Runtime Subnet-Sensing Adaptation and Configuration-Aware Cloud Calibration for Resource-Constrained IoT Edges
2026.11~
On-device intelligence enables battery-powered Internet of Things sensors to detect abnormal conditions without continuously transmitting raw measurements. Deploying such intelligence typically involves training a task-specific model and optimizing it for the target device's computational and memory constraints. Once-for-all networks extend this approach by enabling hardware-compatible SubNet selection from a shared SuperNet, while computation offloading provides access to remote processing at the cost of communication energy and network-dependent latency. However, a deployment-time configuration may not remain energy-efficient as signal characteristics, available resources, and the remaining energy budget change during operation. This paper proposes a runtime adaptation framework that jointly controls sensing and subnet execution to extend sensor lifetime under detection reliability and response-time constraints. A lightweight on-device controller selects the sampling rate and active subnet according to signal statistics, recent prediction uncertainty, and the energy budget required for the remaining mission duration. The framework allocates sensing and computational effort to intervals requiring closer inspection while reducing expenditure during stable operation. Each reported prediction is accompanied by a compact configuration identifier, enabling the cloud to apply configuration-aware confidence calibration rather than interpret outputs from different operating modes identically. Cloud processing supports risk-aware interpretation without serving as a prerequisite for local inference or adaptation. The resulting design couples energy-aware on-device adaptation with configuration-aware cloud interpretation, targeting longer unattended operation without relying on routine raw-data offloading.Â