Efficient detection of different fire scenarios or nuisance incidents using deep learning methods

Fire is a major disaster in buildings, consequences of which can be minimized or even prevented by appropriate measures. Traditional smoke detectors usually create an alarm without distinguishing between a fire or nuisance. Frequent false alarms result in unnecessary evacuations, costly fire-fighter responses, and waste of extinguishing agents. Early and accurate fire detection is crucial. Therefore, DL (deep learning) models are developed to distinguish smokes of cotton, wood, N-heptane, polyurethane foam, sunflower oil, cigarettes, printed circuit board (PCB), aerosols of paraffin, PAO (Polyalphaolefins), DEHS (Di-Ethyl-Hexyl-Sebacat), cement or plaster powders, fabric and mixtures of some of them. An existing non-complex high-sensitivity optical cell has been improved and adapted. DL models are trained using orientation-dependent scattering of light from particles at wavelengths of 405 and 980 nm. With paraffin aerosol, the highest smoke density at which the cell saturates is about 1.4 % obs/m. Five different DL models are created, trained with fire events with slowest fire growth rates and evaluated against unknown fire events with fastest fire growth rates. To consider the component tolerances of the amplifier electronics, Monte Carlo methods are performed, and the test data are manipulated accordingly. The F1 scores of the largest DL model for individual particle and fire event type discriminations surpass 88.0 % and 99.4 %, respectively, with augmented data. Due to its high sensitivity at low particle concentrations, transferring the methods from this study to smoke detectors in buildings can significantly decrease false alarms and enables precise localization of fire instances via source type prediction.