New transfer learning fault diagnosis method of rolling bearing based on ADC-CNN and LATL under variable conditions

Convolutional neural network with transfer learning are effective methods for rolling bearing unsupervised learning fault diagnosis. In view of the problem that 1D-CNN cannot give full play to the feature extraction, an improved Adaptive Dimension Convert convolutional neural network (ADC-CNN) is proposed, which can adaptively process one-dimensional vibration signals into two-dimension matrices and input them into 2D-CNN for learning, making full use of the ability of CNN to extract two-dimensional data features. In order to further reduce the data distribution distance between source domain and target domain, the training method of transfer learning is improved by Layered Alternately Transfer Learning (LATL), which layering calculate the CORAL and MK-MMD loss function alternately. To verify the reliability of the proposed method, we carry out experimental verification on the rolling bearing datasets of CWRU and PU. Compared with the traditional 1D-CNN model, the diagnostic classifier accuracy of the proposed ADC-CNN-FLATL is improved by 9% per transfer mission on PU dataset on average, which proves the validity of the proposed method.