Interpretable Constrained Multiplicative Convolution Network for Smart Bearing Edge Intelligent Diagnosis

Abstract

In recent years, deep learning models have shown its power ability for representation and mining, while it has also suffered from high computational costs, limiting their deployment on resource-constrained edge devices. In particular, the insufficiency of model interpretability further undermines the effect for intelligent fault diagnosis in the rail transit application of Industry 4.0. Focusing on the issues of unclear feature extraction mechanisms and addressing fault diagnosis for smart bearings in resource-constrained environments with certain accuracy, this study proposes a lightweight edge intelligent diagnostic method, represented as a constrained multiplicative convolution network (C-MCN). Benefiting from signal processing and deep learning, an improved Wiener filter kernel (IWFK) is designed as the first layer for this interpretable network architecture, where the multiplication layer and the convolution layer are simultaneously operating with few key parameters. In particular, considering the principle of signal decomposition, a residual energy ratio constraint is purposefully introduced to interpretable learning. This will contribute to pre-event interpretability of the lightweight model. Thereby, the fault-sensitive mode information can be well-matched with optimized filter kernel parameters. Experimental and comparative results demonstrate that the lightweight C-MCN optimizes the central position of the filtering kernels through the introduction of residual constraints, thereby achieving complete extraction of mode features. While maintaining high fault recognition accuracy, the proposed model significantly enhances diagnostic efficiency. Furthermore, edge monitoring experiments verify the high efficiency and accuracy of C-MCN in signal decomposition and fault diagnosis, fully highlighting its broad application potential in intelligent edge-based bearing fault diagnosis.