距离相关系数融合GPR模型的卫星异常检测方法

卫星在轨运行期间，遥测数据表现形式通常为多维时间序列。高斯过程回归(GPR)模型可以为重要的遥测参数提供动态门限，及时发现隐藏在工程阈值内的故障征兆，但是高维卫星数据使得GPR模型具有局限性。因此，为获取与多个遥测参数相关的动态门限，在GPR模型的基础上，融合距离相关系数对预测变量进行选择，减少信息冗余和计算量，提高模型的可解释性，并估计模型的泛化误差以设置更合理的预测区间，提高模型的泛化能力，检测数据流的持续异常。对实际在轨卫星数据进行仿真实验，验证了距离相关系数融合GPR模型的卫星异常检测方法可以在卫星故障早期检测到数据异常，而且提高了模型的预测性能，降低了虚警率。 

A satellite anomaly detection method based on distance correlation coefficient and GPR model

During the orbital operation of the satellite, the telemetry data is usually represented by multidimensional time series. The Gaussian Process Regression (GPR) model can provide dynamic thresholds for important telemetry parameters and timely discover failure symptoms hidden within the engineering threshold. However, high dimensional satellite data makes GPR model limited. Therefore, in order to obtain the dynamic threshold related to multiple telemetry parameters, based on the GPR model, the distance correlation coefficient is combined to select predictive variables, reduce the information redundancy and the amount of calculation, and improve the interpretability of the model.The generalization error of the model is estimated to set a more reasonable prediction interval, to improve the generalization ability and detect the continuous abnormality of the data stream. Simulation experiments on actual orbiting satellite data verify that this method can detect data anomalies in the early failure of the satellite, improve the prediction performance of the model and reduce the false alarm rate.