Edge AI for Real-Time Physiological State Detection: A Dual-Use Architecture for Enhancing Combatant Resilience

Contemporary military operations require rapid, decentralised decision-making based on reliable combatant telemetry, yet cloud-dependent monitoring systems remain vulnerable to electronic warfare and spectrum denial. This paper presents ArmyAI, a proof-of-concept dual-use Edge AI architecture for offline multi-label physiological state detection on commercial off-the-shelf (COTS) microcontrollers. The system is built around a low-power neural processing platform and encodes sensor telemetry as a compact image representation processed by a multi-label Convolutional Neural Network with independent per-state outputs, trained on a combined synthetic and small pilot dataset. The model achieves strong multi-label classification performance on a held-out test set, and its quantised form is compact enough to reside entirely within the constrained flash and activation memory of an embedded NPU. The intended deployment is a local-area network of wearables linked by a low-power LoRa LAN, with planned reporting to an exercise-control team during training events; both are design-stage and have not been implemented or measured on hardware in the present prototype. We discuss the architecture, methodology, results, limitations, and the alignment of the approach with European dual-use funding mechanisms.