FR-LLM: Multi-task large language model with signal-to-text encoding and adaptive optimization for joint fault diagnosis and RUL prediction

Abstract

Fault diagnosis and remaining useful life prediction are crucial for ensuring the reliability and safety of rotating machinery, yet most existing methods address them separately, limiting adaptability. We propose FR-LLM, a unified multi-task large language model that jointly performs both tasks within a single framework. Raw vibration signals are transformed into structured textual prompts through signal-to-text encoding, where frequency-domain features support fault diagnosis and multi-domain statistical features with empirical mode decomposition capture degradation for life prediction. An adaptive Convergence Balancer dynamically adjusts task-specific loss weights to mitigate conflicts in multi-task optimization, while a low-rank adaptation strategy reduces computational demands. Experiments on the XJTU-SY and IMS bearing datasets show that FR-LLM consistently outperforms single-task approaches and existing language model baselines in accuracy, generalization, and efficiency. Ablation studies further highlight the contributions of the Convergence Balancer and low-rank adaptation to robustness and stability. These results demonstrate that FR-LLM offers a practical and interpretable solution for predictive maintenance, advancing the application of large language models in industrial prognostics.