Satellite Telemetry Anomaly Detection Based on Gradient Boosting Regression with Feature Selection

A data-driven satellite telemetry data anomaly detection method is proposed. The gradient boosting regression algorithm combined with feature selection, including feature scoring and recursive lowest-score feature elimination, can automatically mine the correlative telemetry variables through iterations and establish a nonlinear regression model for their functional association, which can be used as a health baseline for anomaly detection of telemetry data. This method requires no expert to specify correlative telemetry variables based on domain knowledge beforehand. It has the advantage of self-adaption for satellite operating conditions, which can overcome the problem of functional association altering under different operating conditions caused by orbit or sunshine condition changes. The validity and effectiveness of the method is verified by the telemetry data of the power subsystem.