Study On Signal Processing Of Hall Thruster Based On Neural Network

As a power device of space industry,the development of Hall Thruster(HT)will affect the development of space industry in China.The low-frequency oscillation is the inherent property of HT in orbit motion,and the voltage will change greatly with the change of loop current during the oscillation,which will damage the power supply of spacecraft.When HT is in orbit,the morphology of discharge channels will deteriorate over time.Therefore,it is of great significance to study the channel wall erosion to understand the physical mechanism of low frequency oscillation and to predict the lifetime of HT.In recent years,neural network has risen rapidly in all walks of life due to its better computing power and better algorithms.Neural network models are established in various fields to classify or predict.In view of its good predictive ability,neural network models with different algorithms can be established to predict some parameters in HT that are difficult to measure.This article is based on a one-dimensional quasi neutral fluid model,building dynamic simulation platform,and understand the dynamic characteristic of the HT on-orbit working,the simulation process,it can be seen that the discharge current and discharge voltage have opposite trends over time,can also be observed in the electric field intensity,the density of atoms,ion density and other parameters with the change of time and space.In order to break the deterministic relationship between wall erosion and discharge current low-frequency oscillation,the fluid model was improved by adding disturbance.The influence of channel wall erosion on low-frequency oscillation was studied by changing the cross-sectional area.It was found that the parameters such as frequency and amplitude of discharge low-frequency oscillation would change with the change of cross-sectional area.According to the information of low frequency oscillation,the wall erosion of discharge channel is predicted by the information inversion.The wall erosion at the exit of the channel was taken as the research object,and a total of 491 sets of sample data were collected.The sample waveform was processed by using the method of smoothing and denoising.After comparing the training results before and after denoising,it was found that a better model could be established after the denoising of the samples.In this paper,a neural network nonlinear model is established to predict the erosion of the channel wall.Elman neural network was established to study the effects of the number of hidden layer nodes,training function and training times on the model.Compared with BP neural network and RBF neural network,the results show that the training error of Elman neural network algorithm is small and stable,and the effect of multiple predictions is consistent,among which the root mean square error,mean absolute error and mean absolute percentage error are 0.0084,0.0637 and 0.045%respectively.Finally,ground tests are carried out to verify that Elman neural network can predict wall erosion.