Fire smoke and interference sources identification using scattering spectrum

Conventional photoelectric smoke detectors frequently present challenges in complex environments owing to their reliance on single-angle and single-wavelength scattering signals, which can result in high false alarm rates and limited accuracy. To address these limitations, this study proposes an advanced method for distinguishing between fire smoke and interference sources by leveraging multi-angle and multi-wavelength scattering spectral analysis. The primary objective of this study is to enhance the reliability of detection, reduce the occurrence of false alarms, and improve generalization capability in real-world scenarios. These contributions will facilitate the development of next-generation intelligent fire detection systems.

The system incorporated a broadband LED light source with controlled illumination characteristics and an AS7341 multi-channel spectral sensor capable of simultaneous detection across eight discrete wavelength channels (415, 445, 480, 515, 555, 590, 630, and 680 nm). Scattering signals were systematically acquired at 10 discrete angles ranging from 40° to 85°, with an interval of 5°. Spectral data were synchronously captured across all channels at each angle, forming a comprehensive wavelength-versus-angle feature matrix. The experimental design encompassed five types of standard fires (smoldering cotton, wood, polyurethane, heptane flame, and paper fire) and five typical interference sources (water mist, cooking fumes, cement dust, loest dust, and A2 test dust). The model's capacity for generalization was subsequently assessed using a set of five non-standard fire categories. A range of machine learning algorithms were utilized to construct classification models, including random forest, XGBoost, support vector machine, k-nearest neighbors (KNN), and logistic regression. The feature selection and hyperparameter tuning steps were implemented to enhance the model's performance and interpretability.

Under controlled chamber conditions, the 50° scattering angle consistently yielded optimal performance across all tested scenarios. At this angle, the XGBoost model demonstrated an accuracy of 100% in discriminating standard fires from interference sources. For non-standard fires, the KNN model demonstrated an accuracy of 97.2%, with a recall rate exceeding 95%, suggesting robust generalization capability and minimal false negatives. Within the 40°~55° angular range, all models demonstrated accuracy levels exceeding 85% and recall rates surpassing 80%, thereby substantiating the resilience of the proposed spectral feature representation. A comparative analysis revealed that models utilizing multi-angle spectral features significantly outperformed conventional single-angle methods across all evaluation metrics, including precision, F1-score, and AUC. The system's performance demonstrated stability across a series of experimental trials, thereby substantiating its capacity for consistent detection under controlled conditions.

The integration of angle-resolved scattering spectroscopy and machine learning has proven to be a highly effective and reliable solution for accurate fire smoke detection amidst diverse interference sources. The proposed method offers significant advantages in reducing false alarms and enhancing detection sensitivity, particularly in challenging operational conditions. This research establishes a robust framework for intelligent fire detection systems with stronger generalization abilities and higher operational precision. Subsequent endeavors will concentrate on real-world validation, system miniaturization, and the development of adaptive learning mechanisms for continuous performance enhancement in dynamic environments. The findings of this study bear significant implications for practical applications in smart buildings, industrial fire safety, and public infrastructure protection.