A Low-Cost Multimodal Detector for Emergency Signals and Closed-Loop Search Guidance: A Prototype Study Using Controlled RF, Audio, and Vibration Tests

Abstract

Rapid survivor search after structural collapse, flooding, avalanche burial, or other disasters is often limited because trapped people may produce only weak, irregular, and non-coded cues such as tapping, whistles, short voice calls, or incidental electromagnetic pulses. Existing rescue technologies are valuable, but many depend on certified distress beacons, specialized radar systems, trained teams, or infrastructure that may be unavailable in educational settings or early volunteer response. This study examined whether a low-cost, explainable, student-built prototype could support a bounded search workflow by combining a long-range RF cue channel with close-range audio and vibration confirmation.

The prototype integrated a low-frequency RF receive channel centered near 457 kHz with microphone and piezo-vibration channels. Its firmware used three interpretable criteria—energy, band consistency, and rhythm—and applied channel-specific voting rules through a simple operator interface. Evaluation used controlled educational test scenarios: a low-power laboratory reference source near 457 kHz for line-of-sight baseline trials, structured tapping and whistle/voice tests for close-range confirmation, and a supervised impulsive broadband source coupled to metal structures only for classifier stress testing. Construction and operating details for any signal source that could be misused are deliberately omitted. A software-defined radio receiver was used only for passive spectral observation, including passive monitoring near 121.5 MHz, and did not transmit or participate in detection decisions.

The baseline RF test achieved 58 detections in 60 trials at 100 m line of sight, or 96.7% detection. During one hour of silent monitoring, no false alarms were observed, corresponding to an approximate upper 95% false-alarm bound of 3 events per hour. In structured confirmation tests, knock detection ranged from 60.0% to 95.0% depending on the contacted material, and whistle/voice confirmation was strongest at 1–5 m but declined at longer corridor distances. In classifier-ablation trials, the full three-criterion method achieved 90.0% detection at the shorter building path scale and 70.0% at the longer path scale, both with zero false alarms in those runs. In a small pilot search drill, brief training reduced mean time to first detection from 6.83 ± 1.82 min to 4.20 ± 0.46 min and reduced mean path length from 156 ± 31 m to 104.7 ± 7.5 m.

These results support the feasibility of a low-cost, explainable multimodal rescue-assistance prototype for education, drills, and supervised research. The system is not a certified rescue device; its contribution is a reproducible receiver-and-classifier framework linking coarse cue capture, near-field confirmation, and operator guidance.