Fault diagnosis of gearbox under open set and cross working condition based on transfer learning

With the continuous development of industry and aerospace technology, the working conditions and failure modes of rotating machinery are becoming increasingly diversified and complex, and reliability and safety problems are becoming increasingly prominent. It is critical to research efficient fault detection techniques since many working condition data lack fault labels and the failure modes amongst various working conditions are asymmetric. Take the gearbox as the case verification object, set up cross-working conditions, and open set fault diagnosis scenarios. A method is proposed to address the issue of lacking fault labels under the target working condition. It takes into account the ability of migration learning to facilitate cross-domain knowledge application. Specifically, migration learning is used to transfer knowledge from the source working condition to the target working condition, and the cross entropy classification loss function is used to identify known fault types. However, transfer learning has the problem that the greater the field difference is, the worse the effect is. It is difficult to solve the open set problem of asymmetric fault modes under cross-working conditions. In order to identify the known and unknown classes of target working circumstances, a method utilizing a convolutional neural network to extract similar data characteristics between working conditions is proposed. The two classification loss functions are then used in this process. The joint loss function is proposed to train the diagnosis model and realize the joint migration of fault features from the source domain to the target domain. The results of the case analysis show that the method can realize cross-working condition fault diagnosis under an open set, and the average diagnostic accuracy is more than 90%. © 2024 Beijing University of Aeronautics and Astronautics (BUAA). All rights reserved.