Abstract. Condition monitoring and failure prediction in wind turbines have become an increasingly important research area due to their substantial economic impact. Accurate early detection of developing faults enables more efficient maintenance planning and minimizes costly downtime. However, predicting failures from operational wind farm data remains challenging. Real-world datasets are often affected by measurement noise, incomplete expert knowledge, and extraneous operating conditions, all of which complicate the identification and classification of emerging problems. This work presents a methodology designed to address one critical obstacle: measurement errors caused by faulty or unreliable sensors. Such errors can substantially degrade the performance of normal-behavior models (NBMs), thereby hindering the detection of anomalies and incipient failures. To mitigate this issue, we introduce an approach based on masked autoencoders (MAEs) that selectively suppresses signals deemed unreliable by domain experts or automated diagnostics. The proposed method is evaluated using four datasets from real operational wind farms. We analyze the impact of sensor-induced errors on NBM performance and demonstrate how the MAE framework improves robustness in the presence of corrupted measurements. Furthermore, it is shown that the methodology achieves a high failure prediction accuracy even in contexts with substantial numbers of sensor errors. The results highlight the potential of the method to improve the accuracy of data-driven failure prediction systems in practical wind turbine drive train applications.
Chesterman et al. (2026) studied this question.