FD-Zero framework: LLM-based diagnosis framework for zero-shot mechanical time-domain signals

Abstract

This study addresses the issues of insufficient interpretability in diagnostic results and excessive reliance on labeled data in deep learning for mechanical fault diagnosis. The authors propose a framework for bearing fault diagnosis based on zero-shot learning, termed FD-Zero. First, to tackle the challenge posed by ultralong context in time-domain signals, this framework designs a Transformer-based vibration signal encoder. The raw time-domain vibration signals are compressed using multiscale residual convolution blocks (MRCBs) to extract features and encode them into high-dimensional representations. Subsequently, an improved contrastive learning method spatio temporal contrastive learning (ST-CL) is employed to enhance the temporal information during signal encoding, aligning text vectors with signal vectors within a semantic space. FD-Zero integrates domain knowledge transfer with cross-modal alignment techniques to achieve high-accuracy bearing fault diagnosis under zero-shot scenarios while providing interpretable decision-making support through text-signal semantic associations. Experimental results demonstrate the superior performance of our proposed framework in fault diagnosis. The superior performance of this research framework in fault diagnosis has been proved through experiments. The zero-shot fault diagnosis capability of FD-Zero has been verified through zero-shot fault classification experiments. The ablation experiment proved the superiority of MRCB and ST-CL.