A deep reinforcement learning-based intelligent fault diagnosis framework for rolling bearings under imbalanced datasets

Abstract

Deep learning is a commonly employed technique for fault diagnosis; however, its effectiveness is contingent upon the presence of balanced data. In real-world industrial settings, the collected fault data from mechanical equipment often lacks balance with normal data, resulting in overfitting, reduced generalization, and diminished accuracy of the deep learning approach. Consequently, this study introduces a novel diagnostic framework, namely Deep Reinforcement Learning (DRL) based on Advantage Actor–Critic (A2C), which autonomously extracts profound and pivotal features from data samples, enabling precise decision-making. In this study, we employ the Synthetic Minority Over-sampling Technique (SMOTE) to create a reinforcement learning environment that facilitates balanced data support for model training. Additionally, we utilize the DenseNet network, enhanced by the multi-scale mixed attention mechanism module, as both the policy and value network for the A2C agent. This allows for the extraction of crucial features while retaining important information. Furthermore, multiple A2C agents are executed in parallel to carry out diagnostic tasks, thereby expediting convergence and ensuring stability. The proposed approach is then evaluated and analyzed using two bearing datasets, and its performance is compared to that of alternative methods. The experimental findings demonstrate that the proposed framework exhibits superior diagnostic accuracy and overall performance.