In this study, we investigate the active fault tolerant control(AFTC) of unmanned surface vehicles (USV) under actuator faults and environmental disturbances using deep reinforcement learning(DRL). We realize the perception of the environment and continuous control through end-to-end learning based on neural networks(NN). The contents of AFTC include 1) actuator fault estimation(FE) based on long short-term memory(LSTM) and 2) a fault tolerant control(FTC) framework based on a deep deterministic policy gradient to enable the USV to learn autonomous decision making through training. We introduce mixed noise based on Gaussian noise and Ornstein–Uhlenbeck noise into the training, which increases the exploration space and improves the generalization ability of the USV. An experiential replay method based on quantum computing is designed to distinguish the importance of samples through preparation and depreciation operations, enabling a guided learning process. Finally, we verify the universality and effectiveness of the AFTC framework by tracking straight and sinusoidal trajectories.