Wind Turbine Fault Detection Through Autoencoder-Based Neural Network and FMSA

Amid the global shift toward clean energy, wind power has emerged as a critical pillar of the modern energy matrix. To improve the reliability and maintainability of wind farms, this work proposes a novel hybrid fault detection approach that combines expert-driven diagnostic knowledge with data-driven modeling. The framework integrates autoencoder-based neural networks with Failure Mode and Symptoms Analysis, leveraging the strengths of both methodologies to enhance anomaly detection, feature selection, and fault localization. The methodology comprises five main stages: (i) the identification of failure modes and their observable symptoms using FMSA, (ii) the acquisition and preprocessing of SCADA monitoring data, (iii) the development of dedicated autoencoder models trained exclusively on healthy operational data, (iv) the implementation of an anomaly detection strategy based on the reconstruction error and a persistence-based rule to reduce false positives, and (v) evaluation using performance metrics. The approach adopts a fault-specific modeling strategy, in which each turbine and failure mode is associated with a customized autoencoder. The methodology was first validated using OpenFAST 3.5 simulated data with induced faults comprising normal conditions and a 1% mass imbalance fault on a blade, enabling the verification of its effectiveness under controlled conditions. Subsequently, the methodology was applied to a real-world SCADA data case study from wind turbines operated by EDP, employing historical operational data from turbines, including thermal measurements and operational variables such as wind speed and generated power. The proposed system achieved 99% classification accuracy on simulated data detect anomalies up to 60 days before reported failures in real operational conditions, successfully identifying degradations in components such as the transformer, gearbox, generator, and hydraulic group. The integration of FMSA improves feature selection and fault localization, enhancing both the interpretability and precision of the detection system. This hybrid approach demonstrates the potential to support predictive maintenance in complex industrial environments.