Date of Award

Spring 5-2026

Language

English

Document Type

Honors Thesis

Degree Name

Bachelor of Science

Department

Electrical Engineering

Advisor/Committee Chair

Dr. Hany Elgala

Committee Member

Dr. Gary Saulnier

Abstract

The proliferation of wireless technologies has produced a dense, increasingly heterogeneous radio-frequency (RF) environment that surrounds modern life. While regulatory bodies such as the Federal Communications Commission (FCC) and the World Health Organization (WHO) maintain exposure guidelines, the tools that ordinary people, educators, and small institutions can use to actually observe RF exposure in their own spaces remain expensive, specialized, and largely confined to professional laboratories. Spectrum analyzers and calibrated field probes routinely cost between five hundred and one hundred thousand dollars per unit, placing continuous local monitoring out of reach for most environments where exposure questions are most natural to ask. This thesis presents the design, implementation, and validation of a low-cost, distributed RF exposure monitoring system that performs anomaly classification entirely on-device using TinyML. Each sensor node combines a dipole antenna, an RF detector integrated circuit, and an Arduino MKR WAN 1310 microcontroller, with a per-node hardware cost of $109.22. Sensor nodes sample power in the 730–766 MHz cellular band (LTE Bands 12, 13, and 14), compute four engineered statistical features from each one-second window, and pass those features through a frozen decision tree classifier compiled directly into the firmware. Each node transmits its classification — LOW, MEDIUM, or HIGH — along with the underlying features over LoRa to a Raspberry Pi gateway, which renders a real-time dashboard showing live signal traces, per-node exposure status, and a room-level RF heatmap. The decision tree was trained on 14,600 labeled samples collected across three real-world environments and achieved 98% accuracy with macro-averaged precision, recall, and F1 of 0.98 on a held-out 80/20 stratified split. A controlled Faraday cage experiment, conducted in collaboration with the SINE Lab at the University at Albany, characterized the sensitivity boundary of the low-cost detector by modeling its detection probability as a Gaussian centered at −55 dBm with a standard deviation of 8 dBm. All six top-level system requirements were met or exceeded, and the platform demonstrates that meaningful, real-time RF exposure awareness can be delivered at hobbyist-level cost without sacrificing scientific rigor. The work contributes a reproducible reference architecture for RF monitoring and an open template for embedded ML deployment in resource-constrained sensing applications.

Download

Share

COinS