Research On Health State Prediction Of Satellite Attitude Control System Based On LSTM

With the rapid development of China’s satellite business,the health management of satellite system is paid more and more attention.Satellite attitude control system is the core and the subsystem with the highest failure rate.Its safe operation is the key factor to ensure the stable operation of satellite.However,due to the complex working environment and strong correlation between components,it is difficult to rescue the satellite once it breaks down.At present,health status assessment and prediction of satellite attitude control system are mostly based on traditional models and expert experience methods,which are inefficient and subjective.Therefore,based on the characteristics of large scale,high dimension and strong correlation of satellite telemetry data,this dissertation intends to use time series prediction and deep learning algorithms to evaluate and predict the health status of satellite attitude control system.It mainly includes the following contents:(1)According to the actual satellite telemetry data provided by the project,this study corresponds to the basic structure of the satellite attitude control system,and divides the telemetry parameters of each component into core parameters and conventional parameters.In this study,the health state assessment system of satellite attitude control system is constructed according to the actual needs of the subject,and the health assessment index set of satellite telemetry data is built.(2)In view of problems such as loss,noise interference,outliers and large amount of data in the actual satellite telemetry data,the pre-processing and dimension reduction of satellite telemetry data are studied.Firstly,the telemetry data are processed by deweighting,sorting cleaning,outliers elimination,noise filtering and data compression,and the variation law is analyzed based on morphological characteristics.Then,the correlation analysis of satellite telemetry data is carried out based on MIC,and XGBoost is used to calculate the feature importance to achieve dimension reduction,so as to obtain highly correlated multidimensional input telemetry data for different prediction targets.(3)The attention-based LSTM-Dropout model is proposed to predict the core parameters for the problems of strong correlation and large dimension of satellite data.The model input was multi-dimensional telemetry time series data obtained based on multivariable dynamic sliding window.MSE,RMSE,MAE,MAPE,NMSE and Loss curves were selected as error evaluation indexes.By comparing with the RNN,LSTM,GRU,Bi LSTM and LSTM-Dropout basic models,the stable,periodic and fluctuating telemetry parameters are predicted respectively.The results show that the prediction accuracy of the proposed model is improved by about 70.56%,68.04%and 97.71% respectively compared with the model with the worst effect,and the performance of the proposed model is improved by about 17.208%,9.586% and 82.928% compared with the model with the best effect.This shows that the proposed model has lower prediction error,higher model confidence and higher prediction precision for different morphological characteristics of telemetry parameters,with better prediction performance for fluctuating telemetry parameters.(4)Proposed the satellite attitude control system health status evaluation and prediction model based on cloud model and Bi LSTM.The cloud model is used for comprehensive evaluation of complex system,and the conversion between quantitative value and qualitative concept is realized after the weight vector is obtained by entropy weight method.RNN,LSTM,GRU and Bi LSTM models are used to predict the system health degree layer by layer through AHP structure.The comparison experiments show that the Bi LSTM model improves the accuracy by about 71.453% with lower input data complexity and does not depend on the experience of experts,which provides better prediction results than the traditional health assessment methods.