Dynamic wavelet optimization and hierarchical reinforcement learning for fault diagnosis in wind turbines under variable operating conditions

This article presents a dynamic wavelet optimisation and hierarchical reinforcement learning (DWO-HRL) method to address the poor generalisation and high misjudgement rates in wind turbine fault diagnosis under variable operating conditions. First, a signal-characteristics-driven dynamic wavelet basis optimisation (SCD-DWBO) mechanism is proposed to extract both low-frequency contours and high-frequency transients through adaptive adjustment of wavelet scales. Then, an hybrid deep temporal architecture (HDTA) is designed to capture spatial and temporal features by integrating stacked convolutional layers with long short-term memory (LSTM) units. Finally, a task-driven hierarchical Q-learning (TD-HQL) strategy is designed to enhance diagnostic accuracy and robustness by synchronising policy gradients across dual layers. Experimental results show that the proposed method outperforms state-of-the-art methods in both performance and reliability.