Lightweight MDSCA-Net: An end-to-end CAN bus fault diagnosis framework

Abstract

Controller area network (CAN) buses are widely used as low-cost, highly flexible field buses in various scenarios, such as in vehicle networks for automobiles and communication networks for industrial sites. They typically operate in harsh environments, and faults inevitably occur. CAN bus faults cannot be efficiently diagnosed via traditional manual detection. Herein, we propose a lightweight MDSCA-Net for CAN bus fault diagnosis. Deep separable convolution (DSConv) is used in the model instead of ordinary convolution to reduce the number of parameters and floating-point operations. Additionally, the noise immunity of the model is improved by designing a multiscale denoising module (MDM). A multiscale deep separable convolutional fusion SE at tention (MDSCSA) module is designed to capture the channel dimension details of the features. Furthermore, a spatial attention module (SAM) is utilized to capture the spatial dimension details of the features. Finally, a residual (Res) module stabilizes the model performance. Experimental results on the CAND dataset indicated that the proposed method achieved a diagnostic accuracy of 99% in a noise-free environment, and compared with other fault diagnosis methods, it had better noise immunity and robustness in a noisy environment, which is of considerable practical significance for ensuring the stable operation of CAN buses.