Partly interpretable transformer through binary arborescent filter for intelligent bearing fault diagnosis

• Published in: Measurement : journal of the International Measurement Confederation 2022-11, Vol.203, p.111950, Article 111950
• Main Authors: Jiao, Zhiyuan, Pan, Liren, Fan, Wei, Xu, Zhenying, Chen, Chao
• Format: Article
• Language: English
• Subjects: Binary arborescent filter; Deep learning; Fault diagnosis; Interpretability; Transformer
• ISSN: 0263-2241; 1873-412X
• DOI: 10.1016/j.measurement.2022.111950

•BAFT possess high classification accuracy in bearing fault classification.•BAFT demonstrates a certain noise immunity.•BAFT enhances the interpretability and prevents a full black box of the model.•BAFT demonstrates the capability of the generalization ability. Deep learning (DL) has been widely studied in the field of bearing fault diagnosis and provides some advantages when applied to rich recorded data. However, DL models remain commonly uninterpretable and are merely black boxes, hampering their wide use in bearing fault diagnosis. To classify the bearing fault effectively and understand the learned representations which are hidden inside these models, the binary arborescent filter is embedded in the Transformer in this paper. With the help of the binary arborescent filter, a novel tokenizer is constructed instead of the original one which is only used for the natural language process. We show how the feature constructed by the tokenizer can be interpreted as classifiers that determine different fault types. Therefore, based on the Binary arborescent filter Transformer, a new end-to-end fault diagnostic framework is developed to boost the diagnostic performance of the conventional DL-based bearing fault diagnosis (BFD) models. Experimental studies showed the anti-noise validity and superior performance of the proposed BFD model.