Probabilistic Prediction Based Anomaly Detection Method For Spacecraft Telemetry Data

Due to the special space environment,insufficient design verification,the risk of production and processing technology,as well as damage accumulation effect,onorbit abnormal events occur frequently.The reliability and mission safety of spacecraft system are affected accordingly.As an essential basis for determining the operation status and working performance,the real-time detection of anomalous modes contained in the telemetry data can provide great benefits for higher monitoring ability of ground long-time management.It can also conducive to complete the flight mission and prolong the service life of spacecraft.Telemetry data have the characteristics of time series,pseudo-periodicity,and uncertainty,etc.Anomalous modes of telemetry data are temporally correlated and aggregated.Owing to the advantages of time series modeling,dynamic threshold generation,on-line anomaly detection and strong interpretability,the probabilistic prediction methods have been applied to anomaly detection for spacecraft telemetry data.However,in view of the anomaly detection for collective anomaly modes with temporal property in univariate telemetry parameter and the anomaly detection for multivariate parameters with complex correlations,this kind of method still faces some challenges.To be specific,the difficulty in adaptive selection of prediction confidence level,the inadequate performance of anomaly detection for univariate telemetry parameter and multivariate telemetry parameters are three main challenges in this domain.Therefore,this dissertation focuses on the issues mentioned above and the main contributions are as follows.(1)The selection of prediction confidence level in probability-based anomaly detection method generally depends on conventional empirical value.To address this issue,an improved selection method of prediction confidence level under unsupervised conditions is studied.First,the Unsupervised Receiver Operating Characteristic(Un ROC)curve is designed to reveal the correlation between the potential detection rate and false positive rate of the prediction intervals under different prediction confidence levels.Then,an optimization model for prediction confidence level is constructed with the proposed Enhanced Youden(E-Youden)index as its objective function.This model is utilized to obtain the optimal prediction confidence level of the Un ROC curve to generate effective dynamic thresholds.Compared with other selection methods of prediction confidence level,experimental results show that the proposed method can achieve better prediction confidence level.The corresponding anomaly detection thresholds have better discrimination performance of positive and negative samples and strong prediction ability of sample labels.It provides an effective solution to decrease the influence of unreasonable selection about prediction confidence level on the sample classification performance of the anomaly detection thresholds.(2)To promote probabilistic prediction-based anomaly detection method in anomaly quantitative representation of univariate parameter and multi-step prediction feature fusion modeling,a univariate parameter anomaly detection method based on discrete feature construction and fusion is studied.Firstly,combined with the effective dynamic thresholds provided by the modified probabilistic prediction model with better prediction confidence level,a discretization method based on equal-width discretization and statistical analysis is proposed.Each prediction error is divided into several discretization intervals corresponding to different anomalies.In this way,the quantitative characterization ability of anomalies within univariate parameter is enhanced.Then,based on Markov chain and majority voting integration method,discrete feature fusion at multi-time scales referring to multi-step and multi-window is implemented to realize the time series modeling of multi-step prediction features.Compared with other collective anomaly detection methods for the univariate parameter,experimental results show that the proposed method has better performance on anomaly detection and detecting stability for collective anomaly modes under different manifestations and different durations.It provides an effective solution for improving the comprehensive performance and application ability of probabilstic prediction-based anomaly detection for univariate telemetry parameter.(3)To promote probabilistic prediction-based anomaly detection method in prediction performance of high-dimensional input space and multi-feature fusion modeling,an anomaly detection method for multivariate parameters based on spatiotemporal feature extraction and fusion is proposed.Firstly,the spatial features of multivariate parameters are extracted based on factor analysis method,which reduces the input spatial dimensions and reveals the internal structure relationship of multivariate parameters.In this way,the probability prediction models for these spatial features can be formulated which have better prediction confidence levels.Based on these probability prediction models,spatio-temporal feature of multivariate can be extracted.Then,anomaly descriptors for multivariate parameters are formed by discretizing the spatio-temporal features,and the association modeling of multifeature is realized based on feature crosses with multi-dimensional Markov chains.Compared with other anomaly detection methods for multivariate parameters,experimental results show that the proposed method has better capability of anomaly detection and interpretation for multivariate telemetry parameters with complex correlation.It provides an effective solution for improving the adaptability and applicability of the probabilistic prediction method for the anomaly detection of multivariate telemetry parameters.