Learning-based sliding mode control of discrete markov jump systems with actuator faults via action-critic-disturbance neural network approach

The fault-tolerant control problem for partially knowable discrete-time Markov jump systems is investigated by combining sliding mode control with reinforcement learning. Firstly, a disturbance neural network and a fault observer are utilized to approximate external disturbances and estimate actuator faults. Secondly, an integral sliding mode function is devised and the accessibility of the sliding surface is achieved under an adaptive fault-compensated control that combines disturbance estimation and fault estimation. Thirdly, the reinforcement learning method with action-critic-disturbance neural network framework is designed for sliding mode dynamics that are not subject to actuator faults, enabling partially unknown sliding mode dynamics to achieve stability with optimal performance. In consideration of the attributes of discrete-time Markov jump systems, the cost function in the mathematical expectation sense is constructed and the critic network is designed for each mode. Finally, the integrated control strategy is used in the automobile throttle fault model to demonstrate its effectiveness.