Enhanced Fault Diagnosis Using Large Language Models and Probabilistic Label Fusion

Fault diagnosis is crucial for ensuring the reliability of manufacturing systems. However, a large proportion of unlabeled data remains unused, which presents a challenge that limits fault diagnosis performance. While some semi-supervised methods focus on leveraging unlabeled data, they rely solely on statistical properties and lack domain knowledge, leading to performance bottlenecks. In this paper, we propose a fault diagnosis framework that leverages Large Language Models (LLMs), which contain vast amounts of prior knowledge, to enhance Classification performance. We train a base classifier on labeled data, generate pseudo-labels for unlabeled data, and fine-tune LLaMA3 to infer another set of pseudo-labels. By merging these labels and adjusting sample weights, we create an augmented training set, which improves model performance. Our approach outperforms traditional methods as demonstrated on the Tennessee Eastman Process dataset.