Adversarial generative transfer learning with self-selective attention transformer for multi-class fault diagnosis under variable-speed conditions

Accurate rotating machinery fault diagnosis under variable-speed conditions remains challenging because changes in operating speed produce nonstationary vibration signals and significant domain shifts. To address this issue, this paper proposes an adversarial generative transfer learning approach with self-selective attention transformer. First, raw vibration signals are represented in the time–frequency domain using short-time Fourier transform (STFT) and processed to obtain discriminative latent features. As variable-speed datasets contain complex and redundant feature dependencies, a transformer-based attention mechanism is incorporated into the feature extraction layers to enhance feature representation learning and enhance feature discrimination under speed variations. A major challenge in transfer learning-based fault diagnosis is the absence of fault samples at previously unseen target speeds. Moreover, Maximum Mean Discrepancy (MMD)-based alignment alone is often insufficient to effectively capture complex nonlinear domain discrepancies. In addition, existing deep generative architectures are primarily limited to fault detection and lack the capability for multi-class fault classification. To overcome these limitations, the proposed framework introduces multiple fault-specific generators to enhance this paradigm, enabling source-domain mappings from normal features to distinct fault distributions through joint adversarial and MMD-based optimization. The learned mappings are subsequently adapted to the target domain using only healthy target samples by utilizing transfer learning, enabling multi-class fault classification in unseen target operating speeds without requiring target-domain fault data. Experimental validation on a laboratory-scale wind turbine testbed with four Normal (healthy), Wearing, Mashed, and Bearing fault conditions demonstrates that the proposed approach consistently outperforms baseline methods in diagnostic accuracy under variable-speed operation.