MKD-SFasterNet: A Lightweight Edge Computing Architecture for Mechanical Fault Diagnosis

In the era of intelligent manufacturing, faced with the explosive growth of equipment and sensor data, deep learning-based fault diagnosis methods with significant feature extraction capabilities have been widely applied. However, as deep learning networks become increasingly complex, their computational and storage requirements increase, which poses a challenge for resource-constrained industrial scenarios. To address the problem, this study proposes a novel lightweight edge computing architecture, namely multilevel knowledge distillation improved slim FasterNet (MKD-SFasterNet). Firstly, the lightweight network FasterNet is optimized by using channel shuffle and group convolution to minimize the parameters and computational complexity of the model while improving the diagnostic accuracy. Secondly, the MKD method incorporates label decoupling and feature semantic information. It uses the gradient conflict elimination strategy to optimize the loss function, which further improves the diagnostic accuracy of the model without adding additional resources. Finally, the experiment was conducted using two mechanical fault datasets. The proposed method achieves the highest diagnostic accuracies of 94.95% and 94.25% compared to other lightweight networks, while the network’s number of parameters, floating-point calculations, and storage requirements are reduced by 98, 28, and 22 times compared to ResNet. Moreover, inference on edge devices takes only 4 ms, which proves its efficiency and superiority.