| With the rapid development of our country in the field of aerospace,the requirements for launch vehicles are higher and higher.The reliability of the rocket servo control system is directly related to the success and failure of the whole launch mission.Therefore,it is necessary to diagnose the fault of the rocket servo control system.This paper first had introduced the theoretical basis of simplifying the rocket servo control system into an inverted pendulum system,deduced and established the mathematical model of the three-stage inverted pendulum,designed the LQR controller of the three-stage inverted pendulum,and realized the stable control of the three-stage inverted pendulum.Secondly Elman neural network model and training method had been introduced.Elman neural network had been trained as the state observer of inverted pendulum system.On this basis,three kinds of nozzle sensor fault types of rocket servo control system had been simulated,and the sensor fault data had been obtained.The working principle and training method of deep confidence neural network had been analyzed in detail.The influence of super parameter setting of deep confidence neural network on the accuracy of network diagnosis had been studied,and the optimal value of network super parameter under sensor fault data had been determined.Aiming at the problems of slow network training speed and low network diagnosis accuracy for highdimensional data,this paper had analyzed the optimization principle of momentum gradient algorithm and RMSprop algorithm,and summarized the steps of Adam algorithm combining the two algorithms,and proposed an optimization method combined Adam algorithm with independent adaptive learning rate algorithm,which can improved the network by dynamically adjusting the size of learning rate the training effect of collaterals.Finally,the method was used to optimize the parameters of depth confidence neural network.The results show that the diagnosis accuracy of the optimized depth confidence neural network was higher than that of the standard depth confidence neural network,and the training iterations was lower. |
