Semisupervised fault detection for wind turbines based on explainable artificial intelligence and transformer autoencoder

Abstract

Deep learning-based fault diagnosis for wind turbine blades has been widely adopted owing to its end-to-end feature extraction capability, yet it continues to face considerable challenges. First, wind turbines predominantly operate under normal conditions, resulting in an acute scarcity of labeled fault data. Second, the inherent black-box nature of neural network decision-making processes makes it difficult to earn the trust of field operation and maintenance personnel, thereby hindering industrial deployment. To address these issues, this paper proposes a novel semi-supervised fault detection method. Specifically, a synergistic integration of explainable artificial intelligence (XAI) and a Transformer-based autoencoder is constructed. A hybrid feature selection pipeline combining variance-threshold filtering and Shapley Additive Explanations (SHAP)-guided feature importance ranking is employed to eliminate redundant and noisy sensor channels, retaining only the features most valuable for fault diagnosis. An adaptive 3σ dynamic threshold strategy is then applied to analyze the reconstruction error distribution, minimizing false alarm rates while distinguishing between normal operational fluctuations and genuine fault signatures. The proposed method consistently outperforms the Long Short-Term Memory Autoencoder and Bidirectional LSTM Autoencoder baselines across all evaluation metrics, enabling accurate and effective fault detection.