DAGAN-FD based fault diagnosis under unbalanced data

Aiming at the problem of inaccurate extraction of fault features and high misclassification rate of bearings in unbalanced data, a rolling bearing fault diagnosis method based on improved Data Augmentation Generative Adversarial Networks DAGAN Data Augmentation Generative Adversarial Networks is proposed. Firstly the fault data are reconstructed using Gram s corner field then DAGAN-FD is trained to augment the fault image data and realistic generation samples are output to achieve the purpose of balancing the data and finally the attention mechanism convolutional neural network model SE-CNN is used for fault diagnosis. The CWRU dataset is used for experimental validation and the diagnosis rate of the model for the test set containing multiple fault types is increased from 92.1% to 99.5% after the DAGAN-FD is refined and expanded for the unbalanced dataset proving the superiority of the DAGAN-FD model as well as the SE-CNN network.