Anomaly Discovery Method For Satellite Power Telemetry Data

The satellite power system is an important part of the satellite.It is of great significance to detect anomalies through the telemetry data of the satellite power system.In this paper,we take the telemetry data of a satellite power system as the research object and analyze the advantages and disadvantages of the existing machine learning and data mining algorithms.We design and implement satellite power telemetry data missing value filling,anomaly detection and anomaly prediction algorithms.For timely and effective detection of anomalies,we carry out the following research.There exists missing values and long consecutive missing time in the telemetry data.To solve this problem,a missing value filling method based on gradient boosting regression tree is proposed.According to the principle of maximizing correlation and minimum redundancy,this method uses genetic algorithm to select features from the perspective of information entropy,firstly;Then,the selected features are used as training data,and the feature to be filled with missing values are used as the sample label;Finally,the missing values are filled with the predicted values of the trained gradient boosting regression tree.Compared with the common missing value filling algorithms,this method can fill the missing values with less error and closer to the real values.For the high-dimensional periodic time series telemetry data of satellite power,a novel representative feature auto encoder RFAE model is proposed and used for unsupervised anomaly detection.RFAE improves the loss function and training algorithm,so that the model can learn the representative features of samples with the same phase;Then RFAE reconstructs the samples based on the representative features and detects whether the samples are abnormal based on the reconstruction error.In the experimental part,it was first verified through synthetic data that the RFAE algorithm can effectively detect anomalies in high-dimensional periodic time series data,and then experiments were performed using real telemetry data of a satellite power system from June to October 2014.The accuracy rate of RFAE anomaly detection reached 99%,the detection result is significantly better than other current anomaly detection algorithms,and has higher application value.For the disadvantages of passivity and lag in anomaly detection,an anomaly prediction method MTS-CNN is proposed to predict the anomaly of satellite power system.MTS-CNN is based on convolutional neural network.It treats the multivariate time series data as a single channel image over a period of time.The input of MTS-CNN is the multivariate time series data in different time windows.MTS-CNN extracts time correlation features and variable correlation features.These features are fused by the parametric feature fusion method.Finally,MTS-CNN classifies the data at a certain moment into an alert state or a non-alert state according to the merged features,which achieves the purpose of anomaly prediction.In the experimental part,it is verified by the public datasets that the MTS-CNN can effectively predict the anomalies of multivariate time series,firstly.Then we use the telemetry data of a satellite power system to conduct experiments.The anomaly prediction result is also better than other algorithms,which has a high application value.