Research On Spacecraft Fault Prediction Algorithm Based On Machine Learning

Spacecraft is a system with complex physical structure,which is composed of top technology.It has important influence in military application,manned space and deep space exploration.It is helpful to establish a prediction system to ensure the reliability of spacecraft in orbit operation,ensure its long-term health work,reduce the occurrence rate of system failure and improve its operation life.A large number of telemetry data will be generated when the spacecraft is in orbit,which can reflect the operation status of the spacecraft and the condit ion of the equipment.In this paper,the telemetry data of spacecraft is taken as the research object.Through the analysis of telemetry data prediction methods in recent years,the characteristics of different models are summarized.The spacecraft teleme try data provided by the partner has the characteristics of dense data and slow data change.In this paper,a set of data preprocessing process is constructed.Outliers are eliminated by 3σ principle,missing values are filled by KNN method,dimension differences among parameters are eliminated by data normalization,features are constructed by sliding window,and quantitative indexes of prediction model are given.For the support vector regression model,this paper uses telemetry d ata to train and verify the model.In order to improve the prediction performance of SVR model,particle swarm optimization algorithm is used to optimize the regularization constant.Aiming at the phenomenon of periodic signal stacking in telemetry data,this paper uses empirical mode decomposition algorithm to decompose the original data into multiple intrinsic mode components,then uses sequence model to predict each intrinsic mode component,and finally reconstructs each predicted component.In the sequence model,this paper uses the long and short term memory network model,and establishes the EMD-LSTM model through regularization technology,dropout method and adaptive learning rate setting.The experimental results show that the goodness of fit index of SVR model is0.5563,and the index of PSO-SVR combination prediction model is 0.8917,which has a great degree of performance improvement.The goodness of fit index of LSTM model is 0.2473,and the prediction index of EMD-LSTM model is 0.5658.