Anomaly detection of condition monitoring with predicted uncertainty for aerospace applications

The condition monitoring plays an important role for the system reliability and safety of aerospace engineering. Especially, to detect anomaly with the monitoring data is the very primary and critical task for the system health management (SHM). Due to the high efficiency, easy usage, and predictive performance, the predicted model is applied to realize anomaly detection for monitoring data of complex system. An integrated prediction algorithm is proposed for the outlier detection to the univariate time series. The presented method uses the Least Square Support Vector Machine (LS-SVM) to achieve multi-step prediction. A confidence interval for the normal output range is obtained by estimating the prediction error as well as its probability distribution. As a result, the uncertainty representation can be achieved by improving the point estimator with the LS-SVM. By comparing the updating sample with the predicted by the proposed method, the anomaly detection can be conducted with high real-time performance with the prediction strategy. The experiment results with the open source data sets as well as the real satellite monitoring data sets show the effectiveness of this proposed method.