Prior work has applied deep-learning techniques to industrial control-system (ICS) data, showing effective detection of anomalous behavior. However, these methods often lack interpretability and are vulnerable to evasion attacks. To address these issues, we propose a novel approach that uses structurally sparse graph representations to learn relationships that are directly relevant to the physical and logical structure of the ICS. Our method requires fewer parameters, trains only on existing physical connections, and achieves detection performance comparable to state-of-the-art baselines while operating with low latency and a small model footprint. Additionally, it enables model attribution and improves robustness against evasive behaviors.