Extreme weather events are occurring with increasing frequency, the research on the resilience of power systems under extreme natural disasters has received more and more attention. This paper proposes a high resilience decision-making method based on deep reinforcement learning, and the operation state and line fault state of distribution network under extreme disasters are regarded as the observation state set. In the current environment observation state, self-learning agent seeks feasible decision-making strategies for action, and defines the return function of self-learning agent for action evaluation. Based on the observed state data, the deep reinforcement learning (DRL) training is carried out based on the dueling deep Q network (DDQN). The agent selects the action by trial and error learning, and the trial and error experience is stored in the evaluation function Q matrix to realize the nonlinear mapping from the state to the real-time fault recovery strategy of active distribution network. Finally, the improved IEEE 33 node system is taken as a typical case, based on Monte Carlo method, the random fault scene is simulated, and the random optimization decision of the fault recovery generated by the proposed method is analyzed. The case study show that through the coordinated and optimized control of distributed generation, tie switches and interruptible loads in the active distribution network, power supply capacity in extreme disasters can be effectively improved.