Artificial Intelligence and Electrical & Electronics Engineering: AIEEE Open Access
Joint Contactless Temperature, Humidity, and Occupancy Sensing via Wi-Fi Channel State Information on ESP32 Nodes
Abstract
Saurav Chaudhari, Ketan Pise, Dinesh Fukate and Shantanu Gawande
Smart buildings increasingly rely on dense instrumentation to monitor indoor temperature, humidity, and occupancy for energy-efficient hvac control, yet conventional sensor deployments incur significant hardware, wiring, and maintenance costs. This paper proposes a joint contactless sensing framework that estimates ambient temperature, relative humidity, and occupancy state from wi-fi channel state information (csi) using low-cost esp32-wroom-32 nodes. Building on refractive-index models of microwave propagation, we extend the gladstone–dale relation to incorporate water-vapor pressure and derive sensitivity expressions that link multi-subcarrier csi amplitude and phase to both temperature and humidity. Because the direct refractive-index phase term is far below the phase-noise floor of commodity esp32 hardware, we explicitly identify and model the secondary mechanisms—multipath-interference amplitude modulation and temperature-dependent transceiver effects—that make the environmental signal learnable, and we exploit human- induced multipath and doppler perturbations for occupancy inference. A multi-task learning pipeline is developed in two stages: (1) hybrid csi feature extraction using statistical descriptors and discrete wavelet transform coefficients across 30 ofdm subcarriers, and (2) a shared gradient-boosting representation feeding three task-specific heads for temperature regression, humidity regression, and occupancy classification. All evaluation uses a strictly temporal (leakage-controlled) protocol. Experiments with esp32-wroom-32 devices in a climate chamber (15–35 �?�c, 30–80% rh) and three real-world indoor environments (laboratory, office, residential), each monitored for 21 days, achieve mean absolute error (mae) of 0.72 ± 0.05 �?�c for temperature, 4.6 ± 0.4% for relative humidity, and 95.1% F1-score for binary occupancy detection, with 47 ms on-device inference latency. Relative to single-task baselines, multi-task learning improves temperature MAE by 13% and humidity MAE by 18% (p < 0.01). A 21-day-per-site deployment shows stable operation under routine activity and HVAC cycles with drift-controlled calibration. We report a full bill of materials, power budget, and node- placement guidance, and we state plainly that cross-building generalization remains an open limitation. The system enables cost-effective, maintenance-light ambient and occupancy-aware sensing using existing Wi-Fi infrastructure.