Adjust to reality: LLM-driven test-time semantic adjustment for zero-shot fault diagnosis

Zero-shot fault diagnosis methods identify unseen faults by predicting semantic knowledge from samples. However, existing studies require labor-intensive tasks of training domain experts and annotating semantic knowledge in a specified format. Moreover, due to the domain shift problem (DSP), models trained solely on seen faults experience a decrease in performance when transferred to unseen faults. To reduce the annotation burden and address the DSP, we propose a test-time semantic adjustment method driven by the large language models (LLMs), which focuses on annotating and optimizing semantic knowledge. Firstly, with carefully designed prompts, semantic knowledge is automatically annotated by the LLM based on unstructured professional corpora. Secondly, to overcome the DSP caused by the lack of unseen faults, this study introduces the concept of test-time adjustment for zero-shot diagnosis. Specifically, we design a dual-view semantic knowledge adjustment strategy that employs information on unseen faults from unlabeled test data to adjust the semantic knowledge. This simple yet effective strategy can also be applied to other zero-shot diagnosis methods. Last but not least, we propose the class-agnostic feature extraction to enhance the cross-category transferability of extracted features for preventing overfitting to seen faults. We conduct experiments on the Tennessee-Eastman process (TEP) and a real thermal power plant (TPP), and the proposed method achieves an average improvement of 11.34% in terms of the accuracy of unseen faults on the TEP dataset, and 6.87% on the TPP dataset.