Lightweight Attention Networks for Onboard Satellite Anomaly Detection Under Memory Constraints

Abstract

The rapid expansion of low-Earth orbit satellite constellations has created urgent demand for autonomous onboard fault detection systems capable of operating within the severe memory and computational budgets imposed by space-grade embedded processors. This paper proposes a Lightweight Attention Network (LAN) architecture that integrates depthwise separable convolution (DSC) layers with a squeeze-and-excitation (SE) channel attention module to enable real-time anomaly detection in multivariate satellite telemetry streams under a quantized model footprint of 113 KB. Experiments on the NASA SMAP and MSL spacecraft telemetry benchmarks demonstrate that LAN achieves an F1 score of 0.891, outperforming LSTM-based baselines by 6.3 percentage points while reducing inference memory consumption by 78%. Deployment simulation on a Cortex-M7 processor confirms real-time feasibility at 14.2 ms per inference window. These results establish LAN as a practical solution for next-generation autonomous satellite health monitoring under strict resource constraints.