An Interpretable Fault Diagnosis Framework based on Capsule Network with Statistical Features

Intelligent fault diagnosis algorithms can provide strong support for the predictive maintenance of equipment and avoid economic losses caused by equipment failure. However, the application of intelligent fault diagnosis methods is limited in practice industrial scenarios due to the lack of interpretability. To address the above problem, an interpretable intelligent fault diagnosis framework based on the capsule network is proposed. First, a feature set is constructed by extracting the time- and frequency-domain features through the statistics algorithms. Second, a classifier is established by employing stacked capsule layers, in which the dynamic routing algorithm is used to perform feature cluster and output diagnosis results. Finally, a post hoc interpretability method based on the coupling matrix of the capsule layer is proposed to clarify the decision basis of the model and to improve the transparency of the diagnosis process. The effectiveness of the proposed method is verified in an experiment rig of rolling bearings. The experiment results show that the proposed method can successfully locate the discriminative feature and improve the credibility of diagnosis results on the premise of high diagnosis accuracy.