A multi-wavelength multi-task learning framework for risk-aware fire source classification and smoke density prediction.

Traditional smoke detectors suffer from high false alarm rates and lack the ability to differentiate fire sources or quantify smoke density, limiting their effectiveness in diverse indoor environments. While recent studies have explored single-task approaches for fire classification or smoke estimation, these methods typically require separate models, resulting in increased computational overhead. This study proposes a novel multi-task learning (MTL) framework that simultaneously performs fire source classification and smoke density prediction using multi-wavelength optical sensing at 460, 530, 660, and 940 nm. The system was validated under UL 268 standard fire scenarios, including smoldering fires (filter paper and wood), flaming fires (polyurethane foam), and cooking nuisance sources (hamburger patties). Among the evaluated architectures, the CNN-LSTM model achieved the best performance, with a classification accuracy of 97% and a smoke density prediction error of 0.668 MSE. Compared to single-task learning approaches, the MTL framework reduced memory usage by 45% and achieved inference times 82% faster, demonstrating its suitability for real-time edge-based fire detection systems. Ablation studies further revealed that the 940 nm near-infrared wavelength provides discriminative features for distinguishing fire smoke from representative nuisance aerosols. Overall, the proposed approach enables reliable early detection with reduced false alarms in indoor environments.