An adaptive statistical and clustering-based sensor sleep management framework for energy-efficient IoT networks

Abstract

Wireless IoT deployments spend most of their energy keeping every node awake even when their measurements are statistically redundant. This paper presents the Adaptive Sensor Sleep Management (ASSM) framework, which analyses temperature and humidity streams online to decide how many nodes need to stay awake. ASSM combines a sliding-window statistical engine, a K-means stage that selects spatially representative nodes, and a phase detector that separates stable from transitional periods using the derivative of the smoothed signal, with all thresholds estimated online from past samples only. During stable phases a small subset of nodes remains active; during transitions every node is woken. We evaluate ASSM on a realistic 100-node model built from an 86,401-record Solapur-mimicking dataset by adding per-node calibration offset, measurement noise and drift, and we benchmark it against random, round-robin, fixed-duty, variance-adaptive and prediction-based baselines with an ablation and a multi-metric fidelity analysis. Across eight random seeds ASSM keeps the network mean to about 0.19% and significantly outperforms random selection (p = 0.008), although stronger baselines—fixed-duty, variance-adaptive and dual-prediction—achieve lower mean error; ASSM’s value is that it is training-free, interpretable and keeps no per-node model. A full-link energy model including switching and gateway-compute overhead shows a net saving of about 55% at the operating point, remaining above ~ 46% under communication-overhead stress, and the controller scales near-linearly to 10,000 nodes. Rare local events depend on periodic-wake and threshold-override safeguards rather than the sleeping subset.