Actor-Critic deep reinforcement learning-based integrated fault diagnosis method for nuclear power plants

Although the international community has carried out a lot of research work on fault diagnosis, the nonlinear and high coupling characteristics of nuclear reactors and complex transient conditions put forward higher requirements for fault diagnosis algorithms. In this paper, an integrated method based on reinforcement learning is proposed, in which the base models of the integrated model are pre-trained machine learning units, and Twin delayed deep deterministic policy gradient is used to learn dynamic weight update strategies to adjust the diagnosis probability of the base models based on One-vs-Rest method. An idea of enhanced value propagation is designed and applied to the reinforcement learning. Experimental results show that the fault diagnosis integrated framework based on reinforcement learning has higher accuracy and lower misdiagnosis rate, the enhanced value propagation is proved to be able to obtain higher fault diagnosis accuracy with faster convergence and stronger optimization ability.