Presence-Gated VOC Sensing for Urban Search and Rescue Applications

Description

This dataset contains development-phase logs used to train and evaluate J48, Random Forest, and a simple probability-level hybrid classifier for a portable volatile-signature detector intended for urban search-and-rescue triage. Data were collected in tethered mode while the ESP32 firmware executed repeating intake–hold–purge cycles. During each hold window the device computed a per-hold feature vector and mirrored it to a host PC along with synchronized raw/derived channels and metadata.

Steps to reproduce

-Place the device on a bench, connect USB, and start the logging utility (1 Hz dashboard + per-hold features). -Set cycle timings (e.g., intake 2 s, hold 3 s, purge 3 s) and confirm fan setpoint. -Record fresh air for ≥10 cycles with no source present. -For Ammonia, Propane_Gas, and Injured, present the source in front of the inlet during intake, remove before hold, and capture 10–15 cycles per session (cap sources between cycles; add extra purge if residual odor persists). -Save raw, feature, and meta files for each session; document distance to inlet and any anomalies. -Use the provided folds to train J48, Random Forest, and the hybrid probability-fusion method; select operating thresholds with Youden’s J on training folds; report confusion matrices, ROC/PR curves, and macrometrics on held-out folds. -Apply the radar presence flag as a decision gate only at inference if used.

Institutions

• Huaiyin Institute of Technology

  Jiangsu, Huaian

Categories

Licence