다중 에이전트 대규모 언어 모델을 활용한 데이터 효율적 베어링 고장 진단

키워드

확장요약문

Bearing fault diagnosis is essential for the reliable operation of rotating machinery. However, conventional data-driven methods require large amounts of labeled fault data, which are difficult to obtain in industrial settings, and provide limited insight into the evidence behind each prediction. Large language models (LLMs) can integrate heterogeneous information and generate natural-language reasoning, but converting vibration signals into text discards numerical and spatial information essential for diagnosis. This study proposes SMART-FDLLM, a retrieval-augmented multi-agent LLM framework for few-shot bearing fault diagnosis. The first agent extracts time- and frequency-domain statistical features from vibration signals. The second agent encodes their wavelet scalograms into visual embeddings using a pretrained DINOv2 model. The third agent retrieves similar reference cases directly in the signal domain, using both numerical and visual similarity instead of text embeddings. An LLM then compares the current sample with the retrieved cases and generates an initial diagnosis with an explanation and confidence level. The fourth agent verifies whether the predicted label is consistent with the extracted evidence and whether each claim in the explanation is supported by the signal information. Depending on the result, the initial response is retained, its explanation is rewritten, or the diagnosis is re-evaluated. The framework was evaluated on three public datasets covering fault severity (CWRU), real damage on different bearings (PU), and load variation with unknown-condition rejection (MFPT), as well as a 15-class in-house dataset including single and compound faults. Under a 4-shot setting with 5-fold cross-validation, the framework achieved accuracies of 97.7%, 99.5%, 97.7%, and 98.0%, respectively. Self-verification improved accuracy on all datasets by 0.5 to 3.4 percentage points by correcting misclassified samples. These results demonstrate that the proposed framework enables reliable and evidence-grounded bearing fault diagnosis with limited labeled data.