Rolling bearing fault diagnosis based on multi-label zero-shot learning

Abstract

In recent years, the data-driven method develops rapidly in field of rolling bearing fault diagnosis, but facing fault types without historical records in engineering practice, there are still some deficiencies, such as, insufficient learning of fault characteristics and higher misdiagnosis rate. Here, aiming at the above problems, a multi-label zero-shot learning (MLZSL) fault diagnosis method was proposed. Firstly, visible and unseen samples were pre-processed with short-time Fourier transform (STFT), and the obtained time-frequency images were input into the residual depthwise separable convolutional neural network (RDSCNN) for feature extraction. Then, visible fault features were used to train an attribute learning network, the attribute vector of unseen fault samples was predicted with the attribute learning network. Finally, the diagnosis of unseen faults was realized. Fault diagnosis tests under the condition of zero sample were designed. The results showed that MLZSL can migrate attributes of visible faults to unseen ones, and effectively diagnose unseen faults.