Hybrid Ensemble Based Machine Learning for Smart Building Fire Detection Using Multi Modal Sensor Data

Fire disasters are one the most challenging accidents that can take place in any urban buildings like houses, offices, hospitals, colleges and industries. These accidents which the world faces now, have never been more frequent and fatal, leading to innumerable loses, damage of expensive equipment and unparalleled human lives. The concrete landscapes are threatened by fire disasters, which have prolifically outnumbered in the last decade, both in intensity and frequency. Thus, to minimize the impact of fire disasters, adoption of well planned, intelligent and robust fire detection technology harnessing the niches of machine learning is necessary for early warning and coordinated prevention and response approach. In this research a novel hybrid ensemble technology based machine algorithm using maximum averaging voting classifier has been designed for fire detection in buildings. The proposed model uses feature engineering pre-processing techniques followed by a synergistic integration of four classifiers namely, logistic regression, support vector machine (SVM), Decision tree and Naive Bayes classifier to yield better prediction and improved robustness. A database from NIST has been chosen to validate the research under different fire scenarios. Results indicate an improved classification accuracy of the proposed ensemble technique as compared to reported literatures. After validating the algorithm, the firmware has been implemented on a laboratory developed prototype of smart multi sensor, embedded fire detection node. The designed smart hardware is successfully able to transmit the sensed data wirelessly onto the cloud platform for further data analytics in real time with high precision and reduced root mean square error (MAE).