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

Area of Science:

• * Engineering
• * Artificial Intelligence
• * Automotive Systems

Background:

• * Incipient fault diagnosis in traction control systems (TCS) faces challenges like weak fault signatures, low signal-to-noise ratios, and data variability.
• * Existing methods struggle with condition-induced diversity and cross-domain shifts, hindering accurate fault detection.

Purpose of the Study:

• * To propose an adaptive diversity-perception prompt framework with memory quantization for unsupervised domain adaptation in TCS fault diagnosis.
• * To enhance the accuracy and robustness of incipient fault detection in complex operational conditions.

Main Methods:

• * An adaptive deformation network (ADN) models and compensates for feature distortions to align representations.
• * An expert prompts learning strategy (EPL) achieves class-conditional domain alignment using unsupervised prototypes and learnable prompts.
• * A memory-based quantization for noise compression (MQNC) module discretizes features to suppress noise while retaining fault semantics.

Main Results:

• * The proposed framework demonstrated superior performance on online high-speed train datasets compared to conventional methods.
• * Validation confirmed the framework's effectiveness in handling weak fault signatures and domain shifts.
• * Ablation studies confirmed the robustness and effectiveness of individual components.

Conclusions:

• * The adaptive diversity-perception prompt framework offers a robust solution for incipient fault diagnosis in TCS under unsupervised domain adaptation.
• * The integration of feature alignment, expert prompting, and noise compression significantly improves diagnostic accuracy and reliability.