Anomaly Detection Method For Sequential Telemetry Data

Anomaly detection plays an important role in Prognostics and System Health Management.The effective and timely discovery of abnormal states for in-orbit satellites is an important task to monitor performance of satellite,detect faults,identify the root cause of the fault,and increase safety of satellite.In this thesis,based on several existing data mining and machine learning methods,we analyzed the temporal telemetry data of a specific satellite during 2014 to 2015 and studied the feature extraction and anomaly detection algorithms which are used to discover potential anomalies within temporal telemetry data.Due to the complex physical structure and rugged working environment of the satellite,the telemetry data collected by sensor is very big and has high dimensions.Meanwhile,it also shows non-stationary and nonlinear characteristics.The characteristics of abnormal data are not easy to extract.In order to solve these problems,a feature extraction algorithm based on empirical mode decomposition(EMD)and sample entropy(SE)is proposed.EMD is firstly employed to decompose the variation of each parameter.Then,the sample entropy of trend item for each parameter is calculated,which is considered as feature data.The experimental results show that the presented method could not only reduce the dimension of telemetry data,but also deal with nonlinear and non-stationary problems of temporal telemetry data.Thereupon,it could effectively extract abnormal feature of temporal telemetry data and reflect the state changes of it.Real-time anomaly detection algorithm often performs poor,when it is difficult to determine the threshold of anomaly detection and the reliability of abnormal detection could not be guaranteed.In order to solve this problem and improve the reliability of abnormal detection,an anomaly detection algorithm of single parameter based on short-time prediction and dynamic threshold is proposed.This algorithm firstly optimizes the parameters of kernel extreme learning machine(KELM)by means of differential evolution(DE)and then utilizes the optimized KELM to short-time prediction of single parameter dynamic threshold.Afterwards,the proportional coefficient method(PCM)optimized by DE is utilized to construct dynamic threshold interval.Finally,in order to improve the stability of the model and enhance the accuracy,the integration strategy is used to construct DE-PCM-OKELM model.This algorithm could predict the dynamic threshold interval in a short time,further identify the normal data and abnormal data,and thereby realize the anomaly detection of telemetry data.The experimental results show that the model has better accuracy of anomaly detection,it could explicitly describe the variation range of telemetry parameter,and identify the occurrence of telemetry parameter anomalies in time.Since the anomaly detection algorithm of single parameter is not suitable for the situation of the large scale and high dimension of telemetry data.Meanwhile,these methods could not accurately detect the occurrence of anomalies of multi-parameter.In order to deal with this situation,an anomaly detection method of multi-parameter based on angle deviation is proposed.It firstly applies Shared Nearest Neighbors(SNN)algorithm to construct reference point sets in high-dimensional data space.Then it selects the feature attributes associated with the anomaly by employing angle deviation attribute selected method based on angle replacing distance.Next,the normalized Mahalanobis distance is used to calculate the anomaly scores of points.Finally,combining with the statistical knowledge,calculate the threshold of anomaly scores and classify the data sets.The experimental results indicate that the accuracy of the proposed algorithm could reach more than 95% under the condition of lack of the field knowledge.Simultaneously,the robustness of it is higher and it can detect the anomaly of satellite subsystem timely and effectively.