CGASNet: A Generalized Zero-Shot Learning Compound Fault Diagnosis Approach for Bearings

In deep learning-based compound fault diagnosis of bearings, collecting and labeling enough compound fault samples is unrealistic. However, single fault samples are usually available. Therefore it is a challenging and practical application to employ single fault samples to train the diagnostic model and then identify single fault and compound fault simultaneously. For this purpose, we present a generalized zero-shot learning compound fault diagnosis (GZSLCFD) approach, termed contrast generation and adaptive smoothing network (CGASNet). First, we present a fresh fault semantic constructing approach based on the statistical indicator features of original vibration data aligned with the extracted features. Secondly, a feature extractor based on a deep residual contraction network is devised for extracting fault features from wavelet images of vibration signals. Then, we train a contrast embedding generation module using the semantics and the extracted features. Finally, we apply the adaptive smoothing approach to design three sub-networks in the fault identification module, which together accomplish the recognition of seen single faults and unseen compound faults samples. Extensive comparative experiments are conducted on a self-built bench to validate the superiority of the proposal approach. The experimental analysis revealed that with no compound fault samples, the generalized zero-shot learning compound fault diagnosis achieved an accuracy of 83.15%.