PC-LLM: Physics-Conditioned Large Language Models for Interpretable Bearing Fault Diagnosis in Industry 4.0

Abstract

Vibration-based bearing fault diagnosis is fundamental to condition-based maintenance of rotating machinery, yet generalizable automated diagnosis under varying operating conditions remains an open challenge. Large Language Models (LLMs) applied to this task achieve high accuracy by encoding vibration features as text prompts under matched conditions, yet degrade substantially under cross-condition transfer owing to absent physics-informed spectral context and fault-condition entanglement in the fine-tuned parameters. To address these limitations, this paper presents PC-LLM, a three-component framework comprising a Physics-Conditioned Hidden-state Memory (PCHM) branch, a Condition-Disentangled Adaptation (CDA) mechanism, and a Severity-Aware Structured Reporting (SSR) layer. The PCHM branch encodes orderdomain carrier and analytic envelope spectra into learnable prefix tokens prepended to the LLM hidden-state pathway, while the CDA mechanism decouples fault-discriminative from condition-specific representations via adversarial gradient reversal and Frobenius-norm decorrelation, enabling zero-shot cross-load generalization. The SSR layer fuses language-model and classifier scores, classifies fault severity into four grades, and produces diagnostic reports with spectral evidence and actionable maintenance recommendations. Evaluated across four open-source LLM backbones on the Case Western Reserve University (CWRU) bearing dataset under within-distribution, zero-shot cross-load, and cross-component settings, PC-LLM consistently achieves near-ceiling fault classification accuracy, substantially improves cross-load generalization, and provides interpretable diagnostic evidence not attainable from singlelabel baselines.