Large Language Model for Bearing Fault Diagnosis Based on Multi-view Structured Signal Tokens

Abstract

Bearing fault diagnosis is a key task in industrial big-data-driven intelligent operation and maintenance. However, the non-stationarity and distribution discrepancies of vibration signals under complex operating conditions limit the generalization capability of diagnostic models. To address the challenges of sample learning difficulty, insufficient cross-condition adaptability, and limited cross-component generalization, this paper proposes a LLM-based bearing fault diagnosis method using multi-view structured signal tokens (MVST-LLM). The proposed method takes continuous wavelet transform time-frequency images as the core representation, performs cross-modal fusion of multi-domain features, and constructs robust multi-view structured signal tokens by combining statistical feature modulation with a reliability-gated mechanism. Meanwhile, a knowledge-guided prompt is designed to incorporate prior knowledge such as bearing structural parameters and operating conditions. LoRA fine-tuning is further employed to align the signal tokens with the textual semantic space. Experimental results on three types of generalization tasks using the CWRU dataset demonstrate that the proposed method can effectively improve the generalization ability of large language models in complex diagnostic scenarios.