Research On Data-driven Spacecraft Telemetry Data Anomaly Detection Method

With the rapid development of economy,technology and production,the structural design of the spacecraft system is becoming more and more complex,and the functions are more comprehensive.It presents a large scale,comprehensive integration,and deep intelligence.With the increase of orbit spacecrafts,the incidence of failures is significantly improved.Telemetry data is the only way for ground management personnel or users to judge and understand the state of spacecraft.Therefore,using data-driven method to achieve high-efficiency,accurate and timely detection of spacecraft telemetry data anomaly has very important practical significance and application value.The paper carried out research on unit and multi-variation anomaly detection for spacecraft telemetry data.The main work and innovations are as follows:This paper proposes a LUBE interval prediction anomaly detection model based on multi-objective optimization.Firstly,based on the traditional LUBE model,an anomaly detection framework is proposed.In the framework of anomaly detection,a method to eliminate the error of the model itself is introduced,and the uncertainty of the model is characterized by an improved k-fold cross-validation method.Aiming at the multiobjective optimization based on the prediction interval evaluation(interval width and interval coverage)and abnormal detection index(detection rate and false detection rate)of the model,it is considered as a multi-objective optimization problem.The LUBE interval prediction model introduces the Knee point idea.Compared with the traditional method,the improved method performs better on the prediction effects and abnormality detection evaluation indexes on multiple data sets such as public data and telemetry data.This paper proposes a spacecraft battery anomaly detection method based on deep belief network(DBN)model.It learns the evolution of spacecraft battery health state by constructing a correlation neural network model of temperature,current,pressure,charge and discharge time and voltage.Regularity,when the battery health state is significantly degraded,the multivariate correlation of the model will be abnormal.Therefore,the model predicts the difference between the discharge voltage value and the actual observation value to determine whether the spacecraft battery health state is abnormal.Based on the actual telemetry data of a satellite battery in orbit for three years,it is verified that the DBN model proposed in this paper can describe the intrinsic correlation of the battery system more accurately,and give the magnitude of the abnormality of the space state of the spacecraft battery in time.The detection result is more traditional.The multivariate anomaly detection method is more reliable and effective,and the detection effect is more intuitive.This paper proposes a fusion dimension reduction reconstruction multivariate anomaly detection method,which extracts the global structure and local structure features of complex systems.By introducing the local structure maintenance idea of manifold algorithm,the traditional dimensionality reduction reconstruction method is improved only.Pay attention to the defects of the global structure,making the feature extraction of multivariate parameters more comprehensive.For the multivariate anomaly detection of complex systems such as attitude control,the proposed algorithm can mine the internal features of the data well,and the anomaly detection effect is better than the traditional PCA method.