Research On Prediction Algorithm Of Machine Learning Methods For Spacecraft In Orbit Mutation Status

The spacecraft refers to various aircrafts,which fly in the space beyond the earth's atmosphere owing to celestial mechanics.In this paper,we just consider the artificial satellite.As a highly dynamic and complex system,the artificial satellite has its features such as the expensive,high maintence costs,which also destermines it as the great strategic significance for our country.If not handled timely,the abnormal changes of the satellite would resulte in the accidents,and cause severe damage.Since the abnormal change of the key parameter always change with the abnormal change of the satellite,the historical data of the key parameters can be used to predict the abnormal tendency in order to reduce the possibility of irreversible fault.The main contents and contributions of this paper are as follows:(1)The characteristics of the time series of measured data from satellite which provided by the partner is briefly analyzed.According to these characteristics,the data preprocessing process,which includs removal of outliers,interpolation compensation,high frequency noise reduction,is proposed.Besides,the quantitative determination index of model prediction accuracyis in machine learning is illustrated.(2)The support vector regression(SVR)can forecast the tendency on the basis of the measured data effeciently but short time interval here.Aiming at the problem of SVR' parameter optimization,the genetic algorithm and the particle swarm algorithm of the intelligent optimization algorithm are used to optimize the free parameters of SVR.In the particle swarm optimization algorithm,the fixed inertia weight is replaced by the decreasing inertia weight function,and the function with the shrinkage factor is added to the speed iteration,which both can improve its optimal search performance resulting that the prediction model is more accurate.(3)The long-term memory network(LSTM)as a kind of the deep learning model can learn the scenario and forecast the tendency on account of the measured data.In contrast to the SVR,the LSTM can effectively predict the medium and long time tendency.Recurrent neural network(RNN)as a conventional algorithm can obtain certain short-time memory.However,owing to the existence of the gradient dispersion in RNN,the Long-term dependence on time series cannot be solved well.In this paper,the LSTM is proposed to improve the long-term dependence from the structural source.Simultaneously,a better prediction model of the LSTM can be establish through the optimization techniques such as anti-overfitting,mini-batch and adaptive learning rate.In the depth learning model,the more associated scenario,which is before and after the potential mutation,is excavated and studied in time dimension.Finally,this paper uses two types of machine learing models to predict the temporal information tendency is consist of support vector and recurrent neural networks.From the experimental results,it can be concluded that the shallow learning SVR model can efficiently perform short-term prediction and achieve accuracy indicators.Using the intelligent algorithm combination-forecasting model can improve the accuracy of traditional grid method SVR prediction.Using the deep learing LSTM model can complete the long-term prediction and achieve accuracy indicators.The deep learing model,which combined with the deep learning optimization technology,can make it predict more accurate.