Neural networks are very promising in the design and control of distributed systems due to their powerful computing power, learning and adaptive capabilities, and distributed capabilities. The purpose of this work is to study fault-tolerant control of distributed systems based on neural networks. For nonlinear systems with actuator failures, fault-tolerant control based on a small portion of the network is required. By combining error tracking and forecasting methods to build performance metrics and reduced fault-tolerant systems, this translates into a control problem. The PSO algorithm is introduced to train the judgment neural network, which avoids the selection of the initial vector and improves the training success rate to a certain extent. To mitigate the effects of actuator errors, a neural network error observer is designed to estimate and compensate for errors. Finally, the effectiveness of the proposed fault-tolerant supervisory control method is proved by an example simulation.