A multi-branch convolutional transfer learning diagnostic method for bearings under diverse working conditions and devices

Abstract

Conventional intelligent bearings fault diagnosis methods generally extract fault features with a single channel, which seriously limit the features richness and the diagnostic accuracy. A novel end-to-end diagnosis framework which incorporate multi-scale average processing into the multi-branch convolutional neural network (MBCNN) is established here, thus rich and complementary fault features can be captured from the multi-scale reconstructed signals via multiple parallel residual CNN branches. Besides, with the help of a fine-tuning based transfer learning strategy, the well-trained model on source domain is competent to the related diagnostic tasks of bearings under diverse working conditions and devices. Comparative experiments under 3 cases are conducted, and the results show our approach can be provided with more excellent performances and superior transferability compared with the state-of-the-art ones. For instance, the transfer diagnostic accuracies of MBCNN exceed 98.56% and 98.38% under diverse working conditions and cross-devices, respectively.