DP-MSFF-Net: Dual-Path Multi-Scale Spatiotemporal Feature Fusion for Satellite Telemetry Anomaly Detection

Satellite telemetry exhibits complex temporal dynamics and strong cross-parameter dependencies, complicating both reconstruction and anomaly detection. Conventional methods may inadequately capture channel-specific temporal behavior and cross-parameter dependencies, causing missed on-orbit anomalies, excessive false alarms, and limited generalizability under complex operating conditions. To this end, we propose an anomaly detection framework for multivariate satellite telemetry time-series data based on the Dual-Path Multi-Scale Spatiotemporal Feature Fusion Network (DP-MSFF-Net). The framework organizes telemetry channels through prior-guided grouping and represents each group as a structured spatiotemporal input. Its dual-path encoder combines deep convolutional and convolutional long short-term memory (ConvLSTM) representations through multi-level fusion and gated multi-scale skip connections to capture cross-parameter structure and temporal evolution. Channel-specific anomaly decisions are then derived from group-wise reconstruction using adaptive residual calibration, nonparametric thresholding (NPT), and sequence pruning. The mean Event-wise F0.5 values were 0.134 and 0.326 on Missions 1 and 2, respectively, compared with the highest baseline values of 0.061 and 0.079. The corresponding mean Channel-aware F0.5 values were 0.475 and 0.813 (best baselines: 0.477 and 0.791), while mean Alarming Precision reached 0.411 and 0.818 (best baselines: 0.206 and 0.790).