Embedded Bearing Fault Diagnosis Based on SP-WDCNN

In response to the problems of large size, numerous parameters, poor noise resistance and difficulty in achieving real-time diagnosis in traditional bearing fault diagnosis models, this paper proposes the SP-WDCN model. Taking the original time-domain vibration signal as the input, it simulates the actual industrial environment. After data preprocessing, the model is optimized using techniques such as batch normalization and depthwise separable convolution. Compared with algorithms like WDCNN, this model has a smaller size, lower computational complexity, and significantly shorter inference time. It has the advantages of efficiency, accuracy and robustness in embedded systems. Later experiments show that, compared with models like MLP and ResNet, the training and testing time is shorter, the total number of parameters is fewer, and its accuracy rate for diagnosing 9 types of bearing faults reaches 99% on the ARM Cortex M7 core processor. The complexity of accumulation and addition operations is reduced by 70.4%, the Flash storage occupation is reduced by 93.22%, the inference time is shortened by 84.07%, and the accuracy rate remains nearly unchanged. This provides a highly practical solution for real-time diagnosis of bearing faults in industrial production, and has important theoretical significance and practical application value for promoting the development of industrial intelligence.