Abstract: Given the difference in fault characterization of the planetary gearbox under multiple working conditions, the existing methods have problems of insufficient feature extraction, low generalization, and diagnosis accuracy, an intelligent diagnosis method using multiscale deep attention Q network (MSDAQN) of deep reinforcement learning for planetary gearboxes under multi-working condition is proposed. Classification Markov decision process is defined to describe the diagnosis process and establish the simulation environment of fault diagnosis. The structure of an MSDAQN deep reinforcement learning agent is designed, which extracts multi-scale fault features by a multi-scale convolutional neural network and uses adaptive channel attention to reweight them and highlight key fault information. Finally, using the interaction experience between the developed agent and diagnosis simulation environment, the optimal diagnosis strategy is learned autonomously. The test and analysis of planetary gearbox's multi-working condition test and actual cases show that the proposed method has higher diagnostic accuracy and strong adaptability.

Key words: planetary gearbox, fault diagnosis, deep reinforcement learning, multi-scale features, channel attention