| With the rapid development of the world economy, more and more countries begin to pay more attention to the development of space science and technology. Now, it has been a symbol of a country's comprehensive strength. In the space environment of the spacecraft, due to the disturbances of the space torques. It is inevitable that there will be unpredictable faults, which will directly determine the success or failure of the space mission. In order to ensure the safety, reliability of the spacecraft and the successful development of the space mission, it is essential to carry out intelligent fault diagnosis of spacecraft. Spacecraft attitude control system is one of the most complex subsystems, and its reliability and safety are the sufficient conditions for the normal operation of the spacecraft. The wavelet transform has a good analysis characteristic of multi-resolution analysis of time-frequency two-dimensional signal, while the neural network(NN) has the ability of self-learning, and it can also approximate any nonlinear system. In the fault diagnosis, the wavelet neural network(WNN) is constructed by combining the wavelet transform with the neural network. This method can accelerate the training speed of the neural network, and can quickly detect and locate the faults in the system.In order to know more about the regularity of the spacecraft fault occurrence, the on-orbit spacecraft faults during the last few decades are analyzed in this paper. The failure rate and causes of power subsystem of the spacecraft is analyzed. With the purpose of the spacecraft attitude control system that has the higher fault rate. The closed-loop simulation model of the spacecraft attitude control systems is established based on the analysis of kinematic and dynamic characteristics of spacecraft attitude. Based on BP and wavelet neural networks algorithm structure, the advantages and disadvantages, then the architecture of the single hidden layer recurrent fuzzy wavelet neural network(SLFRWNN), the self-recurrent wavelet neural network(SRWNN), the self-recurrent consequent part fuzzy wavelet neural network(SRCPFWNN) and the single hidden layer feed-forward wavelet neural network(SLFWNN) are proposed. The weights of the SLFRWNN are updated to adjust the information of the hidden layer, and improve the accuracy of the network and the approximation performance of the dynamic nonlinear system. The mother wavelet layer of the SRWNN is composed of self feedback neurons that stored the past information of wavelets to improve the generalization and the dynamic simulation of the network. The SRCPFWNN is a feed forward multi-layer neural network, which uses the SRWNN embed into a TSK fuzzy model to construct a self-recurrent consequent part with the aim of increasing the accuracy of the model as well as the flexibility. It can also realize function approximation and system identification with high accuracy. The adaptive state observer based on the SRWNN, the SRCPFWNN and the SLFRWNN are designed. These neural network adaptive state observers are proved stability by introducing Lyapunov stability theory.The adaptive robust fault diagnosis observers-based the SLFRWNN, SRWNN, SRCPFWNN and the SLFWNN are proposed for actuators and sensors with the higher failure rate in spacecraft attitude control systems. In fact, the nonlinear unknown input detection observer is designed to make a part of the unknown inputs decouple with the residual signal. Fault isolation is achieved by using the SLFWNN to design an adaptive nonlinear unknown input observer. Finally, the effectiveness and superiority of our proposed adaptive robust fault diagnosis state observers are verified by simulating for fault model and comparing with other neural network fault detection observer. The fault diagnosis scheme of the neural network is achieved from the system level to the component level. The real-time online detection and diagnosis of sensor and actuator faults in the spacecraft attitude control system is achieved. |
PDF Full Text Request