محمدرضا هاديان

Fault diagnosis of rotating machinery faces challenges in generalizability an‎d efficiency due to its reliance on manual feature extraction an‎d the need for extensive tuning of deep network architectures. In this research, an intelligent fault diagnosis framework based on Deep Reinforcement Learning (DRL) is proposed, which models the problem as a discrete guessing game, allowing for human-like learning from the environment. This framework utilizes an improved Deep Q-Network (DQN) an‎d maps raw vibration data to fault states end-to-end. To accelerate the convergence of the DRL algorithm, the parameters of a Convolutional Residual Autoencoder are extracted in an unsupervised pre-training process an‎d transferred to the DQN network. Furthermore, the reward function was redesigned by weighting the minority classes to improve diagnostic performance on imbalanced data. Experiments on the CWRU dataset showed that the proposed method achieved an accuracy of 99.2% on balanced data, an‎d its convergence was approximately 20 times faster than the baseline DQN method. On imbalanced data with a 20% imbalance ratio, the accuracy reached 88.7%, which was a significant improvement compared to the baseline method (75.3%). By eliminating the need for manual pre-processing an‎d utilizing the total reward, this framework provides a robust an‎d specialized knowledge independent system that is applicable for online monitoring of industrial machinery.

تشخيص عيب هوشمند , ماشين‌هاي دوار , يادگيري تقويتي عميق , شبكه Q عميق

Intelligent fault diagnosis , Rotating machinery , Deep reinforcement learning , Deep Q network

Mohammadreza Hadian

Dr. Javad Poshtan