A zero-shot fault semantics learning model for compound fault diagnosis

Abstract

Compound fault diagnosis of bearings has always been a challenge, due to the occurrence of various faults with randomness and complexity. Existing deep learning-based methods require numerous compound fault samples for training. However, sufficient training samples for every type of compound fault are usually unavailable in realistic industrial scenarios. In this paper, we propose a zero-shot fault semantics learning model trained on single fault samples to identify unknown compound faults. First, we propose a convolutional autoencoder-based fault semantics construction method to generate the fault semantics for single and compound faults. Second, a CNN-based feature extractor is designed to extract fault features from time–frequency domain vibration signals. Then we design an autoencoder-based fault semantics embedding module to embed the compound fault semantic vectors into the fault feature space as the category centroids. Finally, by the similarity measurement between the compound fault features and the category centroids, the model is able to identify unknown compound faults. Extensive comparative results demonstrate that our method outperforms a series of state-of-the-art models, with the accuracy of 78.40% for three categories of compound faults in the case of 2000 single-fault training samples per category.