Abstract: Aiming at the problem that the traditional rolling bearing fault diagnosis method has insufficient feature extraction and poor diagnosis rate under variable noise, working conditions, and load conditions, a rolling bearing fault diagnosis method of multichannel convolutional neural network (MCNN) is proposed. Firstly, a multi-channel dense connection module is designed, which strengthens the information connection between different convolutional layers through dense connection and effectively extracts fault information. Then, a hollow convolution module with adaptively parametric rectifier linear unit (APReLU) is designed, which assigns different weighting coefficients to each channel to extract more important and critical information. Finally, the Inception module is used to reduce the feature dimensionality and further extract the fault features. Fault diagnosis of rolling bearings is realized by means of multiple classification functions. The method was validated using the bearing dataset of Case Western Reserve University and the gearbox dataset of Southeast University. The experimental results of the bearing data set show that under variable noise conditions, the proposed method achieved an average accuracy of 98.50%. Under varying load conditions, the accuracy ranged from 91.7% to 97.7%, and under varying operating conditions, from 87.79% to 96.71%. In the gearbox data set, the fault diagnosis accuracy is as high as 99.84%. Compared with rolling bearing fault diagnosis other methods, the proposed method has a higher fault diagnosis rate and better generalization performance across different data sets and variable noise, load and working conditions.

Key words: feature extraction, dense connections, convolutional neural networks, Inception module, identify classifications