An Elastic Expandable Fault Diagnosis Method of Three-Phase Motors Using Continual Learning for Class-Added Sample Accumulations

Continual learning is promising in intelligent fault diagnosis of three-phase motors, which allows diagnosis networks to increase diagnosable fault classes without tedious retraining during various mechanical and electrical fault sample accumulations. Existing studies, however, have the following limitations. 1) The representation learning ability of diagnosis networks is nonaugmentable because of predefined and fixed network structures. 2) The impacts of different working conditions are not explicitly considered during continual learning. To overcome the abovementioned limitations, an elastic expandable continual learning framework is proposed for class-added motor fault diagnosis. First, an elastic expansion mechanism is developed to improve the representation learning ability during continual learning by progressively introducing new network branches. In addition, the new branches are pruned concurrently with training to elastically adjust the network sizes, thereby avoiding bringing excessive computational complexity. Then, an adaptive multiscale convolution mechanism is designed to learn robust fault features by fusing working condition information, increasing the diagnostic accuracy of continual learning under complex working conditions. Finally, a new training strategy with a multiobjective loss is formulated to guide continual learning. The effectiveness of the proposed method is demonstrated through motor fault simulation experiments.