Deep Reinforcement Learning Bearing Fault Diagnosis Method Based on Improved Reward Function

In deep reinforcement learning, the traditional reward function is limited in the problem of multi-class sample imbalance. In this paper, an improved reward function based deep reinforcement learning bearing fault diagnosis method is proposed. Firstly, the method improves the definition of reward function, and introduces a weak punishment mechanism to deal with the situation where the diagnosis action is wrong but the fault label is similar. Then, the samples are divided by K-means clustering method, and the reward value is newly defined according to the Euclidean distance from the cluster center and the proportion of sample labels. This method makes the diagnostic model explore and learn more effectively, and can deal with multi-class classification tasks effectively. The experimental results show that the method achieves remarkable performance improvement in multi-class classification tasks, and improves the accuracy and recall rate.