Satellite Anomaly Detection Based On Unsupervised Algorithm

With the increasing number of satellites in orbit and the more complex and diverse tasks they undertake,satellite anomaly detection is more important for the reliability and safety of on-orbit satellites.Traditional anomaly detection methods based on artificial interpretation,threshold and expert knowledge can not cope with the increasingly heavy all-weather,high-dimensional and large amount of data,so it is necessary to develop automatic anomaly detection algorithm.This paper proposes an anomaly detection algorithm based on unsupervised algorithm.The main research work is as follows:1.This paper analyzes the characteristics of satellite telemetry data and preprocesses it.In addition,different evaluation indexes are set according to different satellite anomaly detection task.Data preprocessing includes data cleaning,normalization and feature engineering.Because the abnormal sample is far less than the normal sample,the accuracy can not accurately reflect the performance of the anomaly detection method,while the receiver operating characteristic curve,area value under the curve,recall and precision can better reflect its performance.2.Aiming at the multi-variable point anomaly,this paper proposes a method based on the improved isolated forest.The traditional algorithm of isolation forest has too much randomness,which affects the accuracy of the model.The improved isolation forest proposed in this paper adopts the deterministic attribute value and introduces the heuristic strategy of selecting the root node of isolation tree.The comparative experiments on the telemetry data of multiple subsystems show that the method improves the accuracy,time efficiency and stability.3.A method based on deep learning is proposed for power subsystem.Satellite power subsystem has the characteristics of periodicity,autoregression and complex time sequence.In this paper,a dynamic threshold anomaly detection algorithm based on Long Short-Term Memory is proposed.The experimental results show that this method has a strong ability to fit the telemetry variables of the power subsystem,and the dynamic threshold method can detect the sequential anomalies and point anomalies in time.