Adaptive diversity-perception prompt framework for fault diagnosis with memory quantization

Incipient fault diagnosis in traction control systems (TCS) is challenged by weak fault signatures with low signal-to-noise ratios, condition-induced intra-class diversity, and cross-domain distribution shifts. To address these issues, an adaptive diversity-perception prompt framework for fault diagnosis with memory quantization is proposed, which operates under an unsupervised domain adaptation setting. Specifically, an adaptive deformation network (ADN) explicitly models and compensates for condition-induced feature distortions to achieve diversity-perception representation alignment. Based on the calibrated features, an expert prompts learning strategy (EPL) enables class-conditional domain alignment via unsupervised expert prototypes and learnable prompts without target-domain labels. Furthermore, a memory-based quantization for noise compression (MQNC) module suppresses stochastic noise by discretizing features into stable prototypes while preserving critical fault semantics. The framework was validated with online high-speed train datasets, demonstrating superior performance compared to conventional methods. Ablation studies further validate the effectiveness and robustness of the proposed framework.