shawe-Taylor 2000). Commonly used kernel functions include the Linear Kernel for linear similarity, Polynomial Kernel for introducing polynomial combinations of features, Radial Basis Function (RBF) Kernel for capturing complex nonlinear relationships based on distance, and Sigmoid Kernel for learning nonlinear decision boundaries. The choice of kernel depends on the data and problem characteristics to ensure SVMs can effectively capture underlying patterns. The choice of kernel depends on the nature of the data and the problem at hand. Selecting an appropriate kernel is important to ensure that SVMs can effectively capture the underlying patterns and achieve good classification performance. Training In the maintenance of wind turbines, a crucial step is the analysis of data collected from various sensors and monitoring systems. To effectively analyze this data, it is important to preprocess the dataset and apply certain techniques. In the context of wind turbine maintenance, the following steps are considered: 1. Dataset Split: The first step is to split the dataset into a training set and a test set. This division helps evaluate the performance of the predictive models accurately. Typically, the dataset is divided into an 80% training set and a 20% test set. The training set is used to train the models, while the test set is used to assess their performance. For example, in case of having a dataset of 1000 wind turbine observations, 800 observations are selected for the training set and 200 observations for the test set (Schölkopf et al. 2000). 2. Feature Scaling: Feature scaling is a preprocessing technique applied to normalize or standardize the features in the dataset. It improves the optimization process by ensuring that the range of values across different features is comparable. This helps algorithms converge more quickly and avoids biased influence from features with larger scales (Platt 1999). These steps are essential for effectively training predictive models in the context of wind turbine maintenance and ensuring accurate and reliable results. For example, in case of the wind turbine dataset includes features like wind speed, temperature, and power output. Wind speed ranges from 0 to 50 m/s, temperature ranges from -20 to 40 degrees Celsius, and power output ranges from 0 to 3 MW. By applying feature scaling techniques like normalization or standardization, it will be ensured that all features are on a similar scale, such as ranging from 0 to 1 or having a mean of 0 and a standard deviation of 1. By performing feature scaling, the optimization algorithms used in maintenance tasks, such as gradient descent, can work more effectively and converge towards the minimum of the cost function more efficiently. To summarize, in wind turbine maintenance, splitting the dataset into training and test sets and applying feature scaling are essential steps for accurate analysis. These practices enable better model training and optimization, leading to improved maintenance decision-making and overall turbine performance (Pandit et al. 2023). Testing When applying Support Vector Machines (SVM) to the predictive maintenance of wind turbines, specifically focusing on the slipping event of the Rotor Nacelle Assembly (RNA) due to degradation of the yaw motor brake pads based on 6 months’ data to achieve a prediction period of 14 days, the analysis may involve the following considerations: Accuracy Assessment: The SVM model can be trained to classify instances of RNA slipping accurately. The accuracy of the model is evaluated by comparing the predicted RNA slipping status (e.g., slip or no slip) with the actual slip status obtained from maintenance records or sensor data. Confusion Matrix: A confusion matrix provides a detailed breakdown of the SVM model's predictions and the actual nacelle drift status. It helps identify the different types of classification errors made by the model, such as instances where RNA slipping was present but not detected (false negatives) or instances where no slipping was present but incorrectly classified as slipping (false positives). Receiver Operating Characteristic (ROC) Curve and Area Under the Curve (AUC): By plotting the true positive rate against the false positive rate at various classification thresholds, the SVM model's performance in detecting nacelle drift can be visualized using a ROC curve. The AUC quantifies the model's ability to distinguish between instances of drift and non-drift, with a higher AUC indicating better performance. Cross-validation: Cross-validation is crucial in assessing the SVM model's generalization capability for RNA slipping detection. The data can be split into multiple subsets, and the model is trained and evaluated on different combinations of these subsets. Cross-validation helps estimate how well the model performs on unseen wind turbine data, providing insights into its robustness and reliability. By conducting a thorough analysis of the SVM model's accuracy, confusion matrix, ROC curve, AUC, cross-validation results, and feature importance, maintenance teams can assess the model's effectiveness in detecting nacelle drift in wind turbines. This analysis helps optimize maintenance schedules, reduce downtime, and ensure the optimal performance and longevity of wind turbine assets (Marti-Puig et al. 2019). 2.2.2 Long-Short-Term Memory (LSTM) One of the most advanced models out there to forecast time series is the Long-Short-Term Memory (LSTM) Neural Network. The LSTM cell contributes to long-term memory in an even more performant way because it allows even more parameters to be learned (Lindemann et al. 2021). This makes it the most powerful Recurrent Neural Network (RNN) to do forecasting, especially when you have a longer-term trend in your data. LSTMs are one of the state-of-the-art models for forecasting at the moment. In this study, a LSTM architecture is applied with two layers of 50 units and a final dense layer for the forecast. The LSTM architecture provides a 30 time-period forecast with a lookback-period of 60 time-units. 3. System description Wind turbines have hundreds of sensors connected to the main controller and supervisory control and data acquisition (SCADA) system; however, there are components or sub-systems that are unsupervised and that can lead directly or indirectly to shut-downs of the wind turbine. For example, the yaw system is an integral component of the wind turbine control system. Its primary function is to optimize the alignment of the rotor-nacelle-assembly (RNA) with the prevailing wind direction and to counteract the natural tendency of the RNA to deviate from its position during operation. Significant forces act on the RNA as the rotor behaves akin to a massive gyroscope with an inherent offset from the center of the RNA. Consequently, gyroscopic forces exert a persistent influence, compelling the RNA to veer towards a downwind orientation, which, if left unchecked, would act as a destabilizing factor as described in Danish Wind Industry Assiciation (2003). Therefore, the yaw system is tasked with countering these gyroscopic forces, ensuring that the RNA maintains its ideal alignment with the wind direction while preventing undesired deviations. To ensure these two tasks (active positioning into the wind and station keeping) the yaw system must have sufficient driving and braking forces. A yaw system is typically composed of a number of yaw drive assemblies, consisting of electric motor–gearbox assemblies connected to a drive, located in a circle around the yaw bearing that connects the RNA to the wind turbine tower and a break that is usually either electric at the end of each yaw drive assembly or hydraulic with a break disk connected to the yaw bearing. For the case considered here a fail-safe (energized to release) break is placed behind each electric motor. Due to the gear ratio of 1 : 2157 and the presence of 8 yaw drives, a relative small braking force of 40 Nm per break leads to a total braking torque of 690.25 kNm. The yaw bearing is a pre-stressed friction bearing. Figure 1 illustrates the gyroscopic effect acting on the yaw system of the turbine. In case the RNA position is not aligned to the wind direction, the yaw system is activated and the RNA is rotated towards the wind. The level of allowable yaw angle deviation to trigger the start of the yaw system depends on the average wind speed. The higher the average wind speed, the lower the allowed deviation of yaw angle to activate the system. Figure 2 depicts the allowed maximum yaw angle deviations as a function of wind speeds while Table 1 lists these values. Any point between the given discreet values are interpolated by the wind turbine controller and reacted accordingly. Table 1 Allowed maximum yaw angle deviations and delay times. Wind speed level Wind speed above Triggering yaw angle Delay time Level 1 2.5 m/s 15° 10s Level 2 6 m/s 12° 8s Level 3 8 m/s 10° 4s Level 4 12 m/s 8° 2s Level 5 18 m/s 6° 1s Level 6 25 m/s 4° 0s In practice, the wind direction changes very quickly or the sensors can briefly give an incorrect signal. For these reasons, an inertia is built into the whole system. The yaw motors are not switched on unless the allowable yaw deviation lasts for a certain duration. This time lag is called the switch-on delay. Typically, these delay times are reduced for higher wind speeds as shown in Fig. 3 and listed in the last column of Table 1 . In addition, there is a yaw error limit which, if exceeded, triggers an alarm that will stop the turbine. Just like the other parameters the maximum yaw error depends on the wind speed too. For wind speeds different than the discreet values given the maximum yaw errors are determined by interpolation. Once the signal to stop the turbine is triggered no time delay is applied. Figure 4 depicts the maximum yaw error or the maximum yaw deviation angle allowed before triggering the stop mechanism. Table 2 lists these values. Table 2 Maximum yaw errors to stop turbine. Parameter Wind speed Maximum yaw error to stop turbine Max error wind speed 1 0 m/s 40° Max error wind speed 2 10 m/s 35° Max error wind speed 3 20 m/s 30° Max error wind speed 4 30 m/s 20° The present work essentially aims at predicting the failure incidents of the brake pads of the yaw control brakes due to wear out. The brake pads are not directly supervised by a sensor and as a result, for wind speeds in the vicinity of the nominal operation wind speed of the turbine, uncontrolled yaw movements in the range of up to 1° per second are possible since the brakes cannot keep the RNA sufficiently stationary when worn out. This leads to extended yaw corrections hence increased wearing of the yaw system. Additionally, misaligned durations of yaw angles are increased. Yaw misalignment is one of the most critical design load case (DLC) for wind turbines (DLC 1.4 and 3.3 (IEC 61400-1 2019)) and operating under this condition shortens the life time of the wind turbine. Under normal operation conditions the brakes, as part of the yaw system, keep the nacelle in the desired position, as shown in Fig. 5 . Only in case of the required correction of yaw position the RNA is rotated based on the following sequence: open brakes – activate motor (clockwise or counterclockwise) – deactivate motor – close brakes. The turbine studied in this work is equipped with a pre-stressed friction bearing limiting any uncontrolled motion in moments of open brakes hence inactivate durations of yaw motor. Any yaw movement of the RNA with closed yaw brakes is exceptional. As the brake pads of the yaw brakes are wearing out over time, the RNA starts to yaw under load conditions in the range of millidegrees even though the yaw brakes are engaged, see Fig. 6 . Such movements increase with the wind speed, gustiness, and power production level of the wind turbine. The wind turbine controller does not monitor this behavior, only if such a movement causes a yaw misalignment above a pre-set value the controller activates the yaw system to turn the RNA back towards the wind direction. In case the yaw misalignment becomes greater than the maximum yaw error of the operation condition, the controller triggers a safety stop. As the algorithm to determine yaw misalignment is a double layered control with thresholds for safety stop depending on the wind speed level, a misalignment above the threshold may occur due to slipping. Since slipping increases with increasing wind speed, different thresholds for yaw activation or safety stop can be reached even earlier. Per slipping event the RNA moves 6.75 millidegree in 99.3% of the cases. This value is also the smallest movement detectable by the wind turbine controller. The main goal is to define a threshold of acceptable movements per 10 minutes over which the brake pads should be replaced as the brake wear-outs are not supervised. Using the slipping events per 10-minute-periods detected end of November 2021 for a range of average wind speeds between 4.5 and 14.5 m/s, a quintic polynomial fit by regression analysis resulted in the function $$y=0.0349{x}^{5}-1.5785{x}^{4}+28.223{x}^{3}-249.06{x}^{2}+1090.1x-1815.7$$ 1 where y = number of slipping events and x = wind speed cluster used in 0.5 m/s steps. It is noted that the above and following polynomial fits do not satisfy the condition that \(y=0\) when \(x=0\) ; however, this is immaterial as the number of slips at zero wind speed is irrelevant to any subsequent calculation. Figure 7 shows the data points and Eq. ( 1 ) for the range of wind speeds considered. Average and normalized maximum yaw errors in 10-minute-intervals for wind speed clusters between 3.5 and 14 m/s are plotted in Fig. 8 . Since the average and maximum yaw errors tend to zero with increasing wind speed, the yaw error triggering wind speed level is set to 18 m/s, giving a threshold value implying a safety factor of two. Therefore, 3° or 444 slipping event per 10 minutes at 18 m/s is a reasonable threshold for triggering the exchange of brake pads. In order to define a polynomial function satisfying the corresponding values for lower wind speeds, Eq. ( 1 ) is to be scaled down to fit within the defined threshold of 444 events: $$444=0.0349{x}^{5}-1.5785{x}^{4}+28.223{x}^{3}-249.06{x}^{2}+1090.1x-1815.7$$ 2 Solving for x gives x \(\approx\) 15.4563 m/s so that to obtain y= 444 for x = 18 m/s, a shift of 18-15.4563=2.5437 m/s is necessary. Applying this shift to the polynomial results in $$y=0.0349{\left(x-2.5437\right)}^{5}-1.5785{\left(x-2.5437\right)}^{4}+28.223{\left(x-2.5437\right)}^{3}-249.06{\left(x-2.5437\right)}^{2}+1090.1\left(x-2.5437\right)-1815.7$$ 3 which now satisfies y(18) = 444 as aimed. Before incorporating Eq. ( 3 ) into the ML model, the first step involves the preprocessing the available wind turbine data. This data typically includes wind speed, turbine operational parameters, and condition data. The output from Eq. ( 3 ), which predicts the likelihood of slippage events at different wind speeds, is used as a feature in this dataset. This feature enriches the model by providing insights into the critical thresholds at which maintenance actions are necessary. 4. Arrangements and Results 4.1 Evaluation metrics In order to assess the models, we use the measures of precision, recall, accuracy and F-measure, which are computed from the contents of a classic confusion matrix of the classification predictions. True positive and false positive cases are denoted as TP and FP, while true negative and false negative are denoted as TN and FN respectively. In order to fit the classification evaluation in incident detection problem, we assign the classes no-incident and incident. Precision is the ratio of the predicted true positive cases (TP) to the sum of true positives (TP) and false positives (FP) as $$\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}= \frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{P}}$$ Recall is the ratio of the true positive cases to the sum of true positives (TP) and false negatives (FN): $$\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}= \frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{N}}$$ Accuracy is the ratio of the total number of predictions that were correct. $$\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}= \frac{\text{T}\text{P}+\text{T}\text{N}}{\text{T}\text{P}+\text{F}\text{P}+\text{T}\text{N}+\text{F}\text{N}}$$ Precision or recall alone cannot describe a classifier's efficiency; therefore, F-measure is introduced as a combination of these two metrics. It is defined as twice the harmonic mean of precision and recall and is the metric which is most frequently referred. $$\text{F}-\text{m}\text{e}\text{a}\text{s}\text{u}\text{r}\text{e}= \frac{2 \times \text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n} \times \text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}{\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}+ \text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}$$ A value closer to unity indicates better combined precision and recall of the classifier, whereas lower values imply lower accuracy or precision or both. 4.2 SVM Based on the provided results for identifying and predicting using different SVM kernels (Polynomial, RBF, Sigmoid), the need for maintenance events due to worn yaw motor brake pads causing extensive RNA slipping beyond pre-defined limits the following conclusions can be drawn. 1. Precision: The RBF kernel demonstrates the highest precision (75.30%), followed by the Polynomial kernel (68.86%) and the Sigmoid kernel (50.75%). Precision represents the percentage of correctly identified maintenance events among all instances predicted as maintenance events. Higher precision indicates a lower rate of false positives, meaning that the Polynomial kernel performs better in accurately identifying true maintenance events. 2. Recall: The RBF kernel also exhibits the highest recall (74.10%), followed by the Polynomial kernel (67.80%), and the Sigmoid kernel (50.75%). Recall represents the percentage of correctly identified maintenance events out of all actual maintenance events. Higher recall indicates a lower rate of false negatives, implying that the Polynomial kernel is better at capturing a larger portion of actual maintenance events. 3. Accuracy: The accuracy is highest for the RBF kernels (74.10%), followed by the Polynomial kernel (67.80%) and the Sigmoid kernel (50.75%). Accuracy represents the overall percentage of correctly classified instances, regardless of the class. However, it is important to note that accuracy alone might not be sufficient to evaluate the performance of the model, especially in imbalanced datasets. 4. F-measure: The F-measure, which combines precision and recall into a single metric, is highest for the RBF kernel (73.74%), followed by the Polynomial kernel (67.32%) and the Sigmoid kernel (50.75%). The F-measure provides a balanced assessment of both precision and recall, indicating the Polynomial kernel's superior performance in maintaining a good trade-off between identifying true maintenance events and minimizing false positives and false negatives. Based on these results, the RBF kernel, considering its higher precision, appears to be the most effective among the three kernels in identifying maintenance needs related to yaw brake pad degradation, recall, and F-measure. Table 3 Result analyses per kernel Polynomial kernel RBF kernel Sigmoid kernel Precision (%) 68.86 75.30 50.75 Recall (%) 67.80 74.10 50.75 Accuracy (%) 67.80 74.10 50.75 F-measure (%) 67.32 73.74 50.75 The developed code in python is able to predict RNA slipping events with the results shown in Table 3. 4.3 LSTM LSTM models are adept at handling time-series data, making them well-suited for predictive maintenance tasks. The preprocessed dataset, now including the polynomial function's output, is fed into the LSTM model. The model learns from the patterns in the data, including the relationship between wind speed, yaw system behavior, and the occurrence of slipping events, as indicated by Eq. (3). Moreover, LSTMs are capable of learning from sequences of data, capturing temporal dependencies that are crucial in predicting maintenance requirements. By training the LSTM model on the dataset that includes the polynomial function's output (Eq. 3), the model can learn how the risk of slippage events evolves over time and under varying operational conditions. In summary, Eq. (3) enhances the LSTM model's capability to predict maintenance requirements by providing a critical piece of information about the yaw system's behavior. This integration leads to a more robust and reliable predictive maintenance system, helping to reduce downtime and improve the efficiency of offshore wind turbines. 4.3.1 Data preparation In order for the LSTM to provide sufficient and more valuable prediction, the initial dataset is reformed to a minute-period so as the final prediction to make a state. Thus, the input dataset to the LSTM model is given in the Fig. 9 below, where zeros point a non-incident and ones an incident case. 4.3.2 Simulation setup - Model training For the initial data, the re-scale parameter is 1/255 and the validation split is set at 20% (training 80%-validation 20%). The final 30 time-units of the initial dataset were not fed for training and validation to avoid the model making predictions and testing in unknown territories. The tested LSTM network is trained for 50 epochs, with the callback that the best model is saved when it appears. Figure 10 below, provides the training and the validation loss, where these are the metrics used to assess how a deep learning model fits the training data and the performance of a deep learning model on the validation set. The purpose of visualizing together the training and validation loss is to diagnose the model’s performance and identify which aspects are needed for tuning if required. Figure 10 shows that training and validation loss both decrease and start stabilizing at a point above 40 epochs. This indicates that the proposed LSTM model has an optimal fit. Table 4 below provides the evaluation metrics for training and validation set, where one can see that the LSTM model performs close to optimal because F-measure is above 96% for both sets. Table 4 Evaluation metrics for training and validation set. Set Precision (%) Recall (%) Accuracy (%) F-measure (%) Training 98.91 96.27 93.42 97.56 Validation 95.99 96.26 93.21 96.12 4.3.3 Experimental results of the integrated methodology Table 5 provides the evaluation metrics of the forecasting of LSTM in test set. These experimental results clearly show that the forecasting performance of the proposed architecture is high, where the prediction accuracy is over 98% and the overall f-measure is above 96%. Table 5 Evaluation metrics for test set. Testing set Precision (%) Recall (%) Accuracy (%) F-measure (%) 96.11 96.37 98.36 96.22 Figure 11 (a) illustrates a representation of the projected incident occurrences, while (b) depicts a combination of both the initial incident data and the forecasted incidents. 5. Conclusions and future work By integrating domain expertise with machine learning techniques, this article reveals the potential to significantly enhance predictive maintenance practices in the offshore wind industry, ultimately leading to increased turbine reliability, reduced operational costs, and a more sustainable energy future. With carefully selected variables from the control system and the development of features by combining relevant variables, unsupervised components or events that can lead to shut down of the wind turbine can be converted into indirectly supervised parts. Consequently, the RTTF can be predicted with a fair accuracy thus increasing the reliability of the wind turbine and its technical availability. An important aspect shown is that by using LSTM the prediction time can be increased to a period mandatory in offshore wind industry due to a vast number of constrains like weather conditions, spare part management, and ship and crew availability. The prediction abilities of SVM for this task are significant lower making LSTM the better choice to reach reliable forecast results. Applying this method to other unsupervised components or sub-systems requires careful development of features based on available variables recorded in the data set from the wind turbine control system. Declarations Author Contributions: Conceptualization: Uwe Lützen; Methodology: Serdar Beji; Formal analysis and investigation: Uwe Lützen; Software, writing—original draft: Uwe Lützen; Writing—review and editing: Serdar Beji; Supervision: Serdar Beji. All authors have read and agreed to the published version of the manuscript. Funding: No funding was received for conducting this study. Data Availability Statement: The data presented in this study are available on request from the corresponding author in aggregated form. The data are not publicly available due to restrictions from the owner of the wind turbine. Acknowledgments: This work was carried out as a part of doctoral studies of the first author at Istanbul Technical University. Conflicts of Interest: The authors declare no conflict of interest. References Cevasco D, Koukoura S, Kolios A J (2021) Reliability, availability, maintainability data review for the dientfification of trends in offshore wind energy applications. Renewable and Sustainable Energy Reviews, vol. 136. https://doi.org/10.1016/j.rser.2020.110414 Chen J-h, Pei A-g, Chen P, Hu Z-q (2021) Study on Gyroscopic Effect of Floating Offshore Wind Turbines. Chinese Ocean Engineering 35: 201-214. https://doi.org/10.1007/s13344-021-0018-z Cortes C, Vapnik V (1995) Support-vector networks. Machine Learning 20: 273-297. https://doi.org/10.1007/BF00994018 Cristianini N, Shawe-Taylor J (2000) An Introduction to Support Vector Machines and Other Kernel-based Learning Methods. Cambridge University Press. https://doi.org/10.1017/CBO9780511801389 Danish Wind Industry Association (2003) Wind Turbines: Upwind or Downwind Machines?," Danish Wind Industry Association. http://www.windpower.org/en/tour/design/updown.htm. Accessed 28 October 2022. Dinwoodie I, Endrerud O-E, Hofmann M, Martin R, Bakken Sperstad I (2015) Reference Cases for Verification of Operation and Maintenance Simulation Models for Offshore Wind Farms. Wind Engineering 39: 1-14. https://doi.org/10.1260/0309-524X.39.1.1 Dinwoodie I, McMIllan D, Revie M, Lazakis I, Dalgic Y (2013) Development of a Combined Operational and Strategic Decision Support Model for Offshore Wind. Energy Procedia 35: 157-166. https://doi.org/10.1016/j.egypro.2013.07.169 Douard F, Domecq C, Lair W (2012) A Probabilistic Approach to Introduce Risk Measurement Indicators to an Offshore Wind Project Evaluation – Improvement to an Existing Tool. Energy Procedia 24: 255-262. https://doi.org/10.1016/j.egypro.2012.06.107  Endrerud O-E, Liyanage J P, Keseric N (2014) Marine Logistics Decision Support for Operation and Maintenance of Offshore Wind Parks with a Multi Method Simulation Model. Winter Simulation Conference, Savannah, Georgia. https://doi.org/10.1109/WSC.2014.7020021 Farrar N O, Ali M H, Dasgupta D (2023) Artificial Intelligence and Machine Learning in Grid Connected Wind Turbine Control Systems: A Comprehensive Review. Energies 16. https://doi.org/10.3390/en16031530 Feuchtwang J, Infield D (2013) Offshore Wind Turbine Maintenance Access: A Closed-Form Probabilistic Method for Calculating Delays Caused by Sea-State. Wind Energy 16: 1049-1066. http://dx.doi.org/10.1002/we.1539 Guyon I, Elisseeff A (2003) An introduction to variable and feature selection. Journal of Machine Learning Research 3: 1157-1182. https://doi.org/10.1162/153244303322753616 GWEC (2020) Global Wind Report 2019. Brussels. Halvorsen E E, Gundegjerde C, Halvorsen I B, Hvattum L M, Nonås L M (2013) Vessel fleet analysis for maintenance operations at offshore wind farms. Energy Procedia 35: 167-176. https://doi.org/10.1016/j.egypro.2013.07.170 Havrlant L, Kreinovich V (2017) A Simple Probabilistic Explanation of Term Frequency-Inverse Document Frequency (tf-idf) Heuristic (and Variations Motivated by This Explanation). International Journal of General Systems 46: 27-36. https://doi.org/10.1080/03081079.2017.1291635 He H, Garcia E A (2009) Learning from imbalanced data. IEEE Transactions on Knowledge and Data Engineering 21: 1263-1284. https://doi.org/10.1109/TKDE.2008.239 Hofmann M (2011) A Review of Decision Support Models for Offshore Wind Farms with an Emphasis on Operation and Maintenance Strategies. Wind Engineering 35: 1-16. https://doi.org/10.1260/0309-524X.35.1.1 IEC (2019) IEC 61400-1: Wind Turbinesepart 1: Design Requirements. International Electrotechnical Commission, Geneva. International Energy Agency (2019) Market Report Series: Renewables 2019. International Energy Agency. Ioannou A, Angus A, Brennan F (2019) Informing parametric risk control policies for operational uncertainties of offshore wind energy assets. Ocean Engineering 177: 1-11. https://doi.org/10.1016/j.oceaneng.2019.02.058 Kang J, Wang Z, Guedes Soares C (2020) Condition-Based Maintenance for Offshore Wind Turbines Based on Support Vectro Machine. Energies 13, no. 3518. https://doi.org/10.3390/en13143518 Karyotakis A (2011) On the Optimisation of Operation and Maintenance Strategies for Offshore Wind Farms. Dissertation, University College London. Kessler E H (2013) Encyclopedia of Management Theory. SAGE Publications ,Thousand Oaks, California. Khalid S, Khalil T, Nasreen S (2014) A Survey of Feature Selection and Feature Extraction Techniques in Machine Learning. Science and Information Conference, London. https://doi.org/10.1109/SAI.2014.6918213 Kumar A (2022) Data Preprocessing Steps in Machine Learning. Data Analytics. https://vitalflux.com/data-preprocessing-steps-in-machine-learning/. Accessed 16 04 2023. Lazakis I, Kahn S (2021) An optimization framework for daily route planning and scheduling of maintenance vessel activities in offshore wind farms. Ocean Engineering 225. https://doi.org/10.1016/j.oceaneng.2021.108752 Lindemann B, Müller T, Vietz H, Jazdi N, Weyrich M (2021) A survey on long short-term memory networks for time series prediction. Science Direct 99: 650-655. https://doi.org/10.1016/j.procir.2021.03.088 Marti-Puig P, Blanco-M A, Cardenas J J, Cusido J, Sole-Casals J. (2019) Feature Selection Algorithms for Wind Turbine Failure Prediction. Energies 12: 453. https://doi.org/10.3390/en12030453 Masoumi M (2023) Machine Learning Solutions for Offshore Wind Farms: A Review of Applications and Impact. Journal of Marine Science and Engineering 11. https://doi.org/10.3390/jmse11101855 Nivet T, Muk-Pavik E. (2015) Optimization of O&M for Offshore Wind Farms Modelling. WMTC Conferences. University College London. https://doi.org/10.5957/WMTC-2015-214 Pandit R K, Astolfi D, Cardenas I D (2023) A Review of Predictive Techniques Used to Support Decision Making for Maintenance Operations of Wind Turbines. Energies 16. https://doi.org/10.3390/en16041654 Papatzimos A K (2019) Data-driven Operations & Maintenance for Offshore Wind Farms: Tools and Methodologies. Dissertation, University of Edinburgh. http://dx.doi.org/10.7488/era/17 Perez Granados D, Ortega Ruiz M A, Moreira Acosta J, Gama Lara S A, Gonzalez Dominguez R A, Paramo Kanetas P J (2023) A Wind Turbine Vibration Monitoring System for Predictive Maintenance Based on Machine Learning Methods Developed under Safely Controlled Laboratory Conditions. Energies 16. https://doi.org/10.3390/en16052290 Platt J (1999) Fast training of support vector machines using sequential minimal optimization. Advances in Kernel Methods: Support Vector Learning, pp. 185-208. Rademakers L W M M, Braam H, M. B. Zaaijer M B, van Bussel G J W (2003) Assessment and Optimisation of Operation and Maintenance of Offshore Wind Turbines. European Wind Energy Conference, Madrid. Rusu E, Onea F (2023) The Expected Dynamics of the European Offshore Wind Sector in the Climate Context. Journal of Marine Science and Engineering 11. https://doi.org/10.3390/jmse11101967 Scheu M, Matha D, Hofmann M, Muskulus M (2012) Maintenance strategies for large offshore wind farms. Energy Procedia 24: 281-288. https://doi.org/10.1016/j.egypro.2012.06.110 Schölkopf B, Smola A J (2002) Learning with kernels: Support vector machines, regularization, optimization, and beyond. MIT Press. https://doi.org/10.7551/mitpress/4175.001.0001 Schölkopf B, Smola A, Williamson R, Bartlett P (2000) New support vector algorithms. Neural Computation 12: 1207-1245. https://doi.org/10.1162/089976600300015565  Setiawan B, Djanali S, Ahmad T (2019) Increasing Accuracy and Completeness of Intrusion Detection Model Using Fusion of Normalization, Feature Selection Method and Support Vector Machine. International Journal of Intelligent Engineering & Systems 12: 378 - 389. https://doi.org/ 10.22266/ijies2019.0831.35 Seyr H, Muskulus M (2019) Decision Support Models for Operations and Maintenance for Offshore Wind Farms: A Review. Applied Science 9. https://doi.org/10.3390/app9020278 Sperstad I B, Stålhane M, Dinwoodie I, Endrerud O-E V, Martin R, Warner E (2017) Testing the robustness of optimal access vessel fleet selection for operation and maintenance of offshore wind farms. Ocean Engineering 145: 334-343. https://doi.org/10.1016/j.oceaneng.2017.09.009 Stahlhane M, Hvattum L M, Skaar V (2015) Optimization of routing and scheduling of vessels to perform maintenance at offshore wind farms. Energy Procedia 80: 92-99. https://doi.org/10.1016/j.egypro.2015.11.411 Tang M, Zhao Q, Ding S X, Wu H, Li L, Long W, Huang B (2020) An Improved LightGBM Algorithm for Online Fault Detection of Wind Turbine Gearboxes. Energies 13. https://doi.org/10.3390/en13040807 Yu X (2016) Modelling Offshore Wind Farm Operation and Maintenance with View to Estimating the Benefits of Condition Monitoring. Dissertation, University of Strathclyde. Footnotes TF-IDF (Term Frequency - Inverse Document Frequency) is a handy algorithm that uses the frequency of words to determine how relevant those words are to a given document . It's a relatively simple but intuitive approach to weighting words, allowing it to act as an efficient starting point for a variety of tasks. (Hacrlant and Kreinovich 2017) Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 15 Jun, 2024 Read the published version in Journal of Ocean Engineering and Marine Energy → Version 1 posted Editorial decision: Revision requested 06 Apr, 2024 Reviews received at journal 02 Apr, 2024 Reviewers agreed at journal 18 Mar, 2024 Reviews received at journal 02 Mar, 2024 Reviewers agreed at journal 22 Feb, 2024 Reviewers agreed at journal 21 Feb, 2024 Reviewers invited by journal 03 Feb, 2024 Editor assigned by journal 30 Jan, 2024 Submission checks completed at journal 30 Jan, 2024 First submitted to journal 28 Jan, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-3906932","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":270060833,"identity":"311c3d37-79c1-4f13-8a6e-2a0d74dccc7a","order_by":0,"name":"Uwe Lützen","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyElEQVRIiWNgGAWjYBACPgglUd/PDGYwE9bCBlFkwTizmRmqhY04LRWMGw4QrUUi//CHjzskmI2P8x+TYKiwTmyQ731AQEsyg+HMMxJsZoeZ2SQYzqQnNrCxGxDUkszbJsED1sLYdhiohYDLQFoO/22TkDBuBmn5R5wWxmbGNgkDA2aQlgZitPA8NmbsbZNIkDjMbGyRcCzduI0tDb8WfvbExx9+ttUl8PcffHjjQ421bD/zMfxaUEECA+FoGQWjYBSMglFABAAAmfw1AFip81UAAAAASUVORK5CYII=","orcid":"","institution":"Istanbul Technical University","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"Uwe","middleName":"","lastName":"Lützen","suffix":""},{"id":270060834,"identity":"ae91c10b-8894-48ed-9b5b-0f4d3f2c0b4e","order_by":1,"name":"Serdar Beji","email":"","orcid":"","institution":"Istanbul Technical University","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Serdar","middleName":"","lastName":"Beji","suffix":""}],"badges":[],"createdAt":"2024-01-28 21:29:10","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3906932/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3906932/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s40722-024-00335-z","type":"published","date":"2024-06-16T00:33:23+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":50501875,"identity":"7efdf87b-d3a6-4441-8ca6-f250453cef59","added_by":"auto","created_at":"2024-02-01 13:37:03","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":22724,"visible":true,"origin":"","legend":"\u003cp\u003eAn illustration of the gyroscopic effect on the yaw system (Chen 2021)\u003cstrong\u003e.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3906932/v1/942d2059b2d58fa99b012f5b.jpg"},{"id":50501625,"identity":"6b21934f-f04f-479c-b592-51e4324cb906","added_by":"auto","created_at":"2024-02-01 13:29:03","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":16699,"visible":true,"origin":"","legend":"\u003cp\u003eAllowed yaw angle deviation versus wind speed.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3906932/v1/cec09edad82133fef7705b08.jpg"},{"id":50501623,"identity":"90c8f7d1-ad76-47de-a8bc-4b210ebb889f","added_by":"auto","created_at":"2024-02-01 13:29:03","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":17200,"visible":true,"origin":"","legend":"\u003cp\u003eInertia delay duration versus wind speed.\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3906932/v1/d0e13df2acf57822696f22d9.jpg"},{"id":50501624,"identity":"3327860c-375a-45f1-8f1d-d58537fd42c7","added_by":"auto","created_at":"2024-02-01 13:29:03","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":16010,"visible":true,"origin":"","legend":"\u003cp\u003eMaximum yaw error to stop turbine versus wind speed.\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3906932/v1/65116d3d471433c6ea319a67.jpg"},{"id":50502100,"identity":"6e0e5d28-0f09-4de9-8baf-fdad2d1fe0c1","added_by":"auto","created_at":"2024-02-01 13:45:03","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":48571,"visible":true,"origin":"","legend":"\u003cp\u003eSlipping events during one hour of turbine operation on 14.06.2021.\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3906932/v1/ee8354e481eb24b895962982.jpg"},{"id":50501630,"identity":"e9107ec4-4e08-4fbc-a14e-f40405ff3729","added_by":"auto","created_at":"2024-02-01 13:29:03","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":85133,"visible":true,"origin":"","legend":"\u003cp\u003eSlipping events during one hour of turbine operation on 24.11.2021.\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3906932/v1/88f54a5a4b1a28c318bba8f3.jpg"},{"id":50501626,"identity":"66c81f54-8f27-43c4-abed-3dae513f21ee","added_by":"auto","created_at":"2024-02-01 13:29:03","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":35979,"visible":true,"origin":"","legend":"\u003cp\u003eNumber of slipping events per wind speed cluster: data and polynomial fit.\u003c/p\u003e","description":"","filename":"7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3906932/v1/b6827d0982f1c7e931b0661a.jpg"},{"id":50501878,"identity":"99b46b57-b113-490e-a1b4-fb1ae23ef980","added_by":"auto","created_at":"2024-02-01 13:37:03","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":47641,"visible":true,"origin":"","legend":"\u003cp\u003eYaw errors in 10-minute-intervals.\u003c/p\u003e","description":"","filename":"8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3906932/v1/1a292aee02461f7a04313b1d.jpg"},{"id":50501876,"identity":"3dbb5d96-242d-4201-86c5-bfd6222cfdb6","added_by":"auto","created_at":"2024-02-01 13:37:03","extension":"jpg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":40011,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Initial dataset of incidents (b) zoom-in into the first 500 time moments.\u003c/p\u003e","description":"","filename":"9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3906932/v1/dd94a30361c5b2cd7d507f95.jpg"},{"id":50501879,"identity":"ea9c19cb-b752-49d3-9b33-2a6230647a5b","added_by":"auto","created_at":"2024-02-01 13:37:03","extension":"jpg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":26257,"visible":true,"origin":"","legend":"\u003cp\u003eTraining and validation loss for LSTM architecture.\u003c/p\u003e","description":"","filename":"10.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3906932/v1/6081213a84d3b1c121a4352f.jpg"},{"id":50501632,"identity":"cbac546f-958f-4b9d-842a-6fb74c4f111f","added_by":"auto","created_at":"2024-02-01 13:29:03","extension":"jpg","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":39891,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Forecasted incidents (b) forecasted (red dots) and initial (blue dots) incidents combined.\u003c/p\u003e","description":"","filename":"11.jpg","url":"https://assets-eu.researchsquare.com/files/rs-3906932/v1/85799ad8a2749c876ae3a2e8.jpg"},{"id":58456499,"identity":"13d326d7-c23e-46ad-9e4a-74b9d274a4fd","added_by":"auto","created_at":"2024-06-17 00:33:28","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1087919,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3906932/v1/a4c9e59c-3d17-4e50-a3c8-ff4c102360d1.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003ePredictive Maintenance for Offshore Wind Turbines through Deep Learning and Online Clustering of Unsupervised Subsystems: A Real-World Implementation\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eIn view of escalating global pursuit of offshore wind energy resources as an important element in the transition from fossil fuels to renewable energy sources, ensuring the reliability of offshore wind technology becomes increasingly obvious. Consequently, entities responsible for the operation and maintenance (O\u0026amp;M) of offshore wind farms are progressively raising their standards by employing artificial intelligence (AI) solutions for improving performance and productivity of these vital installations. Accordingly, increased interest in AI integration has prompted a surge in corporate transactions, notably marked by acquisitions of machine learning and AI sectors as O\u0026amp;M companies strategically position themselves to harness the potential of AI to enhance output, sustainability and feasibility of offshore wind farms.\u003c/p\u003e \u003cp\u003eIn that context, predictive maintenance is quite an important tool for increasing operational availability of offshore wind turbines. Availability of a typical offshore wind turbine can be significantly lower than its onshore counterpart which has a typical availability rate of 95\u0026ndash;97% (Cevasco et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Lower availability rates necessitate new approaches to optimize O\u0026amp;M strategies for offshore wind farms to significantly increase the feasibility and grid stability.\u003c/p\u003e \u003cp\u003eThe need for new approaches to O\u0026amp;M strategies for offshore wind energy was recognized quite early about two decades ago, as Rademakers et al. (\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2003\u003c/span\u003e) drew attention to this point during a wind energy conference. In line with this concern Scheu et al. (\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2012\u003c/span\u003e) developed a simulation for the operational phase of offshore wind farms with a focus on modeling of failures and repair. However, until recently, optimization studies concerning the offshore wind O\u0026amp;M problems were mainly focused on specific components as the root cause of failures, as in Karyotakis (\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2011\u003c/span\u003e) who focused on wind turbine inverters as one of the main causes for the interruption of operation.\u003c/p\u003e \u003cp\u003eSeveral factors, such as weather conditions, different component degradation and wear, influence the selection process for the optimal maintenance strategy. Therefore, decision support strategies (DSS) are important aids towards an informed decision. Kessler (\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2013\u003c/span\u003e) presented several DSS in detail; namely, data-driven, communication-driven, document-driven, knowledge-driven, and model-driven DSS. More specific works concerning decision support models and their development for offshore wind energy O\u0026amp;M strategies are also available. A notable work is due to Hofmann (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2011\u003c/span\u003e) who review 49 different commercial as well as non-commercial decision support models focusing mainly on O\u0026amp;M but also including logistics, power production and overall project costs. Papatzimos (2020) gives a decision support framework for maintenance planning of a single wind turbine based on data management that builds a very good basis for extension to wind farms. Endrerud et al. (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2014\u003c/span\u003e) develop a marine logistic simulation model for the O\u0026amp;M lifecycle phase. Dinwoodie et al. (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2013\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) present the development of a strategic decision support model for operation of wind farms and simulation models for O\u0026amp;M. A comprehensive review on the state of art in offshore wind farm O\u0026amp;M with a focus on decision support models for the scheduling of maintenance is given by Seyr and Muskulus (\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAssessment of the input factors and their impact on offshore wind O\u0026amp;M are an important subjects of interest. In this rather broad area, Feuchtwang and Infield (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2013\u003c/span\u003e) consider the delays for repair and maintenance caused by the sea, Douard et al. (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2012\u003c/span\u003e) and Ioannou et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) treat the risk measurement indicators while Yu (\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2016\u003c/span\u003e) shows the benefits of condition monitoring. Pandit et al. (\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) present a review of predictive techniques to support decision making for O\u0026amp;M of wind turbines focusing on condition monitoring for main components. Optimization of a fleet of vessels is assessed by Halvorsen et al. (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). Stalhane et al. (2015) and Lazakis and Kahn (\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) look into optimal routes and schedules. Sperstad et al. (\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2017\u003c/span\u003e) consider the vessel fleet optimization in relation to O\u0026amp;M cost, electricity price and vessel transit speed.\u003c/p\u003e \u003cp\u003eMachine learning (ML) plays a pivotal role for handling large amount of data collected by the wind turbine controller and/or additional systems like condition monitoring system for supporting predictive maintenance. Perez Granados et al. (\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) present a methodology for predictive maintenance of wind turbines by using vibration monitoring as input. Tang et al. (\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) employ an improved lightGBM algorithm for online fault detection in gearboxes of wind turbines. Farrar et al. (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) give an overview of the use of artificial intelligence and ML in grid connected wind turbine control systems. An important aspect of the application of ML is the feature selection, which is studied by Marti-Puig et al. (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) for finding the optimal predictors via different feature selection algorithms. Quite recently, Masoumi (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) has presented a comprehensive review of machine learning applications for offshore wind farms.\u003c/p\u003e \u003cp\u003eInfluence of different O\u0026amp;M strategies on the sustainability of a wind farm is treated in Nivet and Muk-Pavik (\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). Overall situation of offshore wind energy is laid out by the International Energy Agency (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) as well as the GWEC (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) by describing an outlook for the coming years with an emphasis on that offshore wind energy will become more and more important part of the energy mix. Rusu and Onea (\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) gives a review of the state of offshore wind energy sector in Europe in connection with the climate change.\u003c/p\u003e \u003cp\u003eThe above review of the relevant literature reveals that the significance of predictive maintenance for wind turbine systems, particularly in the context of offshore wind energy generation, has grown considerably. The present work endeavors to make a contribution to the field by introducing and simulating an innovative approach for the predictive maintenance of wind turbines. The central focus of this work lies in the design, development, and practical application of the prediction of failures in unsupervised sub-systems in a real-life wind turbine system, specifically parts of the yaw system, utilizing data-driven approaches to validate the proposed methodology. In particular, the present methodology leverages deep learning models in conjunction with online clustering techniques, thus establishing a robust foundation for predictive maintenance in offshore wind energy systems. Accordingly, a comprehensive framework, methodology, experimental setup, and results are presented here for giving new contributions and suggesting insights into the evolving landscape of predictive maintenance of offshore wind turbines.\u003c/p\u003e \u003cp\u003eReal data of over six months have been taken from a 3MW direct-drive wind turbine for this work. The yaw system of this wind turbine experiences an increasing number of unintended slipping events where the yaw system is not able to station keep the wind turbine due to unsupervised wear out of a component.\u003c/p\u003e \u003cp\u003eSection \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e is dedicated to elucidating the categorization of the data collected from turbines. Also, a rigorous evaluation of the selected methodology, which places significant reliance on the utilization of deep learning models, is provided. This analysis serves to underscore the rationale behind the methodological choices.\u003c/p\u003e \u003cp\u003eSection \u003cspan refid=\"Sec14\" class=\"InternalRef\"\u003e3\u003c/span\u003e describes the formulation of the problem and elucidation of underlying assumptions pivotal to predictive maintenance strategies, specifically the inclusion of unsupervised sub-systems and components into predictive maintenance strategies. A detailed description of these considerations serves as the bedrock for the subsequent development of prediction strategies.\u003c/p\u003e \u003cp\u003eSection \u003cspan refid=\"Sec15\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents the results obtained through the application of the selected predictive strategy for unsupervised sub-systems. These results are scrutinized in detail to offer a comprehensive assessment of their implications and significance within the broader context of predictive maintenance for offshore wind turbines.\u003c/p\u003e \u003cp\u003eSection \u003cspan refid=\"Sec22\" class=\"InternalRef\"\u003e5\u003c/span\u003e, the final section, encapsulates the key findings stemming from the application of the proposed predictive maintenance methodology. Additionally, insightful recommendations for prospective areas of research and development are provided thereby delineating a roadmap for future endeavors in this particular research field.\u003c/p\u003e"},{"header":"2. Applied methodology","content":"\u003cp\u003eAs already indicated, offshore wind turbine maintenance operations reveal a pressing need for more accurate maintenance strategies to optimize operational efficiency and reduce downtime. These strategies are highly dependent on accurate data and precise forecast of expected time to failure (TTF) since a fair forecasting period; namely, a sufficient duration is needed in order to plan maintenance operations early enough to prevent downtimes. As offshore wind farms are only accessible at certain weather conditions, the forecasting period and warnings of upcoming expected problems must be long enough to take adverse weather conditions into account.\u003c/p\u003e\n\u003cp\u003eAchieving a fair or long-enough forecasting period requires advanced engineering approaches. With this viewpoint in mind this section describes the integration of data clustering techniques to Support Vector Machine (SVM) and Long-Short-Term Memory (LSTM) networks for forecasting. This approach offers a novel perspective in the context of offshore wind turbines by emphasizing the critical role of feature engineering in harnessing the full potential of deep learning models. By effectively grouping similar data points through clustering, it is possible to capture the underlying patterns and dependencies within the intricate operational data of offshore wind turbines. This not only improves the model\u0026apos;s ability to learn from the data but also enhances its interpretability, enabling better informed maintenance decisions.\u003c/p\u003e\n\u003cp\u003eWhat sets the proposed contribution apart is the emphasis on leveraging experimental data for training and testing the predictive maintenance model. In the offshore wind energy sector, where real-world data is often characterized by high-dimensional and noisy observations, the utilization of experimental data becomes paramount. Experimental data provides a controlled environment to validate the effectiveness of the clustering framework, ensuring its applicability and reliability in the challenging offshore conditions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.1 Data Preparation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData processing plays a crucial role in machine learning (ML) for the maintenance of wind turbines. Basic aspects of data processing can be listed as follows (Kumar 2022).\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eData Collection: Wind turbines gather continuous data on parameters like the wind speed, temperature, vibration, and power output through sensors and monitoring systems. This data is collected and stored for analysis.\u003c/li\u003e\n \u003cli\u003eData Cleaning: Raw data often contains errors, missing values, or outliers. Data cleaning involves identifying and addressing these issues to ensure the data accuracy and reliability. Techniques such as imputation, outlier detection, and error correction are applied.\u003c/li\u003e\n \u003cli\u003eFeature Extraction: Relevant features or variables are extracted from the raw data to capture essential information for predictive modeling. These features could include wind speed patterns, temperature fluctuations, historical maintenance records, or other parameters crucial for identifying potential issues or predicting failures.\u003c/li\u003e\n \u003cli\u003eFeature Engineering: Engineers create new features from the existing data to improve the predictive power of machine learning models. This step requires expertise and domain knowledge, resulting in features like wind turbulence intensity, power output degradation rates, or health indicators based on vibration patterns.\u003c/li\u003e\n \u003cli\u003eFeature Scaling and Normalization: Features are scaled to ensure they are on a similar scale. Scaling methods like normalization or standardization are used to prevent certain features from dominating the learning process and aid in model convergence.\u003c/li\u003e\n \u003cli\u003eData Integration: Data from multiple turbines may be integrated to create a comprehensive dataset. This allows for a broader analysis of patterns and correlations across turbines, leading to more accurate predictions and maintenance strategies.\u003c/li\u003e\n \u003cli\u003eData Splitting: The processed dataset is divided into training and test sets. The training set is used to train machine learning models, while the test set is used to evaluate their performance. This ensures that the models generalize well to unseen data and can effectively predict maintenance needs.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe above steps are essential for preparing data for predictive maintenance in wind turbines, ensuring that machine learning models can provide accurate and reliable maintenance recommendations. By leveraging these techniques, machine learning models can be trained on high-quality, informative datasets, enabling the prediction of potential failures, identification of maintenance requirements, and optimization of maintenance schedules for wind turbines.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.1.1 Data preprocessing\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;and balancing of the dataset\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData preprocessing is a crucial step in data preparation, involving various techniques to ready raw data for further processing. It is used not only in data mining but also in training machine learning and AI models to ensure accuracy. Data balancing, on the other hand, addresses class imbalance issues within a dataset. Class imbalance occurs when one class significantly outnumbers another, which can lead to challenges in model training, especially for the minority class. Data balancing methods include over-sampling (increasing minority class instances), under-sampling (decreasing majority class instances), or hybrid approaches that combine both. The choice of technique depends on the dataset and problem requirements, but caution is needed to prevent overfitting or bias. Cross-validation and independent testing help assess the impact of data balancing on model performance.\u003c/p\u003e\n\u003cp\u003eThe choice of data balancing technique depends on the specifics of the dataset, the degree of class imbalance, and the requirements of the problem at hand. It\u0026apos;s important to note that data balancing should be performed carefully to avoid overfitting, loss of important information, or introducing bias into the model. Cross-validation and evaluation on independent test sets are essential to assess the impact of data balancing on model performance (He and Garcia 2009).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.1.2 Present application: Dataset for Wind Turbine Maintenance\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset is divided into two classes: those with detected slipping, and those without detected slipping of the yaw system. The total data exhibits a significant class imbalance with class 0 (non-slipping) comprising 968,567 instances and class 1 (slipping) consisting of only 17,832 instances. This severe imbalance poses a challenge for the model\u0026apos;s ability to accurately detect the class 1, which represents the deteriorating fault condition leading to maintenance requirements once a predefined threshold is reached. To address this issue and ensure proper identification of the smaller data set class 1, an under-sampling technique is employed. By selectively reducing the number of instances from the bigger data set class 0, it is aimed to rebalance the dataset and improve the model\u0026apos;s performance in detecting the crucial fault condition that will trigger the maintenance actions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.1.3 Feature engineering\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFeature engineering is a crucial phase in the data processing workflow that focuses on transforming raw data into meaningful features, enabling more efficient training of data models and facilitating accurate inferences. It encompasses a range of techniques aimed at extracting, selecting, and transforming data attributes to enhance the performance and interpretability of machine learning algorithms. During feature engineering, domain knowledge and statistical analysis are employed to derive new features that capture relevant information and patterns in the data. This may involve techniques such as (Khalid et al. 2014)\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eFeature extraction: Converting raw data into more informative representations. For example, extracting statistical measures (e.g., mean, variance) from time-series data or deriving textual features (e.g., word frequency, TF-IDF\u003csup\u003e[1]\u003c/sup\u003e) from unstructured text.\u003c/li\u003e\n \u003cli\u003eFeature selection: Identifying the most relevant features that have the greatest impact on the target variable. This can be done through methods like correlation analysis, statistical tests, or regularization techniques (e.g., L1 regularization).\u003c/li\u003e\n \u003cli\u003eFeature transformation: Modifying the scale, distribution, or structure of features to meet modeling assumptions or improve performance. Common techniques include scaling features to a specific range (e.g., normalization, standardization), applying mathematical functions (e.g., logarithmic, polynomial transformations), or creating interaction terms.\u003c/li\u003e\n \u003cli\u003eFeature encoding: Converting categorical variables into numerical representations suitable for modeling. This can involve techniques like one-hot encoding, label encoding, or ordinal encoding.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eEffective feature engineering enables models to capture relevant patterns and relationships in the data, improving their predictive power and interpretability. It also helps mitigate issues such as overfitting, and the curse of dimensionality. It is essential to approach feature engineering iteratively, evaluating the impact of engineered features on model performance using appropriate validation techniques. By iteratively refining the features, the data can be organized in ways that maximize the efficiency of data models and facilitate robust inferences (Guyon and Elisseeff 2003).\u003c/p\u003e\n\u003cp\u003eFor the present application the feature selection led to the following descriptions (variable names):\u003c/p\u003e\n\u003cp\u003eWTURTurSt: \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; Wind turbine status, with following values:\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; 1 Idling, waiting for wind, ready for startup, no fault\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; 4 Start up sequence started, self testing\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; 6 start up sequence, connecting the inverter to the grid\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; 8 Turbine in operation (no power limitation)\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; 9 Turbine in operation (power limitation, value preset\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; by operator)\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; 21 Manual stop\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; 23 Fault\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; 24 Emergency stop\u003c/p\u003e\n\u003cp\u003eWNACWdSpdFilVal: \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; Wind speed measured on top of the wind turbine\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; nacelle, filtered value averaged from two anemometers (m/s)\u003c/p\u003e\n\u003cp\u003eWTURWActVal: \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; Wind turbine actual generator power (kW)\u003c/p\u003e\n\u003cp\u003eWNACDrillPosActVal: \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; Absolute yaw position of the RNA (Rotor Nacelle\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; Assembly) in \u0026deg;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWROTRotSpdActVal: \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;The speed of the rotor (RPM)\u003c/p\u003e\n\u003cp\u003eWYAWMot1On: \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Yaw motors activated (TRUE) or not activated (FALSE)\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.1.4 Feature scaling or normalization\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFrequently, it is observed that various variables undergo alterations across disparate scales, wherein one variable exhibits a linear progression while another variable demonstrates an exponential trend. To illustrate, remuneration may be quantified in terms of thousands of dollars, whereas age is typically denoted by two-digit figures. Employing scaling techniques facilitates the transformation of such data in a manner that simplifies the discernment of significant associations between variables by algorithms.\u003c/p\u003e\n\u003cp\u003eThe use of normalization is very influential in the application of SVM. Normalization can shorten the learning process and improve the performance; however, it needs to be applied correctly and therefore should be applied mutual to all features selected (Setiawan et al. 2019).\u003c/p\u003e\n\u003cp\u003eThe following scaling and normalizations have been applied:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eNormalization of variables to the interval [0,1]\u003c/li\u003e\n \u003cli\u003eTransform \u0026lsquo;TRUE\u0026rsquo; and \u0026lsquo;FALSE\u0026rsquo; to 1 and 0 of \u0026apos;WYAWMot1On\u0026apos; feature so as to be compatible with the data model\u003c/li\u003e\n \u003cli\u003eCreate events based on given conditions:\u003cul class=\"decimal_type\"\u003e\n \u003cli\u003eWTURTurSt = 8 (or 9)\u003c/li\u003e\n \u003cli\u003eWNACWdSpdFilVal \u0026gt; 7m/s\u003c/li\u003e\n \u003cli\u003eWTURWActVal \u0026gt; 0 kW\u003c/li\u003e\n \u003cli\u003eWYAWMot1On = False\u003c/li\u003e\n \u003cli\u003eWNACDrillPosActVal change of value to the previous data set of typically 0.00675\u0026deg;\u003c/li\u003e\n \u003c/ul\u003e\n \u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eIn case all conditions are fulfilled, the dataset belongs to class 1, otherwise to class 0.\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eApply under-sampling to address class imbalance in the dataset.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003e\u003cstrong\u003e2.2 Methodology for maintenance\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.2.1 Support Vector Machines (SVMs)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSupport vector machine (SVM) classifiers is an effective data analysis method, which is based on the structural risk minimization principle solving a quadratic programming problem (Cortes and Vapnik 1995). By applying the kernel function, the method can map the data into a higher dimensional input space and construct an optimal hyperplane (Kang et al. 2020).\u003c/p\u003e\n\u003cp\u003eSupport Vector Machines (SVMs) are primarily used for classification tasks, but they can also be applied to maintenance systems in various ways (Setiawan et al. 2019). Here are a few examples of how SVMs can be used in maintenance systems:\u003c/p\u003e\n\u003cp\u003e1. Fault detection and diagnosis: SVMs can be utilized to detect and diagnose faults in complex systems. By training an SVM with labeled data representing normal and faulty system behavior, the model can learn to classify new instances as either normal or faulty. This helps in identifying potential maintenance issues early on and triggering appropriate actions.\u003c/p\u003e\n\u003cp\u003e2. Anomaly detection: SVMs can be employed for anomaly detection in maintenance systems. By training the SVM on a dataset of normal system behavior, the model can learn to identify deviations from the norm. Any instances that significantly differ from the learned patterns can be flagged as anomalies, indicating potential maintenance needs.\u003c/p\u003e\n\u003cp\u003e3. Predictive maintenance: SVMs can be used to predict the remaining useful life (RUL) of equipment or components as done in this work. By training an SVM with historical data that includes information about the condition of the equipment and the time until failure, the model can learn to predict the RUL of new instances. This information helps maintenance teams plan proactive maintenance actions, minimizing downtime, maximizing equipment lifespan and increase the output.\u003c/p\u003e\n\u003cp\u003e4. Equipment health monitoring: SVMs can be applied to monitor the health of equipment by analyzing sensor data. For instance, in wind turbine components such as bearings or rotor blades, sensors might capture measurements such as temperature, vibration, or pressure that will be collected by a condition monitoring system (CMS). By training an SVM on this CMS data, it can learn to classify equipment conditions, such as normal, early degradation, or critical failure. This enables maintenance teams to take timely actions based on the SVM\u0026apos;s predictions.\u003c/p\u003e\n\u003cp\u003eIt is important to note that SVMs are just one of many machine learning algorithms that can be used in maintenance systems. The choice of algorithm depends on the specific requirements of the problem at hand, the available data, and the desired outcome.\u003c/p\u003e\n\u003cp\u003eKernels\u003c/p\u003e\n\u003cp\u003eIn Support Vector Machines (SVMs), kernels play a crucial role by allowing the model to operate effectively in high-dimensional feature spaces without explicitly computing the coordinates of data points in those spaces (Sch\u0026ouml;lkopf and Smola 2002). Kernels are functions that measure the similarity between two input vectors, working in the original input space but implicitly mapping data into a higher-dimensional feature space. This transformation can make it easier to separate classes using linear decision boundaries. SVMs can efficiently perform complex nonlinear classifications using kernel functions, enabling them to capture intricate data patterns and relationships (Cristianini and Shawe-Taylor 2000).\u003c/p\u003e\n\u003cp\u003eCommonly used kernel functions include the Linear Kernel for linear similarity, Polynomial Kernel for introducing polynomial combinations of features, Radial Basis Function (RBF) Kernel for capturing complex nonlinear relationships based on distance, and Sigmoid Kernel for learning nonlinear decision boundaries. The choice of kernel depends on the data and problem characteristics to ensure SVMs can effectively capture underlying patterns. The choice of kernel depends on the nature of the data and the problem at hand. Selecting an appropriate kernel is important to ensure that SVMs can effectively capture the underlying patterns and achieve good classification performance.\u003c/p\u003e\n\u003cp\u003eTraining\u003c/p\u003e\n\u003cp\u003eIn the maintenance of wind turbines, a crucial step is the analysis of data collected from various sensors and monitoring systems. To effectively analyze this data, it is important to preprocess the dataset and apply certain techniques. In the context of wind turbine maintenance, the following steps are considered:\u003c/p\u003e\n\u003cp\u003e1. Dataset Split: The first step is to split the dataset into a training set and a test set. This division helps evaluate the performance of the predictive models accurately. Typically, the dataset is divided into an 80% training set and a 20% test set. The training set is used to train the models, while the test set is used to assess their performance. For example, in case of having a dataset of 1000 wind turbine observations, 800 observations are selected for the training set and 200 observations for the test set (Sch\u0026ouml;lkopf et al. 2000).\u003c/p\u003e\n\u003cp\u003e2. Feature Scaling: Feature scaling is a preprocessing technique applied to normalize or standardize the features in the dataset. It improves the optimization process by ensuring that the range of values across different features is comparable. This helps algorithms converge more quickly and avoids biased influence from features with larger scales (Platt 1999). These steps are essential for effectively training predictive models in the context of wind turbine maintenance and ensuring accurate and reliable results.\u003c/p\u003e\n\u003cp\u003eFor example, in case of the wind turbine dataset includes features like wind speed, temperature, and power output. Wind speed ranges from 0 to 50 m/s, temperature ranges from -20 to 40 degrees Celsius, and power output ranges from 0 to 3 MW. By applying feature scaling techniques like normalization or standardization, it will be ensured that all features are on a similar scale, such as ranging from 0 to 1 or having a mean of 0 and a standard deviation of 1.\u003c/p\u003e\n\u003cp\u003eBy performing feature scaling, the optimization algorithms used in maintenance tasks, such as gradient descent, can work more effectively and converge towards the minimum of the cost function more efficiently. To summarize, in wind turbine maintenance, splitting the dataset into training and test sets and applying feature scaling are essential steps for accurate analysis. These practices enable better model training and optimization, leading to improved maintenance decision-making and overall turbine performance (Pandit et al. 2023).\u003c/p\u003e\n\u003cp\u003eTesting\u003c/p\u003e\n\u003cp\u003eWhen applying Support Vector Machines (SVM) to the predictive maintenance of wind turbines, specifically focusing on the slipping event of the Rotor Nacelle Assembly (RNA) due to degradation of the yaw motor brake pads based on 6 months\u0026rsquo; data to achieve a prediction period of 14 days, the analysis may involve the following considerations:\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003eAccuracy Assessment: The SVM model can be trained to classify instances of RNA slipping accurately. The accuracy of the model is evaluated by comparing the predicted RNA slipping status (e.g., slip or no slip) with the actual slip status obtained from maintenance records or sensor data.\u003c/li\u003e\n \u003cli\u003eConfusion Matrix: A confusion matrix provides a detailed breakdown of the SVM model\u0026apos;s predictions and the actual nacelle drift status. It helps identify the different types of classification errors made by the model, such as instances where RNA slipping was present but not detected (false negatives) or instances where no slipping was present but incorrectly classified as slipping (false positives).\u003c/li\u003e\n \u003cli\u003eReceiver Operating Characteristic (ROC) Curve and Area Under the Curve (AUC): By plotting the true positive rate against the false positive rate at various classification thresholds, the SVM model\u0026apos;s performance in detecting nacelle drift can be visualized using a ROC curve. The AUC quantifies the model\u0026apos;s ability to distinguish between instances of drift and non-drift, with a higher AUC indicating better performance.\u003c/li\u003e\n \u003cli\u003eCross-validation: Cross-validation is crucial in assessing the SVM model\u0026apos;s generalization capability for RNA slipping detection. The data can be split into multiple subsets, and the model is trained and evaluated on different combinations of these subsets. Cross-validation helps estimate how well the model performs on unseen wind turbine data, providing insights into its robustness and reliability.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eBy conducting a thorough analysis of the SVM model\u0026apos;s accuracy, confusion matrix, ROC curve, AUC, cross-validation results, and feature importance, maintenance teams can assess the model\u0026apos;s effectiveness in detecting nacelle drift in wind turbines. This analysis helps optimize maintenance schedules, reduce downtime, and ensure the optimal performance and longevity of wind turbine assets (Marti-Puig et al. 2019).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.2.2 Long-Short-Term Memory (LSTM)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOne of the most advanced models out there to forecast time series is the Long-Short-Term Memory (LSTM) Neural Network. The LSTM cell contributes to long-term memory in an even more performant way because it allows even more parameters to be learned (Lindemann et al. 2021). This makes it the most powerful Recurrent Neural Network (RNN) to do forecasting, especially when you have a longer-term trend in your data. LSTMs are one of the state-of-the-art models for forecasting at the moment. In this study, a LSTM architecture is applied with two layers of 50 units and a final dense layer for the forecast. The LSTM architecture provides a 30 time-period forecast with a lookback-period of 60 time-units.\u003c/p\u003e"},{"header":"3. System description","content":"\u003cp\u003eWind turbines have hundreds of sensors connected to the main controller and supervisory control and data acquisition (SCADA) system; however, there are components or sub-systems that are unsupervised and that can lead directly or indirectly to shut-downs of the wind turbine. For example, the yaw system is an integral component of the wind turbine control system. Its primary function is to optimize the alignment of the rotor-nacelle-assembly (RNA) with the prevailing wind direction and to counteract the natural tendency of the RNA to deviate from its position during operation. Significant forces act on the RNA as the rotor behaves akin to a massive gyroscope with an inherent offset from the center of the RNA. Consequently, gyroscopic forces exert a persistent influence, compelling the RNA to veer towards a downwind orientation, which, if left unchecked, would act as a destabilizing factor as described in Danish Wind Industry Assiciation (2003). Therefore, the yaw system is tasked with countering these gyroscopic forces, ensuring that the RNA maintains its ideal alignment with the wind direction while preventing undesired deviations. To ensure these two tasks (active positioning into the wind and station keeping) the yaw system must have sufficient driving and braking forces.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eA yaw system is typically composed of a number of yaw drive assemblies, consisting of electric motor\u0026ndash;gearbox assemblies connected to a drive, located in a circle around the yaw bearing that connects the RNA to the wind turbine tower and a break that is usually either electric at the end of each yaw drive assembly or hydraulic with a break disk connected to the yaw bearing. For the case considered here a fail-safe (energized to release) break is placed behind each electric motor. Due to the gear ratio of 1 : 2157 and the presence of 8 yaw drives, a relative small braking force of 40 Nm per break leads to a total braking torque of 690.25 kNm. The yaw bearing is a pre-stressed friction bearing. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e illustrates the gyroscopic effect acting on the yaw system of the turbine.\u003c/p\u003e \u003cp\u003eIn case the RNA position is not aligned to the wind direction, the yaw system is activated and the RNA is rotated towards the wind. The level of allowable yaw angle deviation to trigger the start of the yaw system depends on the average wind speed. The higher the average wind speed, the lower the allowed deviation of yaw angle to activate the system. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e depicts the allowed maximum yaw angle deviations as a function of wind speeds while Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e lists these values. Any point between the given discreet values are interpolated by the wind turbine controller and reacted accordingly.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAllowed maximum yaw angle deviations and delay times.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\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 \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWind speed level\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWind speed above\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTriggering yaw angle\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDelay time\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLevel 1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.5 m/s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e15\u0026deg;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10s\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLevel 2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6 m/s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12\u0026deg;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8s\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLevel 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8 m/s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10\u0026deg;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4s\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLevel 4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12 m/s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8\u0026deg;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2s\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLevel 5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e18 m/s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e6\u0026deg;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1s\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLevel 6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25 m/s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4\u0026deg;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0s\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\u003eIn practice, the wind direction changes very quickly or the sensors can briefly give an incorrect signal. For these reasons, an inertia is built into the whole system. The yaw motors are not switched on unless the allowable yaw deviation lasts for a certain duration. This time lag is called the switch-on delay. Typically, these delay times are reduced for higher wind speeds as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and listed in the last column of Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn addition, there is a yaw error limit which, if exceeded, triggers an alarm that will stop the turbine. Just like the other parameters the maximum yaw error depends on the wind speed too. For wind speeds different than the discreet values given the maximum yaw errors are determined by interpolation. Once the signal to stop the turbine is triggered no time delay is applied. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e depicts the maximum yaw error or the maximum yaw deviation angle allowed before triggering the stop mechanism. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e lists these values.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eMaximum yaw errors to stop turbine.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\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 \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eParameter\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWind speed\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMaximum yaw error to stop turbine\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMax error wind speed 1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 m/s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e40\u0026deg;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMax error wind speed 2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10 m/s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e35\u0026deg;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMax error wind speed 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e20 m/s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e30\u0026deg;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMax error wind speed 4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e30 m/s\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20\u0026deg;\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\u003eThe present work essentially aims at predicting the failure incidents of the brake pads of the yaw control brakes due to wear out. The brake pads are not directly supervised by a sensor and as a result, for wind speeds in the vicinity of the nominal operation wind speed of the turbine, uncontrolled yaw movements in the range of up to 1\u0026deg; per second are possible since the brakes cannot keep the RNA sufficiently stationary when worn out. This leads to extended yaw corrections hence increased wearing of the yaw system. Additionally, misaligned durations of yaw angles are increased. Yaw misalignment is one of the most critical design load case (DLC) for wind turbines (DLC 1.4 and 3.3 (IEC 61400-1 2019)) and operating under this condition shortens the life time of the wind turbine.\u003c/p\u003e \u003cp\u003eUnder normal operation conditions the brakes, as part of the yaw system, keep the nacelle in the desired position, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. Only in case of the required correction of yaw position the RNA is rotated based on the following sequence: open brakes \u0026ndash; activate motor (clockwise or counterclockwise) \u0026ndash; deactivate motor \u0026ndash; close brakes. The turbine studied in this work is equipped with a pre-stressed friction bearing limiting any uncontrolled motion in moments of open brakes hence inactivate durations of yaw motor. Any yaw movement of the RNA with closed yaw brakes is exceptional.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAs the brake pads of the yaw brakes are wearing out over time, the RNA starts to yaw under load conditions in the range of millidegrees even though the yaw brakes are engaged, see Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. Such movements increase with the wind speed, gustiness, and power production level of the wind turbine. The wind turbine controller does not monitor this behavior, only if such a movement causes a yaw misalignment above a pre-set value the controller activates the yaw system to turn the RNA back towards the wind direction. In case the yaw misalignment becomes greater than the maximum yaw error of the operation condition, the controller triggers a safety stop. As the algorithm to determine yaw misalignment is a double layered control with thresholds for safety stop depending on the wind speed level, a misalignment above the threshold may occur due to slipping. Since slipping increases with increasing wind speed, different thresholds for yaw activation or safety stop can be reached even earlier.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003ePer slipping event the RNA moves 6.75 millidegree in 99.3% of the cases. This value is also the smallest movement detectable by the wind turbine controller. The main goal is to define a threshold of acceptable movements per 10 minutes over which the brake pads should be replaced as the brake wear-outs are not supervised. Using the slipping events per 10-minute-periods detected end of November 2021 for a range of average wind speeds between 4.5 and 14.5 m/s, a quintic polynomial fit by regression analysis resulted in the function\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$y=0.0349{x}^{5}-1.5785{x}^{4}+28.223{x}^{3}-249.06{x}^{2}+1090.1x-1815.7$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere y\u0026thinsp;=\u0026thinsp;number of slipping events and x\u0026thinsp;=\u0026thinsp;wind speed cluster used in 0.5 m/s steps. It is noted that the above and following polynomial fits do not satisfy the condition that \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(y=0\\)\u003c/span\u003e\u003c/span\u003e when \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(x=0\\)\u003c/span\u003e\u003c/span\u003e; however, this is immaterial as the number of slips at zero wind speed is irrelevant to any subsequent calculation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e shows the data points and Eq.\u0026nbsp;(\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) for the range of wind speeds considered. Average and normalized maximum yaw errors in 10-minute-intervals for wind speed clusters between 3.5 and 14 m/s are plotted in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSince the average and maximum yaw errors tend to zero with increasing wind speed, the yaw error triggering wind speed level is set to 18 m/s, giving a threshold value implying a safety factor of two. Therefore, 3\u0026deg; or 444 slipping event per 10 minutes at 18 m/s is a reasonable threshold for triggering the exchange of brake pads. In order to define a polynomial function satisfying the corresponding values for lower wind speeds, Eq.\u0026nbsp;(\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) is to be scaled down to fit within the defined threshold of 444 events:\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$444=0.0349{x}^{5}-1.5785{x}^{4}+28.223{x}^{3}-249.06{x}^{2}+1090.1x-1815.7$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eSolving for x gives x \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\approx\\)\u003c/span\u003e\u003c/span\u003e 15.4563 m/s so that to obtain y= 444 for x\u0026thinsp;=\u0026thinsp;18 m/s, a shift of 18-15.4563=2.5437 m/s is necessary. Applying this shift to the polynomial results in\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$y=0.0349{\\left(x-2.5437\\right)}^{5}-1.5785{\\left(x-2.5437\\right)}^{4}+28.223{\\left(x-2.5437\\right)}^{3}-249.06{\\left(x-2.5437\\right)}^{2}+1090.1\\left(x-2.5437\\right)-1815.7$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhich now satisfies y(18)\u0026thinsp;=\u0026thinsp;444 as aimed.\u003c/p\u003e \u003cp\u003eBefore incorporating Eq.\u0026nbsp;(\u003cspan refid=\"Equ3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) into the ML model, the first step involves the preprocessing the available wind turbine data. This data typically includes wind speed, turbine operational parameters, and condition data. The output from Eq.\u0026nbsp;(\u003cspan refid=\"Equ3\" class=\"InternalRef\"\u003e3\u003c/span\u003e), which predicts the likelihood of slippage events at different wind speeds, is used as a feature in this dataset. This feature enriches the model by providing insights into the critical thresholds at which maintenance actions are necessary.\u003c/p\u003e"},{"header":"4. Arrangements and Results","content":"\u003cdiv id=\"Sec16\"\u003e\n \u003ch2\u003e4.1 Evaluation metrics\u003c/h2\u003e\n \u003cp\u003eIn order to assess the models, we use the measures of precision, recall, accuracy and F-measure, which are computed from the contents of a classic confusion matrix of the classification predictions. True positive and false positive cases are denoted as TP and FP, while true negative and false negative are denoted as TN and FN respectively. In order to fit the classification evaluation in incident detection problem, we assign the classes no-incident and incident.\u003c/p\u003e\n \u003cp\u003ePrecision is the ratio of the predicted true positive cases (TP) to the sum of true positives (TP) and false positives (FP) as\u003c/p\u003e\n \u003cdiv id=\"Equa\"\u003e\n \u003cdiv id=\"FileID_Equa\" name=\"EquationSource\"\u003e$$\\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}= \\frac{\\text{T}\\text{P}}{\\text{T}\\text{P}+\\text{F}\\text{P}}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eRecall is the ratio of the true positive cases to the sum of true positives (TP) and false negatives (FN):\u003c/p\u003e\n \u003cdiv id=\"Equb\"\u003e\n \u003cdiv id=\"FileID_Equb\" name=\"EquationSource\"\u003e$$\\text{R}\\text{e}\\text{c}\\text{a}\\text{l}\\text{l}= \\frac{\\text{T}\\text{P}}{\\text{T}\\text{P}+\\text{F}\\text{N}}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eAccuracy is the ratio of the total number of predictions that were correct.\u003c/p\u003e\n \u003cdiv id=\"Equc\"\u003e\n \u003cdiv id=\"FileID_Equc\" name=\"EquationSource\"\u003e$$\\text{A}\\text{c}\\text{c}\\text{u}\\text{r}\\text{a}\\text{c}\\text{y}= \\frac{\\text{T}\\text{P}+\\text{T}\\text{N}}{\\text{T}\\text{P}+\\text{F}\\text{P}+\\text{T}\\text{N}+\\text{F}\\text{N}}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003ePrecision or recall alone cannot describe a classifier\u0026apos;s efficiency; therefore, F-measure is introduced as a combination of these two metrics. It is defined as twice the harmonic mean of precision and recall and is the metric which is most frequently referred.\u003c/p\u003e\n \u003cdiv id=\"Equd\"\u003e\n \u003cdiv id=\"FileID_Equd\" name=\"EquationSource\"\u003e$$\\text{F}-\\text{m}\\text{e}\\text{a}\\text{s}\\text{u}\\text{r}\\text{e}= \\frac{2 \\times \\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n} \\times \\text{R}\\text{e}\\text{c}\\text{a}\\text{l}\\text{l}}{\\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}+ \\text{R}\\text{e}\\text{c}\\text{a}\\text{l}\\text{l}}$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eA value closer to unity indicates better combined precision and recall of the classifier, whereas lower values imply lower accuracy or precision or both.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec17\"\u003e\n \u003ch2\u003e4.2 SVM\u003c/h2\u003e\n \u003cp\u003eBased on the provided results for identifying and predicting using different SVM kernels (Polynomial, RBF, Sigmoid), the need for maintenance events due to worn yaw motor brake pads causing extensive RNA slipping beyond pre-defined limits the following conclusions can be drawn.\u003c/p\u003e\n \u003cp\u003e1. Precision: The RBF kernel demonstrates the highest precision (75.30%), followed by the Polynomial kernel (68.86%) and the Sigmoid kernel (50.75%). Precision represents the percentage of correctly identified maintenance events among all instances predicted as maintenance events. Higher precision indicates a lower rate of false positives, meaning that the Polynomial kernel performs better in accurately identifying true maintenance events.\u003c/p\u003e\n \u003cp\u003e2. Recall: The RBF kernel also exhibits the highest recall (74.10%), followed by the Polynomial kernel (67.80%), and the Sigmoid kernel (50.75%). Recall represents the percentage of correctly identified maintenance events out of all actual maintenance events. Higher recall indicates a lower rate of false negatives, implying that the Polynomial kernel is better at capturing a larger portion of actual maintenance events.\u003c/p\u003e\n \u003cp\u003e3. Accuracy: The accuracy is highest for the RBF kernels (74.10%), followed by the Polynomial kernel (67.80%) and the Sigmoid kernel (50.75%). Accuracy represents the overall percentage of correctly classified instances, regardless of the class. However, it is important to note that accuracy alone might not be sufficient to evaluate the performance of the model, especially in imbalanced datasets.\u003c/p\u003e\n \u003cp\u003e4. F-measure: The F-measure, which combines precision and recall into a single metric, is highest for the RBF kernel (73.74%), followed by the Polynomial kernel (67.32%) and the Sigmoid kernel (50.75%). The F-measure provides a balanced assessment of both precision and recall, indicating the Polynomial kernel\u0026apos;s superior performance in maintaining a good trade-off between identifying true maintenance events and minimizing false positives and false negatives.\u003c/p\u003e\n \u003cp\u003eBased on these results, the RBF kernel, considering its higher precision, appears to be the most effective among the three kernels in identifying maintenance needs related to yaw brake pad degradation, recall, and F-measure.\u003c/p\u003e\n \u003cdiv\u003e\n \u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 3\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eResult analyses per kernel\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePolynomial kernel\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRBF kernel\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSigmoid kernel\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\u003e\u003cstrong\u003ePrecision (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e68.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e75.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e50.75\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e67.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e74.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e50.75\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e67.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e74.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e50.75\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eF-measure (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e67.32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e73.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e50.75\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eThe developed code in python is able to predict RNA slipping events with the results shown in Table\u0026nbsp;3.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec18\"\u003e\n \u003ch2\u003e4.3 LSTM\u003c/h2\u003e\n \u003cp\u003eLSTM models are adept at handling time-series data, making them well-suited for predictive maintenance tasks. The preprocessed dataset, now including the polynomial function\u0026apos;s output, is fed into the LSTM model. The model learns from the patterns in the data, including the relationship between wind speed, yaw system behavior, and the occurrence of slipping events, as indicated by Eq.\u0026nbsp;(3). Moreover, LSTMs are capable of learning from sequences of data, capturing temporal dependencies that are crucial in predicting maintenance requirements. By training the LSTM model on the dataset that includes the polynomial function\u0026apos;s output (Eq.\u0026nbsp;3), the model can learn how the risk of slippage events evolves over time and under varying operational conditions. In summary, Eq.\u0026nbsp;(3) enhances the LSTM model\u0026apos;s capability to predict maintenance requirements by providing a critical piece of information about the yaw system\u0026apos;s behavior. This integration leads to a more robust and reliable predictive maintenance system, helping to reduce downtime and improve the efficiency of offshore wind turbines.\u003c/p\u003e\n \u003cdiv id=\"Sec19\"\u003e\n \u003ch2\u003e4.3.1 Data preparation\u003c/h2\u003e\n \u003cp\u003eIn order for the LSTM to provide sufficient and more valuable prediction, the initial dataset is reformed to a minute-period so as the final prediction to make a state. Thus, the input dataset to the LSTM model is given in the Fig.\u0026nbsp;9 below, where zeros point a non-incident and ones an incident case.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec20\"\u003e\n \u003ch2\u003e4.3.2 Simulation setup - Model training\u003c/h2\u003e\n \u003cp\u003eFor the initial data, the re-scale parameter is 1/255 and the validation split is set at 20% (training 80%-validation 20%). The final 30 time-units of the initial dataset were not fed for training and validation to avoid the model making predictions and testing in unknown territories. The tested LSTM network is trained for 50 epochs, with the callback that the best model is saved when it appears. Figure\u0026nbsp;10 below, provides the training and the validation loss, where these are the metrics used to assess how a deep learning model fits the training data and the performance of a deep learning model on the validation set.\u003c/p\u003e\n \u003cp\u003eThe purpose of visualizing together the training and validation loss is to diagnose the model\u0026rsquo;s performance and identify which aspects are needed for tuning if required. Figure\u0026nbsp;10 shows that training and validation loss both decrease and start stabilizing at a point above 40 epochs. This indicates that the proposed LSTM model has an optimal fit. Table\u0026nbsp;4 below provides the evaluation metrics for training and validation set, where one can see that the LSTM model performs close to optimal because F-measure is above 96% for both sets.\u003c/p\u003e\n \u003cdiv\u003e\n \u003ctable id=\"Tab4\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 4\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eEvaluation metrics for training and validation set.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSet\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePrecision (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRecall (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAccuracy (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eF-measure (%)\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\u003e\u003cstrong\u003eTraining\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e98.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e96.27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e93.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e97.56\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e95.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e96.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e93.21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e96.12\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec21\"\u003e\n \u003ch2\u003e4.3.3 Experimental results of the integrated methodology\u003c/h2\u003e\n \u003cp\u003eTable\u0026nbsp;5 provides the evaluation metrics of the forecasting of LSTM in test set. These experimental results clearly show that the forecasting performance of the proposed architecture is high, where the prediction accuracy is over 98% and the overall f-measure is above 96%.\u003c/p\u003e\n \u003cdiv\u003e\n \u003ctable id=\"Tab5\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 5\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eEvaluation metrics for test set.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colspan=\"4\"\u003e\n \u003cp\u003eTesting set\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\u003e\u003cstrong\u003ePrecision (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eF-measure (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e96.11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e96.37\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e98.36\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e96.22\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\u003eFigure 11 (a) illustrates a representation of the projected incident occurrences, while (b) depicts a combination of both the initial incident data and the forecasted incidents.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"5. Conclusions and future work","content":"\u003cp\u003eBy integrating domain expertise with machine learning techniques, this article reveals the potential to significantly enhance predictive maintenance practices in the offshore wind industry, ultimately leading to increased turbine reliability, reduced operational costs, and a more sustainable energy future. With carefully selected variables from the control system and the development of features by combining relevant variables, unsupervised components or events that can lead to shut down of the wind turbine can be converted into indirectly supervised parts. Consequently, the RTTF can be predicted with a fair accuracy thus increasing the reliability of the wind turbine and its technical availability. An important aspect shown is that by using LSTM the prediction time can be increased to a period mandatory in offshore wind industry due to a vast number of constrains like weather conditions, spare part management, and ship and crew availability. The prediction abilities of SVM for this task are significant lower making LSTM the better choice to reach reliable forecast results. Applying this method to other unsupervised components or sub-systems requires careful development of features based on available variables recorded in the data set from the wind turbine control system.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor Contributions:\u003c/strong\u003e Conceptualization: Uwe L\u0026uuml;tzen; Methodology: Serdar Beji; Formal analysis and investigation: Uwe L\u0026uuml;tzen; Software, writing\u0026mdash;original draft: Uwe L\u0026uuml;tzen; Writing\u0026mdash;review and editing: Serdar Beji; Supervision: Serdar Beji.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAll authors have read and agreed to the published version of the manuscript.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e No funding was received for conducting this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability Statement:\u003c/strong\u003e The data presented in this study are available on request from the corresponding author in aggregated form. The data are not publicly available due to restrictions from the owner of the wind turbine.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments:\u003c/strong\u003e This work was carried out as a part of doctoral studies of the first author at Istanbul Technical University.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflicts of Interest:\u003c/strong\u003e The authors declare no conflict of interest.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCevasco D, Koukoura S, Kolios A J (2021) Reliability, availability, maintainability data review for the dientfification of trends in offshore wind energy applications. Renewable and Sustainable Energy Reviews, vol. 136. https://doi.org/10.1016/j.rser.2020.110414\u003c/li\u003e\n\u003cli\u003eChen J-h, Pei A-g, Chen P, Hu Z-q (2021) Study on Gyroscopic Effect of Floating Offshore Wind Turbines. Chinese Ocean Engineering 35: 201-214. https://doi.org/10.1007/s13344-021-0018-z\u003c/li\u003e\n\u003cli\u003eCortes C, Vapnik V (1995) Support-vector networks. Machine Learning 20: 273-297. https://doi.org/10.1007/BF00994018\u003c/li\u003e\n\u003cli\u003eCristianini N, Shawe-Taylor J (2000) An Introduction to Support Vector Machines and Other Kernel-based Learning Methods. Cambridge University Press. https://doi.org/10.1017/CBO9780511801389\u003c/li\u003e\n\u003cli\u003eDanish Wind Industry Association (2003) Wind Turbines: Upwind or Downwind Machines?,\u0026quot; Danish Wind Industry Association. http://www.windpower.org/en/tour/design/updown.htm. Accessed 28 October 2022.\u003c/li\u003e\n\u003cli\u003eDinwoodie I, Endrerud O-E, Hofmann M, Martin R, Bakken Sperstad I (2015) Reference Cases for Verification of Operation and Maintenance Simulation Models for Offshore Wind Farms. Wind Engineering 39: 1-14. https://doi.org/10.1260/0309-524X.39.1.1\u003c/li\u003e\n\u003cli\u003eDinwoodie I, McMIllan D, Revie M, Lazakis I, Dalgic Y (2013) Development of a Combined Operational and Strategic Decision Support Model for Offshore Wind. Energy Procedia 35: 157-166. https://doi.org/10.1016/j.egypro.2013.07.169\u003c/li\u003e\n\u003cli\u003eDouard F, Domecq C, Lair W (2012) A Probabilistic Approach to Introduce Risk Measurement Indicators to an Offshore Wind Project Evaluation \u0026ndash; Improvement to an Existing Tool. Energy Procedia 24: 255-262. https://doi.org/10.1016/j.egypro.2012.06.107\u0026amp;nbsp;\u003c/li\u003e\n\u003cli\u003eEndrerud O-E, Liyanage J P, Keseric N (2014) Marine Logistics Decision Support for Operation and Maintenance of Offshore Wind Parks with a Multi Method Simulation Model. Winter Simulation Conference, Savannah, Georgia. https://doi.org/10.1109/WSC.2014.7020021\u003c/li\u003e\n\u003cli\u003eFarrar N O, Ali M H, Dasgupta D (2023) Artificial Intelligence and Machine Learning in Grid Connected Wind Turbine Control Systems: A Comprehensive Review. Energies 16. https://doi.org/10.3390/en16031530\u003c/li\u003e\n\u003cli\u003eFeuchtwang J, Infield D (2013) Offshore Wind Turbine Maintenance Access: A Closed-Form Probabilistic Method for Calculating Delays Caused by Sea-State. Wind Energy 16: 1049-1066. http://dx.doi.org/10.1002/we.1539\u003c/li\u003e\n\u003cli\u003eGuyon I, Elisseeff A (2003) An introduction to variable and feature selection. Journal of Machine Learning Research 3: 1157-1182. https://doi.org/10.1162/153244303322753616\u003c/li\u003e\n\u003cli\u003eGWEC (2020) Global Wind Report 2019. Brussels.\u003c/li\u003e\n\u003cli\u003eHalvorsen E E, Gundegjerde C, Halvorsen I B, Hvattum L M, Non\u0026aring;s L M (2013) Vessel fleet analysis for maintenance operations at offshore wind farms. Energy Procedia 35: 167-176. https://doi.org/10.1016/j.egypro.2013.07.170\u003c/li\u003e\n\u003cli\u003eHavrlant L, Kreinovich V (2017) A Simple Probabilistic Explanation of Term Frequency-Inverse Document Frequency (tf-idf) Heuristic (and Variations Motivated by This Explanation). International Journal of General Systems 46: 27-36. https://doi.org/10.1080/03081079.2017.1291635\u003c/li\u003e\n\u003cli\u003eHe H, Garcia E A (2009) Learning from imbalanced data. IEEE Transactions on Knowledge and Data Engineering 21: 1263-1284. https://doi.org/10.1109/TKDE.2008.239\u003c/li\u003e\n\u003cli\u003eHofmann M (2011) A Review of Decision Support Models for Offshore Wind Farms with an Emphasis on Operation and Maintenance Strategies. Wind Engineering 35: 1-16. https://doi.org/10.1260/0309-524X.35.1.1\u003c/li\u003e\n\u003cli\u003eIEC (2019) IEC 61400-1: Wind Turbinesepart 1: Design Requirements. International Electrotechnical Commission, Geneva.\u003c/li\u003e\n\u003cli\u003eInternational Energy Agency (2019) Market Report Series: Renewables 2019. International Energy Agency.\u003c/li\u003e\n\u003cli\u003eIoannou A, Angus A, Brennan F (2019) Informing parametric risk control policies for operational uncertainties of offshore wind energy assets. Ocean Engineering 177: 1-11. https://doi.org/10.1016/j.oceaneng.2019.02.058\u003c/li\u003e\n\u003cli\u003eKang J, Wang Z, Guedes Soares C (2020) Condition-Based Maintenance for Offshore Wind Turbines Based on Support Vectro Machine. Energies 13, no. 3518. https://doi.org/10.3390/en13143518\u003c/li\u003e\n\u003cli\u003eKaryotakis A (2011) On the Optimisation of Operation and Maintenance Strategies for Offshore Wind Farms. Dissertation, University College London.\u003c/li\u003e\n\u003cli\u003eKessler E H (2013) Encyclopedia of Management Theory. SAGE Publications ,Thousand Oaks, California. \u003c/li\u003e\n\u003cli\u003eKhalid S, Khalil T, Nasreen S (2014) A Survey of Feature Selection and Feature Extraction Techniques in Machine Learning. Science and Information Conference, London. https://doi.org/10.1109/SAI.2014.6918213\u003c/li\u003e\n\u003cli\u003eKumar A (2022) Data Preprocessing Steps in Machine Learning. Data Analytics. https://vitalflux.com/data-preprocessing-steps-in-machine-learning/. Accessed 16 04 2023.\u003c/li\u003e\n\u003cli\u003eLazakis I, Kahn S (2021) An optimization framework for daily route planning and scheduling of maintenance vessel activities in offshore wind farms. Ocean Engineering 225. https://doi.org/10.1016/j.oceaneng.2021.108752\u003c/li\u003e\n\u003cli\u003eLindemann B, M\u0026uuml;ller T, Vietz H, Jazdi N, Weyrich M (2021) A survey on long short-term memory networks for time series prediction. Science Direct 99: 650-655. https://doi.org/10.1016/j.procir.2021.03.088\u003c/li\u003e\n\u003cli\u003eMarti-Puig P, Blanco-M A, Cardenas J J, Cusido J, Sole-Casals J. (2019) Feature Selection Algorithms for Wind Turbine Failure Prediction. Energies 12: 453. https://doi.org/10.3390/en12030453\u003c/li\u003e\n\u003cli\u003eMasoumi M (2023) Machine Learning Solutions for Offshore Wind Farms: A Review of Applications and Impact. Journal of Marine Science and Engineering 11. https://doi.org/10.3390/jmse11101855\u003c/li\u003e\n\u003cli\u003eNivet T, Muk-Pavik E. (2015) Optimization of O\u0026amp;M for Offshore Wind Farms Modelling. WMTC Conferences. University College London. https://doi.org/10.5957/WMTC-2015-214\u003c/li\u003e\n\u003cli\u003ePandit R K, Astolfi D, Cardenas I D (2023) A Review of Predictive Techniques Used to Support Decision Making for Maintenance Operations of Wind Turbines. Energies 16. https://doi.org/10.3390/en16041654\u003c/li\u003e\n\u003cli\u003ePapatzimos A K (2019) Data-driven Operations \u0026amp; Maintenance for Offshore Wind Farms: Tools and Methodologies. Dissertation, University of Edinburgh. http://dx.doi.org/10.7488/era/17\u003c/li\u003e\n\u003cli\u003ePerez Granados D, Ortega Ruiz M A, Moreira Acosta J, Gama Lara S A, Gonzalez Dominguez R A, Paramo Kanetas P J (2023) A Wind Turbine Vibration Monitoring System for Predictive Maintenance Based on Machine Learning Methods Developed under Safely Controlled Laboratory Conditions. Energies 16. https://doi.org/10.3390/en16052290\u003c/li\u003e\n\u003cli\u003ePlatt J (1999) Fast training of support vector machines using sequential minimal optimization. Advances in Kernel Methods: Support Vector Learning, pp. 185-208. \u003c/li\u003e\n\u003cli\u003eRademakers L W M M, Braam H, M. B. Zaaijer M B, van Bussel G J W (2003) Assessment and Optimisation of Operation and Maintenance of Offshore Wind Turbines. European Wind Energy Conference, Madrid. \u003c/li\u003e\n\u003cli\u003eRusu E, Onea F (2023) The Expected Dynamics of the European Offshore Wind Sector in the Climate Context. Journal of Marine Science and Engineering 11. https://doi.org/10.3390/jmse11101967\u003c/li\u003e\n\u003cli\u003eScheu M, Matha D, Hofmann M, Muskulus M (2012) Maintenance strategies for large offshore wind farms. Energy Procedia 24: 281-288. https://doi.org/10.1016/j.egypro.2012.06.110\u003c/li\u003e\n\u003cli\u003eSch\u0026ouml;lkopf B, Smola A J (2002) Learning with kernels: Support vector machines, regularization, optimization, and beyond. MIT Press. https://doi.org/10.7551/mitpress/4175.001.0001\u003c/li\u003e\n\u003cli\u003eSch\u0026ouml;lkopf B, Smola A, Williamson R, Bartlett P (2000) New support vector algorithms. Neural Computation 12: 1207-1245. https://doi.org/10.1162/089976600300015565\u0026amp;nbsp;\u003c/li\u003e\n\u003cli\u003eSetiawan B, Djanali S, Ahmad T (2019) Increasing Accuracy and Completeness of Intrusion Detection Model Using Fusion of Normalization, Feature Selection Method and Support Vector Machine. International Journal of Intelligent Engineering \u0026amp; Systems 12: 378 - 389. https://doi.org/ 10.22266/ijies2019.0831.35\u003c/li\u003e\n\u003cli\u003eSeyr H, Muskulus M (2019) Decision Support Models for Operations and Maintenance for Offshore Wind Farms: A Review. Applied Science 9. https://doi.org/10.3390/app9020278\u003c/li\u003e\n\u003cli\u003eSperstad I B, St\u0026aring;lhane M, Dinwoodie I, Endrerud O-E V, Martin R, Warner E (2017) Testing the robustness of optimal access vessel fleet selection for operation and maintenance of offshore wind farms. Ocean Engineering 145: 334-343. https://doi.org/10.1016/j.oceaneng.2017.09.009\u003c/li\u003e\n\u003cli\u003eStahlhane M, Hvattum L M, Skaar V (2015) Optimization of routing and scheduling of vessels to perform maintenance at offshore wind farms. Energy Procedia 80: 92-99. https://doi.org/10.1016/j.egypro.2015.11.411\u003c/li\u003e\n\u003cli\u003eTang M, Zhao Q, Ding S X, Wu H, Li L, Long W, Huang B (2020) An Improved LightGBM Algorithm for Online Fault Detection of Wind Turbine Gearboxes. Energies 13. https://doi.org/10.3390/en13040807\u003c/li\u003e\n\u003cli\u003eYu X (2016) Modelling Offshore Wind Farm Operation and Maintenance with View to Estimating the Benefits of Condition Monitoring. Dissertation, University of Strathclyde.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Footnotes","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003e TF-IDF (Term Frequency - Inverse Document Frequency) is \u003cb\u003ea handy algorithm that uses the frequency of words to determine how relevant those words are to a given document\u003c/b\u003e. It's a relatively simple but intuitive approach to weighting words, allowing it to act as an efficient starting point for a variety of tasks. (Hacrlant and Kreinovich 2017)\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"journal-of-ocean-engineering-and-marine-energy","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"oeme","sideBox":"Learn more about [Journal of Ocean Engineering and Marine Energy](http://link.springer.com/journal/40722)","snPcode":"40722","submissionUrl":"https://submission.nature.com/new-submission/40722/3","title":"Journal of Ocean Engineering and Marine Energy","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Offshore wind energy, Predictive maintenance, Condition monitoring, Prediction period, unsupervised components, asset degradation, predictive asset degradation patterns","lastPublishedDoi":"10.21203/rs.3.rs-3906932/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3906932/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eEnterprises in increasing numbers allocate substantial expenses to offshore wind energy development as a pivotal component of the global energy transition from fossil fuels, hence the importance of ensuring the reliability of offshore wind technology becomes ever more significant. At the same time, operation and maintenance (O\u0026amp;M) of offshore wind farms are progressively focusing on the integration of artificial intelligence (AI) for enhancing the efficiency and performance of the wind energy facilities. Decision support strategies based on failure predictions are an important element in this trend. As a result, AI is more frequently used to create time-to-failure predictions based on large amount of data collected from sensors deployed to wind turbines. Nevertheless, unsupervised components or subsystems may occasionally lead to failures. This paper presents a real-life example that failures in unsupervised components can be reliably predicted by the use of AI. Two different methods, Support Vector Machine and Long Short Term Memory, are presented and their limitations and advantages discussed.\u003c/p\u003e","manuscriptTitle":"Predictive Maintenance for Offshore Wind Turbines through Deep Learning and Online Clustering of Unsupervised Subsystems: A Real-World Implementation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-02-01 13:28:58","doi":"10.21203/rs.3.rs-3906932/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-04-06T16:58:46+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-04-02T12:36:04+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"27cc0aaf-9e0c-4e99-b5a6-fb717c85139b","date":"2024-03-18T06:59:21+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-03-02T12:32:45+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"125b677f-6cb0-46e7-bf6a-7e5a25cfdf44","date":"2024-02-23T04:05:35+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"590d7e42-36e3-4052-a670-10e00e7e3be4","date":"2024-02-21T20:08:36+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-02-04T04:55:16+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-01-30T14:54:28+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-01-30T14:54:27+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of Ocean Engineering and Marine Energy","date":"2024-01-28T21:26:18+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-ocean-engineering-and-marine-energy","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"oeme","sideBox":"Learn more about [Journal of Ocean Engineering and Marine Energy](http://link.springer.com/journal/40722)","snPcode":"40722","submissionUrl":"https://submission.nature.com/new-submission/40722/3","title":"Journal of Ocean Engineering and Marine Energy","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"bb84962b-c38e-4e26-8326-0029c07dfc65","owner":[],"postedDate":"February 1st, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2024-06-17T00:33:23+00:00","versionOfRecord":{"articleIdentity":"rs-3906932","link":"https://doi.org/10.1007/s40722-024-00335-z","journal":{"identity":"journal-of-ocean-engineering-and-marine-energy","isVorOnly":false,"title":"Journal of Ocean Engineering and Marine Energy"},"publishedOn":"2024-06-16 00:33:23","publishedOnDateReadable":"June 16th, 2024"},"versionCreatedAt":"2024-02-01 13:28:58","video":"","vorDoi":"10.1007/s40722-024-00335-z","vorDoiUrl":"https://doi.org/10.1007/s40722-024-00335-z","workflowStages":[]},"version":"v1","identity":"rs-3906932","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3906932","identity":"rs-3906932","version":["v1"]},"buildId":"cBFmMYwuxLRRLfASyISRj","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}