A label enhancement based positive-unlabeled hybrid network for pump bearing intelligent fault diagnosis

Abstract

Bearings are important support parts for rotating machinery such as pumps, and the application of machine learning algorithms has brought the fault diagnosis of bearings to a more intelligent stage. However, with the scarcity of target fault data and the lack of accurate labeling for critical data, commonly used data-driven fault diagnosis methods had its limitations. Inspired by semi-supervised learning, hypergraph and knowledge distillation theories, a hybrid PUHGNN network based on label augmentation was proposed in this paper. Firstly, a hypergraph neural network (HGNN) structure based on the multi-resolution signal was proposed to measure the correlation at multiple scales to make difference and connections between different labels. Secondly, the HGNN network is improved by combining HGNN and Positive-Unlabeled (PU) Learning ideas to form a new PUHGNN label enhancing mechanism which will solve the lacking of labels. Lastly, a soft-label-based label selection method is proposed to dynamically judge the similarity of samples to reiterate, which will make the otherwise indistinguishable faults more explicit. In experimental session, the CWRU dataset and the enviormental protection pump bearing datasets were applied to conduct unbalance, mislabel, extreme mislabel and ablation experiments. The result shows that the label enhancement is not only necessary but significant in the unbalanced under-labeled datasets, furthermore, the PUHGNN has more obvious enhancement compared to other methods.