A multilayer transfer convolutional neural network for bearing fault diagnosis at variable speed

Intelligent algorithm is widely used in the field of device health monitoring because of its good feature extraction ability. However, the fault features for rotating parts of mechanical equipment often change with the change of rotating speed. Failure labels at various speeds are impossible to obtain. In this case, the traditional deep learning model often fails to achieve good diagnostic results. In this paper, a multilayer transfer network relied on the CNN module named MTCNN is proposed to overcome the above difficulties. The GAP layer is employed in the CNN architecture. The Maximum Mean Discrepancy (MMD) transfer function is applied to the second last layer and the deep coral transfer function is used on the last layer. Finally, a set of test data from the test bench is applied to verify the manifestation of the MTCNN architecture. The performance shows that it has good diagnostic capability and strong domain adaptive capability.