Fault diagnosis of rolling bearing with uneven data distribution based on continuous wavelet transform and deep convolution generated adversarial network

Generally, the recognition models established by deep learning methods need a large amount of training data. Although normal samples are abundant, fault samples are scarce since while faults seldom occur during most of the operation time. The recognition accuracies of abnormal classes are low for the convolutional neural network with such uneven distribution, especially under strong noise environment. In this paper, one data augment method combining continuous wavelet transform and deep convolution generated adversarial network (DCGAN) is proposed to solve the problem of uneven data distribution in the field of rolling bearing fault diagnosis. Firstly, continuous wavelet transform is used to transform one-dimensional time series data into two-dimensional time–frequency images. Then the time–frequency image samples of fault categories were expanded by deep convolutional generative adversarial network. Meanwhile, the image quality evaluation methods of SSIN and PSNR were used to evaluate the quality of the generated images, and the diversity of the generated images was evaluated by comparing the distribution of uneven distribution dataset and even distribution dataset. Finally, the convolutional neural network is used to classify the expanded time–frequency image dataset. Experimental results show that the proposed method can effectively solve the problem of uneven data distribution in the field of fault diagnosis, and has higher accuracy than the traditional method using one-dimensional data expansion.