Abstract: Traditional maintenance methods suffer from the curse of dimensionality when dealing with partially observable and multi-state systems. To address this issue, we propose a maintenance strategy for multi-state partially observable systems based on deep reinforcement learning. Firstly, we construct a maintenance strategy model based on a partially observable Markov decision process. Then, we employ a deep learning framework for solving this model by introducing an improved Double Deep Q-Network (DDQN) algorithm. The improved algorithm optimizes the experience replay process through prioritized experience replay and estimates the value function using deep neural networks, which effectively resolves the low sample utilization problem faced by traditional DDQN during training, thereby enhancing the learning efficiency and convergence speed of the DDQN algorithm. To verify the effectiveness of the model and the efficiency of the improved algorithm, numerical examples are provided based on a real coke oven gas to methanol synthesis process system. The results of the examples show that the improved algorithm significantly outperforms traditional methods in terms of maintenance efficiency and system reliability, which demonstrate the validity of the model and provide decision support for complex systems’ maintenance.