A fault diagnosis strategy for gearboxes based on reinforcement learning with balanced reward and cost- equilibrium sampling

To address the challenge of deep reinforcement learning exhibiting low accuracy in data-imbalanced gearbox fault diagnosis, this paper proposes a Reinforcement Learning Model with Cost-sensitive Sampling and Balanced Reward (RLM-CSBR) from the perspectives of feature representation, data utilization, and learning strategy. To tackle the lack of minority-class feature patterns, a multi-level convolutional deep integrated Q-network is constructed to fully explore deep discriminative features from imbalanced data, thereby maximizing feature-perceptive information. To mitigate model training bias, a balanced reward strategy based on the sample missing rate is designed; this strategy not only guides the agent to prioritize the exploration and learning of sample-scarce categories but also ensures the utilization efficiency of majority-class samples. To solve the problem of insufficient model fitting for minority-class samples, a novel cost-equilibrium matrix is incorporated into the prioritized experience replay mechanism, which prioritizes the selection of high-value experiences learned from critical minority-class samples during training.