Abstract

Maintaining an effective facial seal is critical for the performance of tight-fitting industrial respirators used in high-risk sectors such as mining, manufacturing, and construction. Traditional fit verification methods—Qualitative Fit Testing (QLFT) and Quantitative Fit Testing (QNFT)—are limited to periodic assessments and cannot detect fit degradation during active use. This study presents a real-time fit detection system based on embedded breath sensors and machine learning algorithms. A compact sensor module inside the respirator continuously measures pressure, temperature, and humidity, transmitting data via Bluetooth Low Energy (BLE) to a smartphone for on-device inference. This system functions as a multimodal biosensor: intra-mask pressure tracks flow-driven mechanical dynamics, while temperature and humidity capture the thermal–hygrometric signature of exhaled breath. Their cycle-synchronous patterns provide an indirect yet reliable readout of respirator–face sealing in real time. Data were collected from 20 healthy volunteers under fit and misfit conditions using OSHA-standardized procedures, generating over 10,000 labeled breathing cycles. Statistical features extracted from segmented signals were used to train Random Forest, Support Vector Machine (SVM), and XGBoost classifiers. Model development and validation were conducted using variable-size sliding windows depending on the person’s breathing cycles, k-fold cross-validation, and leave-one-subject-out (LOSO) evaluation. The best-performing models achieved F1 scores approaching or exceeding 95%. This approach enables continuous, non-invasive fit monitoring and real-time alerts during work shifts. Unlike conventional techniques, the system relies on internal physiological signals rather than external particle measurements, providing a scalable, cost-effective, and field-deployable solution to enhance occupational safety and regulatory compliance. © 2025 by the authors.