Lightweight neural networks with anti-colored noise for bearing fault diagnosis using deep separable convolution and transfer learning.

Practical bearing fault diagnosis faces several key challenges: noise interference, performance degradation under varying operating conditions, and slow diagnostic speed. For such engineering problems, this article combines deeply separable convolution to establish three lightweight neural network fault diagnosis models-MobileNet-DLCNN, ShuffleNet-DLCNN and SqueezeNet-DLCNN. At the same time, the failure mechanism of rolling bearings is studied. This article incorporates colored noise into the standard bearing failure datasets from Case Western Reserve University (CWRU) and the American Society for Machinery Failure Prevention Technology (MFPT), and conducts noise resistance training under variable adaptive characteristics, compares the diagnostic ability with the datasets before adding colored noise. Then, the identification and classification abilities of the three lightweight convolution models proposed under each dataset, as well as the transfer learning adaptive ability of the three models between different datasets and working conditions, are verified through comparative experiments. The experimental results show that SqueezeNet-DLCNN has the best diagnostic performance among the three models, and it can complete the recognition and classification of all data in about one minute with an accuracy of 97%. The lightweight convolution designed in this article has the advantages of strong noise resistance, high efficiency, and fast speed.