A generative zero-shot learning method for compound fault intelligent diagnosis of rolling bearings

Abstract In practical industrial applications, the scarcity of compound fault samples poses significant challenges in obtaining sufficient training data, leading to class imbalance and negatively impacting the diagnostic performance of models. To tackle this issue, we propose a generative zero-shot learning method for rolling bearing compound fault intelligent diagnosis, attempting to diagnose unknown target faults using only known fault samples. Specifically, a novel fault attribute description method is designed, which combines fault semantic information with manually defined fault description information, thereby constructing fault category auxiliary information (FCAI) to learn the correlation between single faults and compound faults. Furthermore, the continuous wavelet transform is used to data preprocess the raw vibration signals, and a wide kernel convolutional neural network is constructed to extract deep fault feature information from the samples. Finally, an adversarial training strategy is adopted to learn the mapping relationship between the fault feature information and the FCAI, and compound faults are diagnosed using a distance metric method based on the similarity relationship between fault features. Through experimental validation on the laboratory bearing dataset and the Huazhong University of Science and Technology bearing dataset, the effectiveness and superiority of the proposed method in scenarios lacking compound fault samples.