Research On On-orbit Anomaly Detection Method Based On Causal Inference For Spacecraft

Because a spacecraft operates in a complex space environment for a long time,it is inevitable that performance degradation or functional failure will occur,threatening the safety of the spacecraft on orbit,and even affecting the service life of the spacecraft.A large amount of telemetry data is monitored during the spacecraft’s orbit operation.These data,which record the temperature,power,current,voltage and other information of the spacecraft,are an important basis for analyzing the operation status of the spacecraft and detecting abnormal conditions.Based on the spacecraft in-orbit telemetry data,this paper studies the data-driven spacecraft in-orbit anomaly detection method,and conducts a detailed and systematic study from the perspective of causal inference.The main research content includes the following three aspects.(1)Aiming at the problems of numerous spacecraft telemetry parameters and complex correlations between parameters,an improved multivariate transfer entropy method is proposed to infer the causality relationship between high-dimensional telemetry parameters.This method is based on the principle of information decomposition,using multivariate transfer entropy to preliminarily determine the cause parameters(called interpretation parameters)that affect the changes of specific telemetry parameters,and then use the mean statistical test method to eliminate redundant causality.Examples show that,compared with other methods,the causal relationship obtained by the improved multivariate transfer entropy method is more in line with the actual situation.(2)In view of the difficulty in determining the detection threshold and the difficulty in identifying false anomalies in the current data-driven spacecraft on-orbit anomaly detection,this paper proposes a threshold determination and correction method based on the false alarm rate;proposes a threshold determination and correction algorithm,called EPOT(Error Peaks-Over-Threshold);and proposes a method of identifying and pruning false anomalies based on causality.Case analysis shows that the method based on false alarm rate has wider applicability,but the EPOT method has better effect on non-stationary telemetry parameters.Combining the threshold determination method in this paper with the correction method and the causality-based false anomaly identification method can significantly improve the performance of the state detection algorithm.(3)Aiming at the problem of data-driven spacecraft on-orbit anomaly detection,it is difficult to establish a model with high-precision prediction and good anomaly sensitivity for multivariate telemetry parameters.This paper proposes a CF-LSTM(Causality Features-Long Short-Term Memory)model for spacecraft on-orbit state prediction,and an on-orbit anomaly detection method based on CF-LSTM.Among them,the CF-LSTM model of each telemetry parameter is established based on the identified causality,and the historical information of the telemetry parameter itself is combined with the information of the interpretation parameter to improve the prediction accuracy and has good anomalous sensitivity.The CF-LSTM-based on-orbit state detection method uses historical telemetry data to establish a normal state CF-LSTM model,adopts corresponding threshold determination and correction strategies for different parameter characteristics,and recognizes false anomalies based on causality.Case analysis shows that the anomaly detection method based on CF-LSTM not only significantly improves the detection accuracy,recall and F1-score,etc.,but also is more sensitive to anomalous state and more robust to false anomalous state,indicating the effectiveness of the method.At the end of the paper,we summarized the research content and put forward the content that needs further research.