GMPF-LLM: gateless multi-physical fusion large language model for fault diagnosis

Abstract

The operational reliability of rotating machinery is critical for modern industrial systems. However, existing large language model (LLM)-based fault diagnosis approaches face challenges in processing high-frequency, strong-noise vibration signals, including modal misalignment, cross-modal negative fusion, and limited learning capacity for early-stage weak faults. To address these issues, this paper proposes a gate-free multi-physical field fusion LLM (GMPF-LLM) for fault diagnosis. The method constructs a parallel multi-physical feature extraction and alignment mechanism to effectively decouple steady-state harmonics, background residuals, and transient impulses. A continuous multimodal projector with a gate-free residual fusion structure is employed to mitigate information distortion and noise interference during cross-modal fusion. Furthermore, a difficulty-aware joint optimization strategy using low-rank adaptation (LoRA) and multi-class focal loss enhances recognition of early weak faults under low signal-to-noise ratio conditions. Experiments on the XJTU-SY and GDUPT bearing datasets demonstrate mean diagnostic accuracies of 99.25% and 90.55% with minimal standard deviation, outperforming mainstream methods. Ablation studies further confirm the contributions of each core module. This work provides an effective framework for integrating multi-physical field data with LLMs in industrial fault diagnosis.