CM of wind turbines has emerged as a rapidly expanding research field, encompassing a wide range of methodological approaches. Based on the primary source of information, CM techniques can be categorized into three main groups: (i) SCADA-based approaches, which rely on readily available low-frequency data from the SCADA system of the wind turbine; (ii) vibration-based approaches, which utilize high-frequency signals collected from dedicated accelerometers; and (iii) acoustic emission-based approaches, which analyze high-frequency acoustic signals generated by turbine components . Each category exhibits distinct advantages and limitations and addresses specific needs within the CM domain.

This paper focuses on SCADA-based CM methods that exploit low-frequency operational data. The most commonly used temporal resolution for SCADA data is 10 min, although higher sampling frequencies have also been reported in the literature . Within the SCADA-based paradigm, distinguish between physical-model-based approaches, which rely on explicit representations of turbine dynamics, and data-driven approaches, in which models are learned directly from historical SCADA data. This paper adopts the data-driven approach.

Data-driven CM encompasses a broad spectrum of methodologies. According to , these methods can be divided into traditional data-driven techniques (e.g., trending analysis, statistical methods, and state–curve modeling) and machine-learning- and deep-learning-based approaches, including supervised, semi-supervised, and unsupervised methods. An alternative taxonomy is proposed by , who identify NBM, trending, clustering, damage modeling, alarm assessment, and expert systems as distinct categories. Comprehensive reviews of recent advances in wind turbine CM are provided in , , , and .

This paper adopts the NBM approach, which has demonstrated strong performance in wind turbine health monitoring. A key advantage of NBM methods is that they learn a representation of normal system behavior that can be used for prediction and anomaly detection. NBM techniques can themselves be subdivided into several groups, with statistical methods, shallow machine learning models, and deep learning architectures being the most prominent (see , for an extensive overview).

The statistical group includes conventional modeling and dimensionality-reduction techniques, such as ordinary least squares regression (OLS) , principal component analysis (PCA) , and cointegration analysis .

Shallow machine learning models comprise widely used algorithms that are more flexible than statistical methods, particularly for capturing nonlinear relationships. Examples applied in wind turbine CM include random forests (RFs) , gradient boosting machines (GBMs) , and support vector machines (SVMs) . While these models can achieve higher predictive accuracy, they typically require larger training datasets and offer reduced interpretability compared to statistical approaches.

Deep learning models represent the most recent and rapidly expanding class of NBM techniques. These methods rely on neural network architectures with multiple layers and have been applied extensively in modern CM systems. Common examples include deep neural networks , convolutional neural networks (CNNs) , long short-term memory networks (LSTMs) , and autoencoders (AEs) .

This work addresses failure prediction in the presence of sensor errors, which are conditions under which corrupted sensor signals can significantly degrade predictive performance. Following the ISO definition, a measurement error is the difference between the measured value and the true value of the measurand and can generally be decomposed into random and systematic components . We define a sensor error as any condition causing a sensor to produce such measurement errors. In a similar vein, describe a sensor error or fault as an unexpected deviation in the observed signal in the absence of any underlying anomalous system behavior.

A wide variety of sensor fault modes has been reported in the literature, and several taxonomies have been proposed. identify the following major categories:

Bias is a constant offset from the nominal signal, often caused by miscalibration or physical changes in the sensing element (e.g., a temperature sensor consistently overestimating by 10 °C).

Masked machine learning models form a prominent class of self-supervised learners. Their initial development was driven by natural language processing (NLP), with Bidirectional Encoder Representations from Transformers (BERT) being a well-known example. BERT employs a masked language modeling (MLM) objective, in which parts of the input sequence are masked and subsequently predicted during training .

MAEs have since been adopted in computer vision. For instance, introduce an MAE method for image data based on an asymmetric encoder–decoder architecture. This design differs from ours, which employs a strictly symmetric architecture. Their approach also uses a substantially higher masking ratio (75 %), whereas the ratio in our study is considerably lower. Additionally, the data modality (images versus the tabular data used here) introduces fundamental differences in model structure and learning dynamics. Several other works have extended MAE to computer vision tasks .

In contrast to vision-focused research, our work applies MAE to tabular sensor data, which requires important algorithmic modifications. Several recent papers have investigated this direction. use MAE for missing-value imputation in tabular datasets. Similarly, develop a proportional masking strategy designed to compensate for biases that arise with naive random masking and demonstrate that multi-layer perceptron (MLP)-mixed models often outperform transformer-based alternatives for tabular data. A general masked encoding methodology for tabular data is presented in , with further specialization for time-series applications explored by , who use masking to learn stronger temporal representations.

To our knowledge, the methodology most closely related to the one proposed in this paper is that of , who apply an MAE to streamline the fault detection, isolation, and accommodation (FDIA) pipeline for offshore wind turbines. Despite the conceptual similarity, several important differences remain. First, their AE topology does not include a bottleneck layer, whereas our method explicitly enforces one, and they do not incorporate a feature expansion layer. Their training noise level is also substantially lower (maximum 0.2) compared with ours (0.5). In terms of input structure, they use only masked SCADA data, while we provide both masked SCADA data and an explicit masking matrix. Their validation focuses on blade bending moments with simulated sensor faults, whereas our study evaluates generator and gearbox temperature sensors that exhibit real sensor failures. Furthermore, we additionally assess the ability of the model to detect wind turbine component failures, while concentrate solely on sensor fault detection. As a result, the methodology developed in this paper is tailored to component failure prediction under realistic sensor fault conditions, aligning more closely with operational needs.

This study uses 10 min SCADA data from four operational offshore wind farms (WF1–WF4). The dataset spans three wind turbine types with rated capacities between 3 MW (megawatt) and 10 MW. Across the four wind farms, 136 turbines are included, representing an aggregated observation period of approximately 1000 turbine years. The SCADA data provide measurements of component temperatures and environmental and operational conditions. Only a subset of these signals is used in this work. Table  provides an aggregated overview, with detailed signal lists presented in Tables , , , and . The primary focus of this research is on detecting abnormal behavior in temperature-related signals under the condition that sensor errors can occur. These anomalies may serve as precursors to component failures.

Because anomaly detection and failure prediction for wind turbines must rely on operational SCADA data, which frequently contain gaps, sensor issues, and other imperfections, several preprocessing steps are required before model development. The following steps summarize the procedure adopted in this study:

Data aggregation. The original 10 min SCADA data are aggregated to a 1 h resolution. This reduction in temporal resolution is acceptable because the degradation mechanisms of interest evolve over long timescales (days to months). Aggregation also reduces high-frequency noise, resulting in cleaner input for model training.

A wide range of techniques for anomaly detection and failure prediction have been proposed in the literature, including trending approaches, clustering methods, NBMs, physics-based damage models, alarm-based systems, and expert-rule systems . In this work, we adopt the NBM approach, which is widely used and has demonstrated strong performance in industrial monitoring applications.

Normal-behavior modeling consists of two steps. First, a model is trained exclusively on data representing healthy system behavior. When successful, the model learns the normal relationships among the variables. In the second step, the model is applied to new data to generate predictions that represent the expected (healthy) behavior. Deviations between the predictions and the measured values are interpreted as indicators of abnormal or degraded operation: small deviations correspond to nominal behavior, whereas large deviations suggest emerging faults.

In this study, an MAE is employed as the NBM. This architecture is specifically chosen for its ability to handle measurement errors and missing values caused by sensor failures. By incorporating an explicit mask, the AE can learn normal inter-signal relationships even when some sensor values are unavailable or unreliable.

The input data contain numerous sensor failures that manifest as measurement errors. For validation, 12 cases were selected from WF1, WF2, and WF3. The selection is based on their suitability for the analysis. As explained in the previous section, the IFDW – of which the masked window is a part – must be sufficiently long and contain sufficient healthy data. A sufficient number of samples need to be available for the KS test. This implies that both the sensor error window and the healthy or normal data window surrounding the sensor error window must be large enough. This was the case for 12 cases.

The cases include several of the sensor failure types discussed in Sect. 2. Some sensor errors are readily identifiable because they produce physically impossible values, as illustrated in Figs.  and . These typically correspond to biased sensors or hard faults. Other errors are more subtle, arising from gradual sensor drift. This drift appears as a slow upward or downward trend and may only produce implausible values after an extended period, if at all (see Fig. ). These cases are more difficult to detect, especially because temperature-related signals may also deviate due to legitimate causes, such as component wear or damage.

To make the identification of the sensor errors easier, the MAE was first applied in a non-masking mode, meaning that no observations or signals are switched off. In this configuration, sensor failures manifest as characteristic anomalies in the reconstruction errors. In some instances, the failures appear as isolated outliers, but more commonly they result in prolonged periods during which the reconstruction error deviates consistently from the expected pattern, reflecting biased readings or hard faults, as visible in Figs. , , and . These extended error periods arise because sensor issues are often not noticed immediately, and even once identified, repairs or replacements typically occur only during scheduled maintenance. In some cases, the sensor errors persist for several years before the issue is resolved (see Fig. ). These patterns were identified (in collaboration with industrial partners) and logged (i.e., turbine, signal, start, and end). The log was then used to switch off signals during the run in masking mode.

The sensor errors occur across a broad range of sensors, including hydraulic oil temperature sensors, gearbox oil sump temperature sensors, gearbox A2B planet non-rotor-end (NRE) carrier bearing temperature sensors, gearbox A2B planet rotor-end (RE) carrier bearing temperature sensors, generator phase 3B temperature sensors, generator air outlet 1 temperature sensors, gearbox A1 planet NRE bearing temperature sensors, gearbox A2 planet RE bearing temperature sensors, gearbox B planet NRE carrier bearing temperature sensors, the nacelle temperature sensor, and generator slip-ring temperature sensors.

During the observation period, the wind turbines under study experienced several types of failures. Based on feedback from industrial partners and their assessment of the economic relevance of each failure type, the methodology is validated on generator and gearbox failures. These failures occurred relatively frequently within the wind farm and were associated with long downtimes; outages exceeding 1 month were not uncommon. In all cases, the failures resulted in complete replacement of the affected component, which is a non-trivial operation given the size and complexity of these components.

The industrial partners provided a list comprising 22 failures deemed relevant for this study, including 9 gearbox failures and 13 generator failures. Failures were identified by the industrial partners at the time of occurrence, and their nature was subsequently determined through inspections of the failed components. The gearbox failures were primarily caused by bearing damage, representing relatively conventional cases for which SCADA-based condition monitoring techniques are well established. In contrast, the generator failures involved electrical short circuits. These cases are more challenging to detect, as the failures are electrical in nature and can only be observed indirectly through the warming of certain generator components like the phases. At the time of writing, the exact root causes of the generator failures remain unknown.

The first question examined in this study is whether the reconstruction errors produced by the MAE when at least one sensor has failed remain comparable to those obtained when none of the sensors have failed and the turbine is in a healthy state. Establishing this equivalence is critical: if reconstruction errors differ substantially between these two regimes, it would indicate that the AE has not adequately learned the underlying inter-sensor relationships in the presence of missing or corrupted inputs. In such a case, the reconstruction errors would become unreliable for anomaly detection precisely when sensor faults occur, thereby undermining the purpose of the method.

Figures  and provide qualitative insight into the behavior of the algorithm. The bottom plot in each figure shows the normalized signal that is generated by the faulty sensor. The window during which the sensor has failed is highlighted with a green box. The top plot shows the reconstruction errors for a signal different from the one suffering from the sensor failure. The model generates the orange reconstruction error when the data generated by the faulty signal are not masked and the blue reconstruction error when they are masked.

In Fig. , the failure occurs in the gearbox oil sump temperature sensor (bottom panel). When the failure is not masked, the corrupted measurement propagates through the learned correlations and induces distortions in the reconstruction errors of other signals, for example, the temperature measured at the gearbox rotor end (RE) carrier bearing of planetary stage A1 (orange line in top plot). These distortions render the anomaly detections for these signals unreliable. Specifically, the failed sensor exhibits a sudden upward bias, which in turn produces an inverted (downward) bias in the reconstructed estimates of correlated signals (orange line in top plot). Because the AE models the normal relationships among these variables, an anomalously high value in one variable prompts overestimation of the corresponding “normal” values of others, leading to negatively biased reconstruction errors. When the gearbox oil sump temperature signal is masked when suffering from the sensor error (green box) the reconstruction error is not influenced by the reconstruction bias, and the error remains in line with the errors outside the green box (blue line in top plot). This indicates that the MAE and the masking of the faulty signal work well.

An analogous but reversed pattern is observed in Fig. . Here, the affected sensor, the gearbox RE carrier bearing of planetary stage A2B temperature, shows a pronounced downward bias during the sensor failure window (green box), resulting in upward-biased reconstruction errors in several correlated signals, exemplified in the figure by the reconstruction error for the gearbox RE carrier bearing temperature (orange line in top plot). When the faulty sensor is masked, this upward bias is eliminated, and the reconstruction errors align with the surrounding baseline values (blue line in top plot).

Due to the large number of sensor error validation cases (12 in total), exhaustive visual inspection is impractical. We therefore adopt an automated and formal evaluation based on multiple statistical metrics (see the Methodology section). Figures – report the average p values (horizontal bars) obtained from KS tests for different subsystems, i.e., the generator, gearbox, cooling system, and hydraulics, and for several statistical descriptors of the reconstruction error distributions, i.e., the mean, median, SD, IQR, skewness, and kurtosis. The purpose of these figures is to assess the similarity of reconstruction error distributions, for signals not suffering from sensor errors, between masked windows containing one or more sensor errors and unmasked windows free of such errors. If the distributions are statistically indistinguishable, the corresponding KS test p values should exceed the 2.5 % significance threshold (indicated by the horizontal reference bars). Larger p values indicate weaker evidence against the null hypothesis of identical distributions. By considering multiple metrics that capture different distributional properties, a more comprehensive assessment of similarity is obtained. For clarity and conciseness, only the average p value per component is shown in the figures.

For validation, 12 sensor error cases were identified. The selection procedure prioritized cases that provide high test quality, requiring that (i) the sensor error period be sufficiently long to extract multiple samples and (ii) enough healthy operational data be available in the surrounding time windows to enable a balanced comparison. Details of this procedure are provided in the Methodology section. The results indicate that, for most components and metrics, the average p values clearly exceed the 2.5 % threshold, implying that the null hypothesis (no difference between reconstruction error distributions with and without sensor errors) cannot be rejected. The consistency of these findings across metrics suggests that there is insufficient statistical evidence to conclude that reconstruction errors differ between data affected by sensor errors and unaffected data. Since the applied metrics characterize complementary aspects of the distributions, these results indicate that the MAE effectively compensates for sensor errors.

In some instances, p values approach or slightly fall below the 2.5 % threshold. There are several possible explanations for this. Firstly, when comparing the p values to the Bonferroni-corrected significance levels, which take into account the multiple hypothesis testing, then none of the p values are smaller than the threshold (0.42 %). However, as stated in the Methodology section, this procedure is very conservative. Secondly, as the data originate from operational wind farms rather than controlled laboratory environments, where unobserved operational events may influence measurements, some fluctuations are expected. A third, farm-specific explanation applies to wind farm 4, which exhibits particularly complex sensor error conditions. In this farm, all gearbox-related signals are missing for an extended period exceeding 2 years, likely due to a persistent communication failure, although the root cause cannot be conclusively determined. As a result, only generator-related results are shown in the right subfigure of Fig. . For WF4-T1, the p value for the mean falls below the significance threshold, and the p value for the median lies close to it. Overall, the average p values for wind farm 4 are lower than those observed for the other farms. This reduction reflects the increased difficulty of the scenario, in which approximately half of the input signals are missing. Despite these challenging conditions, the MAE continues to produce reconstruction errors for the generator signals that are broadly representative, albeit with reduced robustness.

In this section, we evaluate the modeling performance of the MAE. The assessment is based on the absolute reconstruction errors computed for data classified as healthy. Unhealthy data are identified using the procedure described in the Methodology section. For each signal, both the mean absolute reconstruction error and the mean relative reconstruction error are calculated. The absolute reconstruction error is defined as the absolute difference between the reconstructed and the original signal. The relative reconstruction error is defined as the ratio between the absolute reconstruction error and the corresponding input value. These metrics are subsequently averaged across components, namely the gearbox and the generator. The results are presented in Fig. . Overall, the mean absolute reconstruction error is below 1 °C for most cases, with the exception of the generator results for WF4, where it is approximately 1.5 °C. Similarly, the mean relative reconstruction error is generally below 1.5 %, except for the WF4 generator results, where it ranges from approximately 2.5 % to 3.0 %. These findings indicate that the model is able to capture normal operating behavior with good accuracy.

The comparatively lower performance observed for WF4 may have several explanations. First, this wind farm provides the smallest number of available signals for modeling. For instance, only one signal is available per generator phase, whereas two signals are available for the other wind farms. These signals are known to contain complementary information. In addition, many turbines in WF4 are known to experience generator bearing slip. The cause and nature of the bearing slip are currently unknown. The phenomenon manifests as frequent, sudden spikes in bearing temperatures, which are difficult for the model to predict and therefore increase the reconstruction error. Because these events are widespread, they are not fully excluded by the unhealthy data filtering procedure and thus remain present in the dataset considered healthy. This may explain the higher absolute and relative reconstruction errors observed for the WF4 generators.

For failure prediction, it is not sufficient that the model effectively masks sensor faults; it must also reliably distinguish between healthy and degraded operating conditions. Real-world data are not generated under lab conditions. This implies that the data, in this case the temperatures, are influenced by many factors, among which there are factors that are unrelated to the failure cases we use for validation. This might cause temperatures to rise suspiciously and result in anomalies that cannot be correlated with the validation failure cases. However, it would be incorrect to label these as false positives since the temperature increases are real. The validation is further complicated by the fact that these unrelated influential factors are often not all known to us and the industrial partners. This can be due to a lack of information on the occurrence of the event or on the relation between the factor and the temperature signals.

The validation of the methodology is based on generator and gearbox failures. These failures were identified by the industrial partners and are based on actual inspection of the component after the failure. The validation methodology, which is based on the ABSM metric, is tailored in such a way that it measures the actual performance of the algorithm in identifying the validation failure cases as well as possible, with as little influence as possible from the other unknown influential factors. For the analysis of the figures, we focus on the evolution of the anomaly concentration metric just prior to the failure and less on the evolutions far removed in time since those can be caused by the other unknown influential factors.

Figures  and illustrate examples of correct detections. The top plot in each figure shows (i) the mean of the reconstruction errors generated by the 200 bootstrapped models (blue line) and (ii) the 95 % prediction intervals (PIs) (yellow zone) for a specific signal, namely generator phase 2 temperature in Fig.  and gearbox RE main bearing temperature in Fig. , respectively. The PIs are generated by calculating the 2.5 % and 97.5 % percentiles of the reconstruction errors generated by the 200 bootstrapped models. These percentiles bound the 95 % PIs. The bottom plot shows the anomaly concentration for the respective signals. An anomaly is identified when the 2.5 % percentile of the reconstruction errors is larger than 0. By using this rule for flagging anomalies, both the trend and the uncertainty are taken into account. The anomaly concentration is the ratio of the number of anomalies versus the number of data points in a 1 d window. This results in a time series with values between 0 and 1.

In Fig. , a gradual degradation trend is visible in the anomaly concentration for the generator phase 2 temperature. The trend starts approximately 18 months prior to the failure, which occurred at year 6.25. In Fig. , temperatures in the gearbox RE main bearing begin to rise around year 5.5, followed by failure at year 5.75. The detection lead time varies substantially across failures, influenced by many factors that are not always observable. In both of these cases, the model identifies the degradation well in advance. The figure also includes several long-duration downtimes unrelated to the gearbox; these are not detected when using gearbox signals, as the associated failure modes are not gearbox related. In Fig.  the anomaly concentration metric also rises around Y2. It is unclear what causes this. There is no evidence that it is caused by seasonal fluctuations since this would also be visible in the following years. An analysis of the event is ongoing.

To obtain a general assessment of the value of the model as NBM, we validate it on historical failure data from operational offshore wind farms. For three of the four wind farms described in the Methodology section, industrial partners provided lists containing the turbine ID, timestamp, and failure type. Validation is performed using the ABSM metric (see Methodology section). A strong NBM should yield ABSM values substantially greater than 1 for the component that ultimately fails. In total, 22 drive-train-related failures were evaluated: 9 gearbox failures and 13 generator failures. These failures were provided by the industrial partner. They constructed the list based on post-failure inspections of the components. The results for wind farm 1 and wind farms 2 and 3 are presented in Tables  and , respectively.

Detection performance is categorized as follows: ABSM >2 indicates a strong detection, 1.25<ABSM≤2 a marginal detection, and ABSM≤1.25 a missed detection. Across all 22 failures, 15 (68.18 %) are strong detections, 5 (22.72 %) are marginal detections, and 2 (9 %) are missed detections, yielding an overall detection rate of 20 out of 22 cases (90.91 %). When split by failure type, the detection rate is 11 of 13 (84.62 %) for generator failures and 9 of 9 (100 %) for gearbox failures.

Overall, the model demonstrates strong discriminative ability. It reliably separates healthy from unhealthy behavior in most cases, with performance somewhat higher for gearbox failures than for generator failures. This difference is expected: several generator failures, particularly in wind farms 1 and 3, stem from electrical issues that are only indirectly reflected in generator-phase temperatures, making early detection more difficult. The two missed detections correspond to generator failures in wind farm 1.