Predictive Maintenance for Wind Turbines: Leveraging Sensor Data with Traditional and Novel Machine Learning Techniques

Predictive Maintenance for Wind Turbines: Leveraging Sensor Data with Traditional and Novel Machine Learning Techniques⋆ Juan Vanerio, Pedro Casas, Axel Weissenfeld AIT Austrian Institute of Technology, Vienna, Austria [email protected] Abstract. Predictive maintenance has emerged as a critical strategy to improve the efficiency and reliability of wind turbines. Using advanced technologies such as machine learning and data analytics, wind farm operators can detect potential problems early, optimize maintenance schedules, and reduce downtime. Through monitoring and analysis of industry standard Supervisory Control and Data Acquisition (SCADA) sensor data, predictive maintenance enables proactive interventions that extend the useful life of turbines and improve operational performance. This paper presents a novel anomaly detection framework for predicting the normal operating range of generator bearing temperature using covariates derived from current and lagged SCADA signals. Synthetic anomalies are injected across the full variable range during training, enabling a self-supervised XGBoost model to distinguish between normal and anomalous behavior. The approach is validated using a one-year SCADA dataset from a wind farm comprising 10 wind turbine generators. Results highlight the framework’s effectiveness in identifying early signs of generator failure, supporting predictive maintenance and improving overall turbine performance and reliability. Keywords: Predictive Maintenance ·Wind Turbines ·Machine Learning ·Data Analytics ·SCADA Systems 1 Introduction Wind turbines are critical assets in renewable energy production, but their operation is often challenged by mechanical failures. Predictive maintenance offers a proactive approach to identify potential issues before they escalate, reducing downtime and maintenance costs while improving operational efficiency. The wind energy sector faces unique challenges, including the remote locations of wind farms (WFs), harsh operational conditions, and the large-scale ⋆This work has been funded by the Austrian Research Promotion Agency (FFG) under grant No. FO999913202 UNDERPIN and by the European Commission under contract No. 101123179 UNDERPIN. deployment of turbines. These factors make maintenance a costly and logistically complex task, accounting for 20-30% of the total power generation expenses [1]. Traditional maintenance practices, such as reactive and periodic maintenance, have significant limitations that predictive maintenance aims to address. For instance, reactive maintenance involves performing maintenance only after a failure occurs, leading to increased downtime and additional costs due to unplanned outages. In the case of periodically scheduled maintenance, the strategy may lead to unnecessary interventions and costs. On the other hand, predictive maintenance is a proactive approach leveraging sensor data analysis and machine learning to identify potential issues and predict component failures in advance, allowing operators to take corrective actions before problems escalate. Predictive maintenance offers significant benefits, including reduced downtime, extended asset lifespan, and enhanced operational efficiency [2]. Timely failure prediction of critical components such as generators is vital for consistent electricity production. For example, previous studies found that 68% of downtime in two Chinese wind farms stemmed from generator, converter, and blade issues [3]. The solution proposed in this paper empowers operators to monitor abnormalities proactively, ensuring early intervention and improved turbine reliability. The literature already presents several effective approaches using different model types, such as LSTM and XGBoost, for prediction tasks based on SCADA data. However, the performance differences between models are generally small and often depend on the specific wind turbine [4]. Our study focuses on a yearlong dataset of SCADA sensor measurements, recorded at 10-minute intervals, from a single wind turbine generator (WTG) within a wind farm. During the observation period, five major fault events were identified through system logs and operator reports. As a result, the turbine experienced reduced or halted operation for a total of 38 days. These findings underscore the importance of early failure prediction to enable timely maintenance, thereby minimizing downtime and associated costs. 2 Methodology This section outlines the methods used to process the data for the proposed predictive maintenance approach, which aims to identify potential failures in advance. 2.1 Data Pre-processing First, incorrect timestamps and inconsistent or out-of-range values in the original data were identified and corrected as part of basic consistency cleanup. Data rows containing erroneous data that could not be corrected are removed. Data points outside the operational wind speed range (5 m/s to 25 m/s), as well as those with a blade pitch angle exceeding 40°, or operating under restricted power limits or grid demands, are excluded. Additionally, values exceeding five standard deviations are removed. To further refine the dataset, an Isolation Forest model is employed to identify and remove outliers. These exclusions ensure that the training dataset only includes periods when the WTG was likely operating under normal conditions. Furthermore, data augmentation is applied by marking all data points within a seven-day window before and after a major failure as away_from_anomalies (afa) with a value of 0, and all other points with a value of 1. This new column reflects the proximity to failure, allowing for the exclusion of data points near failures from the training set, as their feature values may be altered during such events. 2.2 Feature Selection A careful exploration of the dataset, accompanied by a literature review, indicates that the sensor readings that are representative of the WTG health can be predicted from a relatively small set of ambient readings and historical values of the predicted variables. For instance, generator bearing temperature is frequently used to predict generator health [4]. The ambient variables used as input features for prediction are listed in we consider the ambient variables of Table 1. The remaining input features are the previous value of the target variables (which we consider new augmentation variables and append their names with ‘_lag’). The target variables to be learned are described in Table 2. Variable Description Amb_Temp_Avg Average ambient temperature Amb_WindDir_Abs_Avg Absolute wind direction average Amb_WindDir_Relative_Avg Relative wind direction average Amb_WindSpeed_Avg Average wind speed Amb_WindSpeed_Max Maximum wind speed Amb_WindSpeed_Min Minimum wind speed Amb_WindSpeed_Std Standard deviation of wind speed Table 1. Ambient variables used as input features for prediction. By training the model to predict these target variables, each data point can be inferred from its preceding state and the latest ambient sensor readings. This approach enhances robustness by reducing dependence on any single potentially noisy or corrupted input. Moreover, the model implicitly learns the underlying dynamics of the system, knowledge that can be effectively leveraged for anomaly detection and predictive maintenance. Variable Description Grd_Prod_PsblePwr_Avg Possible power production from the grid Grd_Prod_Pwr_Avg Average power production from the grid Blds_PitchAngle_Avg Average pitch angle of the blades Gear_Bear_Temp_Avg Average temperature of the gearbox bearing Gen_RPM_Avg Average rotational speed of the generator Gen_RPM_Std Standard deviation of generator RPM Hyd_Oil_Temp_Avg Average temperature of hydraulic oil Gear_Oil_Temp_Avg Average temperature of gearbox oil Rtr_RPM_Avg Average rotational speed of the rotor HVTrafo_Phase1_Temp_Avg Average temperature of HV transformer phase Table 2. Target variables for prediction. Finally, to perform predictive maintenance, we aim to predict in advance the occurrence of anomalous events. For this, we propose a novel method that directly addresses the high imbalance between normal operation samples and anomalous samples, the latter of which may even not be present in the data. These normal samples are identified by the variable afa taking a value of 1, and are expected to concentrate in specific regions, i.e., to live in some manifold(s) of the data space. Therefore, if we where to uniformly inject synthetically created anomalous samples (tagged with afa equal to 0) spanning the whole data range of the normal samples, the ratio of normal vs anomalous samples across regions of the data space would either see a mixture of normal and anomalous samples (in the manifold region) or a large majority of anomalous samples (outside of it). Figure 1 shows an illustrative examples of the described setup. It can be seen there that the anomalous samples can be located anywhere, even inside the normal data regions. However, differently form normal data points they are expected not to concentrate anywhere, and so different density regions emerge and can be learned. We propose to use a tabular machine learning model to learn to discriminate between these patterns, which maps to discriminate between normal and anomalous data regions. 2.3 Model Training The dataset pruned according to the rules from section 2.1 contains data considered as ‘normal’, and is therefore split into two consecutive fragments of 60% and 40% of the total length for training and test purposes, respectively, to ensure the models generalize well to unseen data. First, a Histogram-Based Gradient Boosting Regressor (HGBR) model is built for predicting each one of the targets from Table 2. These models are implementations of the LightGBM tree-based model [5], and are trained on the Fig. 1. Anomaly detection method. The figure represents a two-dimensional data space, where normal samples observed from the real data are represented as green circles and artificially generated anomalous samples are represented as red triangles. The green region is the real regions where normal samples may be found, and the solid black lines show the observed data range for normal samples. Normal samples naturally concentrate in a few regions while synthetic anomalous samples spread across the data space, which help in identifying the normality region. pre-processed and filtered training dataset. Cross-validation was used to tune hyperparameters and select the best-performing models. Second, a normality detector is deployed to predict the values from afa as a proxy variable for the absence of failures. A common challenge in anomaly detection is the scarcity—or complete absence—of ground truth labels for anomalous events. In this study, while a proxy variable provides some indication of anomalies, these samples are limited in diversity and exhibit high temporal correlation, as they span contiguous time intervals. Consequently, they represent only a subset of possible anomaly types. In contrast, ground truth data for normal operating conditions is typically abundant. To address this imbalance, the proposed method introduces synthetic anomalies that span the entire feature space and leverages the higher density and consistency of normal samples on their underlying manifold to distinguish anomalous from normal behavior. To this end, the input features are first standardized using a quantile scaler to normalize their value ranges. Next, we introduce a novel iterative technique designed to compute a score that reflects the likelihood of imminent failures: 1. Consider each original data value, representing a normal operation sample, as a ‘positive’ instance and perturb it with Gaussian noise at 5% of the feature deviation, 2. Synthesize ‘negative’ samples (representing abnormal operation points) by sampling values uniformly in the data range of each feature in the training dataset and assigning a zero to their afa parameter (i.e., marking them as anomalous), 3. Use the newly created dataset to (iteratively) train an XGBoost Classifier [6] model to predict the negative class, 4. Repeat for a fixed number of steps or if the F1-score on training does not increase. The XGBoost model is selected due to its strong performance in classification tasks on tabular data and its support for incremental (iterative) training. In this work, the model is trained to output the probability that a given data point corresponds to a period of normal operation. This setup, referred to as the Normality Detector (ND), is trained on a dataset in which normal and synthetically generated anomalous samples partially overlap. Consequently, the model’s output probabilities are confined to a narrower subrange within [0, 1], rather than covering the entire spectrum. Despite this compression, lower probability values are still interpreted as indicative of normal operating conditions. To prevent overfitting to the synthetic anomalies, a low-complexity model is used. This encourages rapid regularization and better generalization by effectively isolating regions of normal operation in the feature space, while treating the remaining regions as anomalous. 2.4 Model Evaluation The evaluation of the models for predicting target variables is performed on the previously split testing dataset, where the data is expected to be representative of normal operation. The normality detector is instead deployed on all available data – recall that it was trained only on normal data from the initial 60% of the available time. 3 Preliminary Evaluation For evaluation, this study focuses on multivariate time-series data collected from a single wind turbine generator (WTG) in an onshore wind farm that has been operational since 2009. The dataset comprises SCADA measurements recorded throughout the year 2016 at 10-minute intervals. Where applicable, each time step includes statistical aggregates such as minimum, maximum, average, and standard deviation values for key parameters. During the observation period, five major generator failures were recorded, each flagged by high-severity alarms. To visually represent the available data, Figure 2 shows the scatter plot of the empirical data of ambient wind speed against the WTG produced power. The data dots in blue correspond to data that has been labeled by the rulebased logic as normal operation points, while the red dots show data outside of the normal range. To contrast this information, the green curve shows the theoretically maximum power output according to the vendor specifications. It can be seen that the normal data follows the theoretical curve closely, while the data out of range falls far away from the expected curve. Fig. 2. Power curves in the dataset: Empirical power curve in blue, vendor-provided theoretical maximum capacity power curves in green, red data points do not correspond to normal operation intervals. 3.1 Prediction of target variables The study first acknowledges the information shared between different variables to implement a next-step prediction for multiple variables from a subset of all available variables in the previous data-point. The rationale is that an accurate prediction of the next data point(s) would allow to run the anomaly detection mechanism on the predicted data and therefore identify the anomalies in advance. At the same time, it would confirm that there exist inherent relationships in the data that can be used not only for predictions, but also for detecting out-of-distribution samples. Figure 3 shows the performance of the HGBR model in predicting the target variable for the WTG power output, by plotting the predicted output horizontally and the actual data vertically. It can be seen that the scatter points fall very close to the identity line (in green), representing good quality in the predictions. 3.2 Predictive Maintenance - anomaly detection in advance To illustrate the performance of the normality detection mechanism, Figure 4 presents a time series plot distinguishing normal and abnormal periods. The variable afa, indicating normal operation, is shown in green, while actual recorded anomalies are marked with dashed red lines. The output of the normality detector – a continuous anomaly score – is smoothed using an exponentially weighted Fig. 3. Prediction and actual data for wind turbine power output. moving average (EWA) with a half-life of approximately half a week and is depicted by the orange curve. In this implementation, higher score values indicate a greater likelihood of a failure or anomalous behavior in the near future. The exact threshold to trigger the alarm for upcoming anomalous data can be user-defined according to its own risk aversion, as long as it stays between the minimum and maximum observed values, which in our experiments were 0.5 and 0.9 respectively. These values are induced by the smoothing. A higher threshold triggers less false positives but may trigger uncomfortably close to the actual failure. Finding an optimal value for the threshold could possibly be formulated as a cost optimization problem and is left for future research. The interpretation of said output is that lower values indicate periods of normal operation, while the higher values represent the proximity of anomalies. It can be seen that the normality predictor is able to anticipate the occurrence of failures many days in advance of the failure events. 4 Conclusion As the demand for renewable energy rises, predictive maintenance is poised to play a critical role in ensuring the reliability and efficiency of wind farm operations. This study demonstrates the effectiveness of leveraging SCADA sensor data and machine learning for early fault detection in wind turbines. The proposed methodology integrates data preprocessing, feature engineering, and advanced model training to predict key operational variables and identify anomalies well in advance. By enabling early intervention, this approach Fig. 4. Time series data for the wind turbine. Specific failure points are vertical dashed red lines. The green curve represents normal time periods with value 1 and abnormal periods by proximity to failures with a 0. The orange curve is the EWA filtered output score from the normality detector. supports improved turbine performance, reduced downtime, and more sustainable wind energy production. Future work will focus on deploying the methodology in real-world wind farm monitoring systems, expanding it to multi-turbine scenarios, and incorporating real-time feedback to further enhance accuracy and responsiveness. References 1. María Isabel Blanco. The economics of wind energy. Renewable and sustainable energy reviews, 13(6-7):1372–1382, 2009. 2. Shu Liu, Siwei Ren, and Hongliang Jiang. Predictive maintenance of wind turbines based on digital twin technology. Energy Reports, 9:1344–1352, 2023. 3. Yingying Zhao, Dongsheng Li, Ao Dong, Dahai Kang, Qin Lv, and Li Shang. Fault prediction and diagnosis of wind turbine generators using scada data. Energies, 10(8):1210, 2017. 4. Wisdom Udo and Yar Muhammad. Data-driven predictive maintenance of wind turbine based on scada data. IEEE Access, 9:162370–162388, 2021. 5. Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 6. Tianqi Chen, Tong He, Michael Benesty, Vadim Khotilovich, Yuan Tang, Hyunsu Cho, Kailong Chen, Rory Mitchell, Ignacio Cano, Tianyi Zhou, et al. Xgboost: extreme gradient boosting. R package version 0.4-2, 1(4):1–4, 2015.