Application And Research Of Machine Learning In Payload PHM System

In recent years,with the rapid development of China’s space industry,the number of on-orbit satellites has continued to increase.Taking the space science pilot project of the Chinese Academy of Sciences as an example,since the launch of the Dark Matter Particle Detection Satellite in December 2015,there have been quantum scientific experimental satellites and hard X-rays.Modulated telescope satellites successfully launched.Combining the actual satellite in-orbit operation experience,the prediction and analysis of the satellite telemetry parameter data has become a hot issue in the study of the payload PHM system,and the rapid development of computer technology represented by machine learning is becoming an important breakthrough of the payload PHM technology.This paper firstly studies the time series model and the characteristics of the real satellite’s telemetry parameter data.It shows that the time characteristics of the satellite telemetry parameters accord with the application conditions of the time series model.Then based on the time characteristics of the satellite telemetry parameter data,the satellite telemetry parameter data is obtained.The problem of trend forecasting is transformed into a problem of forecasting the trend of time series,and based on a large number of historical telemetering parameter data of satellites generated during the orbital operation of satellites and quantum scientific satellites of dark matter particles,combined with the actual space of space science pilots of the Chinese Academy of Sciences.In the background of scientific satellite operation and control tasks,the anomaly detection in payload PHM system was studied,and a satellite payload anomaly detection method based on time series model was proposed.Focuses on the study of polynomial fitting extrapolation algorithms,ARMA algorithm,deep learning LSTM model,three machine learning algorithms,and short-term experiments on quantum scientific satellites and dark matter particle detection satellite telemetry parameter data sets.The optimization of the medium,long-term,and long-term prediction experiments has greatly improved the accuracy of model prediction results.Finally,the error rates of the prediction results after the optimization of different algorithm experiments are compared,and their respective application ranges in the field of satellite telemetry parameter data prediction analysis are analyzed.Among them,the polynomial fitting extrapolation algorithm is suitable for telemetry parameters that are relatively slow with time,and can be applied to the short-term prediction of telemetering parameters.The autoregressive moving average algorithm is suitable for processing telemetering parameters with obvious periodic changes,which is suitable for medium-term prediction;The LSTM model has relatively high prediction accuracy,but it also has high requirements on data quality and long training time,which is suitable for medium-and long-term prediction tasks of satellite telemetry parameter data with low real-time requirements.The experimental verification shows that the satellite payload anomaly detection method based on time series model presented in this paper has good practical application value.Through the model tuning of the specific telemetric parameter data set,it can well process the satellite telemetry parameter data.The problem of forecasting trends further effectively explores the relationship between satellite telemetering data and historical anomaly information,providing a certain degree of support and aiding decision-making functions for the space science satellites in a healthy and effective orbital operation.