| Résumé | Proactive diagnosis of spacecraft issues and response to conceivable hazards has attracted considerable interest. Hidden anomalies in satellites can cause overall system degradation. In multivariate time series anomaly detection (AD), several sensors report to a fusion center at each timestamp. Each of these sensors measures a specific feature of the satellites. In this paper, we first leverage the eigenvalues of covariance metrics of a multivariate time series to determine the anomaly score at each timestamp. Then, the anomaly scores are fed to various machine learning models. After that, we employ a Multivariate Variance-based Genetic Ensemble (MVVGE) learning method to ensemble the results of several models based on their corresponding performance. More specifically, we ensemble different Neural Network (NN), Random Forest (RF), and Linear Regression (LR) models based on their associated variance to catch the anomalies of three datasets, i.e., two satellite datasets: i) Soil Moisture Active Passive (SMAP), ii) Mars Science Laboratory rover (MSL), and Server Machine Dataset (SMD). Given that establishing the model confidence region is the bottleneck of our approach, we use an approximated version of the Bayesian NNs (BNN)s for acquiring confidence intervals of NN-based models. For RF and LR confidence intervals, we use an empirical method and bootstrapping, respectively. Simulation results confirm the superiority of our proposed approach compared to other methods. |
|---|
