Machine learning hybrid dynamic best model selection algorithm for real-time fire prediction using IoT-enabled multi-sensor data in buildings

作者：Mujeeb Ali Khan, Weiguo Song, Abbas Khan, Mazhar Ali, Rehmat Karim, Jun Zhang · 发表于：Journal of Safety Science and Resilience · 年份：2025 · DOI：10.1016/j.jnlssr.2025.100236 · 被引用次数：5 · 研究领域：Fire Detection and Safety Systems、Fire dynamics and safety research、Evacuation and Crowd Dynamics

Fire disasters in urban areas, including homes, offices, and industrial facilities, have increased significantly over the past decade, causing extensive damage and loss of life. Integrating intelligent fire detection systems with machine learning (ML) is crucial for providing early warnings and facilitating effective response coordination. In this research, a novel hybrid dynamic best model selection (HDBMS) ML-based algorithm is proposed for IoT-enabled fire detection in buildings, which outperforms traditional static models by providing higher accuracy and adaptability across diverse fire scenarios. The proposed algorithm employs feature selection pre-processing techniques, followed by the synergistic integration of five classifiers: support vector classifier (SVC), logistic regression, random forest, Gaussian Naive Bayes (Gaussian NB), and decision tree, aiming to enhance prediction accuracy and robustness across various fire scenarios. This system dynamically selects the optimal classifier based on real-time performance metrics such as precision, accuracy, F1-score, and recall. This approach was rigorously validated using our developed dataset and real-time sensor data to monitor smoke, temperature, and humidity under various fire scenarios. Following algorithm validation, a laboratory-constructed multi-sensor fire detection node prototype wirelessly feeds sensor data to the ThingSpeak cloud platform for real-time data analysis and communication with ML algorithms in the ...