Satellite Engineering Parameters Anomaly Detection Method Integrating Deep Learning And Transfer Learning

With the prosperous development of National Space Science Pioneer Project,a series of scientific experiment satellites have been successfully launched.It is significant to ensure safe and efficient operation of satellites in orbit and complete corresponding scientific experiment tasks,which is a major challenge in the field of space science.The research of scientific experiment satellites requires a large amount of human,material and financial resources investment.If a satellite fails in orbit,a sequence of consequences is immeasurable.Therefore,it is important to detect healthy state of the satellite.Satellite anomaly detection is aiming to establish some efficient and effective methods to monitor health status on satellites in orbit,so as to detect the appearances of abnormal status on satellites and on-board loads in time.So that relevant personnel can respond quickly and avoid the occurrence of satellite operation safety accidents.In recent years,the continuous development in the field of machine learning and the improvement of computer computing power have provided reliable algorithm support and hardware support for anomaly detection in the satellite field.Meanwhile,the huge volume of telemetry data generated by a number of satellites acts as a reliable data supplier.As a result,anomaly detection in satellite field based on machine learning methods produces practical application meaning.Satellite engineering parameters are direct evidence of whether the satellite is normal or not.In order to obtain an optimal anomaly detection method,based on the engineering parameters from Micius Quantum Experiment Science Satellite------MQESS and “Taiji-1” satellite,two experiment contents have been determined under the full understanding of satellite design characteristics and data itself.Aiming at those engineering parameters from a single satellite,traditional machine learning methods are chosen to detect the abnormal incidents.During this detection process,the simulated dataset with complete anomaly data types is constructed,which will provide strong data foundation for the subsequent research work.As for those engineering parameters both on MQESS and “Taiji-1” satellite,an algorithm integrating deep learning and transfer learning is proposed to detect anomalies.Taking advantage of the simulated dataset constructed in early research task,a method based on domain adaptation is used to extract features in the simulated data,completing the cross-platform satellite engineering parameters anomaly detection research.This paper mainly includes the following aspects.(1)An optimal anomaly detection algorithm on single satellite via performace comparison.Firstly,according to the organization structure of MQESS and the experience of abnormal engineering parameters types occurred in the past,a simulated dataset based on engineering parameters from MQESS was constructed,including three different types data samples.Then applying various traditional machine learning algorithms,such as decision tree,Gradient Boosting Decision Tree,Support Vector Machine,and K-nearest Neighbor,detected anomalies on MQESS dataset.Experiment results show that detection accuracy of all algorithms is higher than 96%.And GBDT performs best,with an average of 99%,showing excellent anomaly detection ability.In terms of detection efficiency,GBDT performs well,meeting reliable requirements of fast and efficient detection in satellite engineering parameters.(2)An algorithm integrating deep learning and transfer learning for anomaly detection on satellite engineering parameters.Conventional anomaly detection methods are designed in a single satellite data,achieving a poor detection accuracy on anomaly detection task in a new satellite.In order to solve the problem,a model integrating deep learning and transfer learning is proposed.It is mainly composed of two parts.Using residual shrinkage network extracts features,so as to obtain some implicit information behind the data.CORAL is added in the network to calculate the second-order statistical feature distance,evaluating similarity between the source domain features and target domain features.The experiment chose MQESS data established in previous work and“Taiji-1” satellite engineering parameters to verify effectiveness.Experiment results show that the accuracy can reach 98.03%.Compared with Gradient Boosting Decision Tree,Surpport Vector Machine and Deep Residual Shrinkage Network,detecting accuracy has been improved by 20.95%,22.51% and 15.02%,respectively.On the other hand,the detection efficiency is also improved because of the strength of GPU.Finally,feature extraction results are visualized.From the analysis,it is more obvious that methods integrating deep learning and transfer learning can effectively reduce the difference between source domain data and target domain data,realizing excellent abnormal detection effect on cross-satellite engineering parameters.