Recurrent Neural Network Based Robust Actuator and Sensor Fault Estimation for Satellite Attitude Control System

For the satellite attitude control system with actuator and sensor fault, the paper proposes a neural network based robust state and fault estimation method. Compared with the traditional model-based approach that relies on the accurate model of the system, we train the recurrent neural network with the inputs and outputs of the current attitude control system to achieve the purpose of modeling and improving the accuracy of the model. Then, through the expansion of the system state vectors, the neural state space model is transformed into a generalized nonlinear system without sensor fault terms. Furthermore, a combination of the generalized unknown input observer scheme with the robust ${H}\infty $ linear parameter-varying (LPV) approach is developed to estimate the system state and actuator faults simultaneously. According to Lyapunov theory, the stability analysis of observer is considered by transforming the dynamic error into the discrete time polytopic LPV form. Finally, some tests are performed on the satellite attitude control system to validate the effectiveness of the proposed method.