Inter-class Metric Learning based Depth-wise Convolutional Neural Network for Gear Fault Diagnosis

This paper proposes an inter-class metric learning-based depth-wise convolutional neural network algorithm to diagnose the failure of the gearbox. According to the characteristics of the given triaxial vibration signals, we constructed our model that can extract features for each channel. Moreover, the proposed method includes a new inter-class maximum mean discrepancy loss function that can maximize the distance between fault types. This can help the network to learn not only learn useful features, but also increase accuracy due to metric differences in latent space. Finally, the robust diagnosis performance can be achieved through a soft voting-based ensemble procedure. The results show that the proposed method successfully classified the health states of the gearbox; the proposed method was ranked in first place in the IEEE International Conference on Prognostic and Health Management (ICPHM) 2023 Data Challenge. In addition, through ablation study, it is validated that each component in our methods contributes to improving diagnosis performance.