Deep learning may encounter challenges in fault diagnosis, such as exploration capability, long-term planning, and handling non-deterministic factors. The paper introduces a novel fault diagnosis method for rolling bearings combining deep Q-network (DQN) with discrete random separation (DRS) frequency spectrum images. DRS can be used to separate the deterministic components and random components of the fault signal and extract the fault feature very effectively. The improved DQN incorporates the atrous spatial pyramid pooling module for multiscale contextual fault feature extraction, enhancing diagnosis accuracy. Various deep networks, including convolutional neural network, ResNet18, traditional DQN, and the improved DQN, are employed with different frequency spectrum images (power spectral density, cepstrum, and DRS) for diagnosing bearing faults. Simulation results demonstrate that combining DRS frequency spectrum images with the improved DQN enhances fault diagnosis accuracy across diverse conditions. Generalization tests reveal strong capability of the proposed method in handling conditions different from the training data. Validation tests on difficult-to-diagnose fault data from the CWRU-bearing dataset demonstrate commendable performance of the improved DQN, even on difficult datasets. Finally, validation tests using the XJTU–SY bearing dataset reaffirm the excellent performance and robust adaptability of the improved DQN in conjunction with DRS frequency spectrum images.