An Innovative Foundation Model for Bearing Prognostics and Health Management Through Pre-Trained Large Language Models

Abstract

In recent years, deep learning has significantly advanced fault diagnosis and remaining useful life (RUL) prediction in prognostics and health management (PHM) through the end-to-end feature extraction capabilities of deep neural networks. However, most existing PHM models are task-specific and scale-limited, lacking the generalization ability of foundation models. This leads to overfitting on small datasets, performance saturation on more complex datasets, and difficulties when handling multiple tasks simultaneously. To address these issues, we have developed an innovative PHM foundation model, Pulse, which is achieved through fine-tuning pre-trained large language models (LLMs). By leveraging the modal adaptability of LLMs, Pulse integrates signal and prompt data modalities to effectively tackle challenges of data scarcity and performance saturation, while enabling knowledge transfer across multiple tasks.  Validated on four bearing datasets, Pulse demonstrates robust generalization and scalability in fault diagnosis and RUL prediction.

Keywords: Rotating machinery, Prognostics and health management (PHM), Fault diagnosis, Remaining useful life prediction, Large language model (LLM), Foundation model