Augmented temporal grid for graph neural networks and anomaly detection in spacecraft

Abstract Graph neural networks (GNN) for spatiotemporal problems incorporate a temporal module operating on a temporal grid of interconnected subsequent time steps. In this work, we explore adding connections to future time steps to the temporal grid in order to extract unique features for the task of anomaly detection in time series. We introduce strategies for determining future connections utilizing the autocorrelation function and different random sampling techniques. The effectiveness of resulting graph neural networks is demonstrated on multiple anomaly detection benchmarks, including spacecraft telemetry datasets. A selection of graph neural network models is further deployed on the AMD-Xilinx Versal AI Core SoC to measure the execution time and resource utilization when processing telemetry data.