A two-stage source-free domain adaptation method for fault diagnosis of wind turbine gearboxes

Abstract In recent years, domain adaptation-based fault diagnosis methods have demonstrated remarkable effectiveness in cross-domain intelligent diagnosis of wind turbine gearboxes. However, their practical applicability is often limited by concerns over data privacy and high transmission costs. We propose a two-stage source-free domain adaptation technique for wind turbine gearbox problem diagnostics to overcome these constraints. In the first stage, transferable fault knowledge is extracted across multiple source domains using a combination of global adversarial training and local class semantic constraints, improving the generalizability of the model. The second stage involves selecting high-quality unlabeled samples from the target domain via K -means clustering and the silhouette coefficient to support model self-training. To further improve the model’s capacity for extracting fault-related features from time-domain vibration signals, a wavelet residual network is constructed. A publicly accessible wind turbine gearbox dataset and an extra dataset gathered from a test platform are used to assess the suggested approach. According to experimental results, the strategy achieves superior fault recognition performance across multiple diagnostic tasks.