Abstract
Intelligent fault diagnosis of bearings is crucial to the safe operation and productivity of mechanical equipment, but it still faces the challenge of difficulty in acquiring real fault data in practical applications. Therefore, this paper proposes a domain adaptive meta-relation network (DAMRN) to achieve diagnostic knowledge transfer from laboratory-simulated faults (human-induced faults) to real scenario faults (natural faults) by fusing meta-learning and domain adaptation techniques. Specifically, firstly, through meta-task scenario training, DAMRN captures task-independent generic features from human-induced fault samples, which gives the model the ability to adapt quickly to the target domain tasks. Secondly, a domain adaptation strategy that complements each other with explicit alignment and implicit confrontation is set up to effectively reduce the domain discrepancy between human-induced faults and natural faults. Finally, this paper experimentally validates DAMRN in two cases (same-machine and cross-machine) of a human-induced fault to a natural fault, and DAMRN outperforms other methods with average accuracies as high as 99.62% and 96.38%, respectively. The success of DAMRN provides a viable solution for practical industrial applications of bearing fault diagnosis.