A reinforcement double deep Q-network with prioritised experience replay for rolling bearing fault diagnosis

Abstract

Abstract In recent years, deep learning has been increasingly applied to fault diagnosis and has attracted significant attention and research interest. Deep reinforcement learning (RL), with its capabilities in feature extraction and interactive learning, is highly suitable for fault diagnosis problems because it can acquire knowledge solely via system feedback. Despite its advantages, this method also has limitations, such as low training efficiency and unstable performance. Therefore, this study presents a novel diagnostic approach based on system feedback for rolling bearing fault diagnosis. This approach builds upon the original deep Q-network (DQN) approach, which incorporates an interactive dual network structure and experience replay optimisation for RL intelligence. This method introduces two major improvements. First, a dual network cyclic update scheme is implemented, assigning each dual network specific responsibilities to ensure training stability. Second, a novel experience playback system is introduced, which improves the efficiency of experience utilisation while circumventing the risk of overfitting. Compared with the original DQN method, the proposed approach and its two enhancement strategies provide significant advances in training efficiency, stability and diagnostic accuracy. Our experimental results indicate that this novel methodology has the potential to make valuable contributions in the area of rotating machinery fault diagnosis.