Abstract: Aiming at the problems of imbalance in sensor samples and low autonomy of fault diagnosis for deep space probes’ attitude control systems, an autonomous fault diagnosis method is proposed based on feature selection and deep Q-network (DQN). In this method, sensor signals are first converted into time-domain and wavelet packet energy feature parameters. A distance-based evaluation criterion is then constructed for feature selection to obtain a more representative feature subset. Subsequently, the fault diagnosis issue is modeled using the imbalanced classification Markov decision process, and a specialized reward function is designed for the imbalanced samples, enabling DQN to focus more on minority-class samples. Finally, through training conducted via environmental interactions, DQN can autonomously learn the optimal diagnostic strategy. The simulation results demonstrate that the proposed method achieves high diagnostic accuracy and stability on multiple imbalanced datasets.