ShuffleNet2MC: A method of light weight fault diagnosis

Abstract

Bearing fault diagnosis plays an important role in the field of modern industry. Although convolution neural network achieves good results, large amount of parameters costs a lot of calculation, which brings challenges to the deployment of fault diagnosis tasks in low computational power equipments. To solve the problems, an novel CNN model ShuffleNet2MC based on improved Shufflenetv2 network is proposed. Firstly, Depthwise convolution and Channel Shuffle are used to reduce the computational cost while ensuring the accuracy of diagnosis computation; Secondly, mixed convolution is used to extract the features of different resolutions through multi-scale and multi-channel method, which improves the accuracy of the model; Finally, K-means quantization is applied to the model, which greatly reduces the GFLOPS of the model and further improves the performance of the model while ensuring that the accuracy is basically unchanged. A large number of experiments on the bearing fault dataset of Western Reserve University show that: The times of floating point operation and classification accuracy of ShufflenetV2 are 0.001GFLOPS and 97.9% respectively in the task of fault diagnosis. Compared with other models, it not only reduces the model parameters and compresses the model, but also gets better classification accuracy.