Contrastive learning-driven event extraction for hoist fault diagnosis based on XLNet

Abstract With the growing demand for intelligent industrial equipment, the accuracy and efficiency of hoist fault diagnosis have become crucial for ensuring production safety. However, the scarcity of fault data and the high annotation cost in real industrial scenarios pose significant challenges for traditional methods, making it difficult to diagnose faults effectively under few-shot or zero-shot conditions. To address this issue, this study proposes a fault diagnosis event extraction method based on the XLNet pretraining model combined with contrastive learning. The proposed method involves three key steps: First, contrastive learning samples are constructed, and event descriptions are rewritten using labelled and unlabeled data. Second, diverse fault samples are generated through data augmentation techniques to expand the dataset. Finally, the loss function is optimized to enhance the training process of the XLNet model. Experimental results demonstrate that in the hoist fault diagnosis task, the F1 score of the proposed method improves by 2.24% after data augmentation. Moreover, compared to other pre-trained models, the XLNet model achieves a maximum enhancement of 8.45% in the F1 score, while the improved loss function further boosts the F1 score by 3.64%. Additionally, the results on the public dataset indicate that the proposed method exhibits strong generalization capabilities.