Abstract:With the development of space technology, satellite telemetry parameters show high-dimensional, coupled, and nonlinear characteristics. Traditional statistical methods can hardly meet the needs of fine-grained health monitoring. Current deep-learning methods for telemetry anomaly detection mainly focus on univariate time-series modeling. They ignore the coupling among parameters and physical consistency constraints. To solve this problem, this paper proposes a Physics-Consistency Multi-Relation Temporal Convolutional Network (PC-MR-TCN). The model uses a Temporal Convolutional Network (TCN) as the backbone. It introduces a relation encoder to model temporal dependencies and inter-variable dependencies. The encoder captures the multi-relation coupling among telemetry parameters. A gating mechanism is used to fuse these representations. In addition, the model constructs weak physical-constraint loss terms, such as communication-link consistency and current-conservation consistency. These terms inject domain knowledge into the training process in a differentiable way. In the detection stage, a comprehensive anomaly score is constructed based on reconstruction error, prediction error, and physical consistency deviation. The proposed method achieves better scores on both a certain satellite platform and public datasets, providing methodological support for on-orbit safety and health detection.
