A RUL prediction method based on multi-scale convolutional feature extraction and deep reinforcement learning

Remaining useful life (RUL) is crucial to enhance the reliability and safety of complex and systematic mechanical equipment. Current machine learning-based methods have poor ability to capture temporal dependencies in the prediction process, and the RUL prediction fluctuates greatly. In this paper, we propose a RUL prediction method combining multi-scale convolutional feature extraction and deep reinforcement learning for partially RUL label missing datasets to achieve effective delivery of temporal dependencies and reduce the significant variations in predicted RUL. Firstly, Generative Adversarial Interpolation Network (GAIN) is used to realize the interpolation of missing vibration data, after which a feature extractor based on multi-scale convolutional module and attention mechanism is used to extract features from the time-frequency matrix of the vibration data, and the extracted features are used for the construction of the state variables of reinforcement learning. The reinforcement learning model is trained by setting the appropriate action space and reward function, so that it can realize the transfer and utilization of timing information in the RUL prediction process. Finally, the proposed method is experimentally validated using the bearing dataset XJTU-SY, and the proposed method outperforms many machine learning-based RUL prediction methods in both RMSE and MAE.