Kernel Self-Adaptive Learning-Based Satellite Telemetry Data Classification

Abstract

Telemetry data, containing the data of multiple subsystems such as power system, implies the on-orbit operation status information of the satellite. We can obtain performance characteristics and fault symptom of the satellite subsystems through analyzing these data. Using classification algorithm we can provide normal data for anomaly detection and find the data from various subsystems which have the same category labels. Then we will get potential knowledge of the satellite and rich expert experience. Because of the importance to national defense and people's livelihood fields, we have higher requirements about the accuracy and stability of satellite data processing. We proposed a kernel self-adaptive method based on Fisher Discriminant Analysis (FDA). Firstly map the data to high-dimensional space by kernel function. Then perform PCA in high-dimensional space. Finally classify data by Fisher Discriminant Analysis method. The selection of the kernel function and its parameters realized by self-adaptive method are based on the data. This method can achieve ideal classification accuracy toward satellite telemetry data according to our experiments. And the accuracy is stable when process data of same type. The performance of this method proved that it is reliable.