SDMC-Net: A Lightweight and Deployable Fault Diagnosis Network Using Self-Distillation and Multiscale-Depth Convolution

Deep learning-based fault diagnosis methods have achieved significant success in intelligent diagnosis. As research progresses, the complexity of intelligent diagnosis models, however, increases, leading to a substantial rise in the number of parameters and computational demands. This complexity poses challenges for deployment on low-cost, resource-constrained edge devices that require real-time operation and low power consumption, severely limiting the practicality of intelligent fault diagnosis. This article proposes a lightweight deployable fault diagnosis network called “SDMC-Net.” First, a group conv-transposed convolution (GCTC) block is designed by using group convolution and transposed convolution, significantly reducing the parameters and the computational consumption compared to standard convolution. Then, an efficient multiscale-depth convolution module with a serial structure embed is proposed, increasing network depth and minimizing the growth of the model size. The module effectively extracts composite features of multiscale and multidepth information, enhancing the generalization of shallow networks. Finally, self-distillation from the last batch is employed during model training to maximize the performance of the proposed method. Experimental results on two mechanical systems demonstrate that, compared to other advanced lightweight diagnosis methods, the proposed method features a more lightweight model architecture while achieving optimal diagnosis accuracy and competitive generalization. In addition, the method has been successfully deployed on edge devices, validating its deployability and advantages in real-time fault diagnosis.