Research on fault diagnosis of rolling bearing based on lightweight convolutional neural network

Convolutional neural network models in rolling bearing fault diagnosis have problems such as relatively large number of parameters, overly complex structure and long computation time. To address the above issues, this paper designs an improved fully connected layer (FC) layer in the multilayer perceptron model that can be used for rolling bearing fault diagnosis. Based on this idea, this paper proposes a "Reshape, Linear, Transpose, Linear" fully connected layer replacement strategy (RLTL) by using the Kronecker decomposition method. This method decomposes the weight matrix, converting one linear operation to two linear operations in the FC layer, which results in a significant reduction in the total number of parameters and calculations. On this foundation, this paper also uses one-dimensional convolution to improve the current mainstream two-dimensional lightweight convolutional neural networks and lightweight strategies, and designs nine alternative modules. The results show that the proposed RLTL module replacement scheme decreases the number of parameters of the multilayer perceptron to less than 1%, reduces the number of calculations to less than 10%, and increases the diagnostic accuracy by more than 30% in a small-sample and noise-free environment. In addition, this paper also verifies and compares the effects of different transformation methods of convolutional neural networks in each alternative module on small sample diagnosis and noise immunity diagnosis of rolling bearings.