Adaptive LIF Spiking Neural Networks for Satellite Anomaly Detection

This work presents a spiking neural network (SNN) for anomaly detection in satellite telemetry. The network was trained and tested on ESA benchmark dataset, containing anonymized telemetry from two satellite missions spanning over 17 years of operation. The model was trained exclusively on nominal data, learning to reconstruct the expected signal based on the previous 40 minutes of telemetry history. Anomalies are detected during inference by comparing prediction to the actual telemetry signal. The architecture includes one input and output signal per channel (6 for Mission 1, 11 for Mission 2), and three hidden layers with 256, 128, and 64 neurons. An adaptive Leaky Integrate-and-Fire (adLIF) neuron model was employed to enhance temporal memory through an adaptive current mechanism. The network was implemented in snnTorch, with quantization effects modeled during inference to support future deployment on FPGA. The proposed model achieved an event-wise F0.5-score of 0.776 on Mission 1 and 0.971 on Mission 2, demonstrating the feasibility of applying SNNs to time-series anomaly detection tasks in resource-constrained space environments.