Web-winding systems are generally large-scale interconnected systems with many motor-driven subsystems. If the centralized fault diagnosis methods are adopted, the information exchanges among the motor driven subsystems will be required. However, the information exchanges are not always available due to the communication restriction, high system cost and so on. Moreover, it will lead to the high computational cost. To solve this problem well, the web-winding system is considered as a synthetic system with several dynamic subsystems subjected to multiple disturbances and actuator faults. Then, the methods of disturbance attenuation based distributed fault diagnosis method and disturbance compensation based distributed fault diagnosis method are developed to estimate the actuator faults. The objective of fault detection, fault isolation and fault estimation can also be realized via these methods. Meanwhile, sufficient conditions of asymptotic stability of the estimation error system are derived based on the Lyapunov theory. Observer gain matrices are obtained by solving the linear matrix inequalities (LMIs). Finally, simulations and analysis are performed on the three-motor web-winding system to verify the effectiveness of the proposed two fault diagnosis methods.

For the multi-agent formation control problem, the current mainstream control methods are based on graph theory and consistency theory, which often require precise modeling of the system and rigorous mathematical reasoning. Based on the MADDPG deep reinforcement learning algorithm, this paper models the formation control problem as a reinforcement learning problem by considering various constraints on the agent, and obtains the behavior strategy of each agent through learning and training. Furthermore, this paper combines the leader-follow method to divide all agents into two categories, and solves the problem that it is difficult for agents to balance the long-term rewards of formation and obstacle avoidance tasks in reinforcement learning methods. The simulation experiments are divided into two types: circular formation control and formation obstacle avoidance. The dynamic movement process of the agent in the corresponding experimental scene is shown, and the effectiveness of the algorithm is verified by combining the formation process error.