A New Bearing-Fault Classification Method Based on Deep Learning*

Bearing fault signal detection plays a vital role in the industrial field, which directly affects the performance and safety of mechanical equipment. The application of CNN (convolutional neural network) for fault signal detection is an emerging method, but this method does not work well on one-dimensional data such as bearing fault signals, mainly because the features of the one-dimensional signal are not distinct compared to the image signal. Secondly, due to the limitation of the special situation, the data of the bearing signal is less, which makes it hard for the deep learning model to fit and converge well. To solve the above problems, this paper proposes a CNN based on improved softmax-loss (ISM-CNN). The constructed CNN can learn more subtle features from the bearing signals, thereby improving the accuracy of bearing signal classification. Besides, the algorithm proposed in this paper expands the training data set to a certain extent, so that the parameters of the ISM-CNN can be better fitted. We validate the effectiveness of the proposed algorithm on the CWRU open dataset and give ablation experiments to prove it. In the 97-category complex bearing signal generation scenario, the proposed algorithm achieves 95% accuracy.