Transnosis: Transformer-Variation-based Bearing Compound Fault Diagnosis with Zero-shot Learning and Domain Adaptation
Abstract
Bearing compound fault diagnosis is an important problem, and signal data from the target bearing are collected and inputted into deep learning model for diagnosis. One challenge of this problem is, to initially train the deep learning model, it is difficult or impossible to obtain data of compound faults, and the training data are usually single fault data. Another challenge is, the training data are usually obtained in working conditions different from the target bearing. To solve the bearing compound fault diagnosis problem addressing the two challenges, we introduce Transnosis, a Transformer-variation-based bearing compound fault diagnosis method with zero-shot learning and domain adaptation. Transnosis is a combination of three solid and time-tested Transformer variations: Cross-domain Transformer (CDTrans), Query2Label and Shifted Patch Tokenization/Locality Self-Attention (SPT/LSA) with dedicated designs. In the design of Transnosis, we propose weight sharing triple-attention decoders to incorporate domain adaptation in decoders of combined CDTrans and Query2Label, multidimensional embedding approach to make signal data enhancement motivated by the SPT technique, and also splitting-for-voting approach for further optimization. We test Transnosis with BJTU-RAO datasets on six tasks, and experimental results demonstrate Transnosis can accurately diagnose compound faults by effectively distinguishing features of different faults and clustering features of the same fault from different working conditions.