Hybrid Fine-Tuning in Large Language Model Learning for Machinery Fault Diagnosis

Large Language Models (LLMs) have already made significant innovations and breakthroughs in the field of natural language processing (NLP). Considering the advances in addressing large-volume sequential data, LLMs can be implemented to deal with the problem of machinery fault diagnosis, which is still at a relatively early stage of exploration. Recently, LLMs have also been explored and applied in the field of time series analysis with excellent performance. Inspired by the use of LLMs in handling time series data, this paper investigates the effectiveness of LLMs in addressing fault diagnosis tasks. A novel hybrid fine-tuning for an LLM-based framework, called LLaMA-HFT, is proposed. The framework integrates the LLaMA model as a network backbone for feature extraction. Moreover, a hybrid fine-tuning (HFT) strategy is proposed for efficient fine-tuning of the model parameters. Experiments are conducted systematically and evaluated using a real-world fault diagnosis dataset. The experimental results show that the proposed method outperforms the other baseline methods in machinery fault diagnosis.