A Lightweight Subtraction-Convolution Network via Adaptive Sparse Feature Extraction for Interpretable Intelligent Edge Diagnosis

From the perspective of signal processing collaborated with deep learning (DL), the interpretability of the features separated from the DL model is an important factor affecting reliability and accuracy. Considering the challenges of large amount of data transmission and large-size model deployment in real-time fault diagnosis, this study proposed a lightweight subtraction-convolution network (SCN) for industrial intelligent edge fault diagnosis. An array of randomly initialized sparse kernels (SKs) is designed to interpretably achieve adaptive sparse spectrum feature separation with the ${L} ^{{1}}$ regularization constraint introduced. Additionally, the depthwise separable convolution (DSC) is subsequently employed as a substitute for the conventional convolution operation to diminish computational burden and design a more lightweight model named SCN-L. Self-made extensive experiments indicated that the proposed SCN and SCN-L have shown great lightweight performance, high accuracy effect, and interpretability. A public dataset is used to illustrate the generalizability of the proposed model. Furthermore, an intelligent edge diagnosis node (EDN) hardware with SCN-L is designed to implement efficient industrial intelligent edge diagnosis. The experimental results show that the proposed model performs efficiently in edge diagnosis, indicating great potential for industrial application.