Development and analysis of quantitative models for predicting false alarms and missed detections in smoke detection systems

Abstract

Substantial amounts of sensor data were processed by Machine Learning (ML) models to expose trends and patterns distinguishing genuine fire alarms from false signals. Such models improve the accuracy of fire alarm systems with the ability to detect infinitesimal changes in smoke readings, environmental conditions, and sensor patterns with history-based training. This improves overall fire safety by reducing false alarms and correcting earlier warnings. To forecast the precision of smoke sensor detection, the current research applied Bagging C and Histogram Gradient Boosting Classification (HGBC) models. These models were hybridized using Victoria Amazonica Optimization (VAO) and Smell Agent Optimization (SAO) to enhance prediction results and develop novel hybrid models with enhanced precision and realism. Comparative results for various phases, like training, testing, and all phases, were obtained to identify the most efficient model. For example, the HGVA model with 0.937 recall and the BAVA model with 0.965 recall were found to be the optimal models in this study, performing better than the BASA model throughout the testing process. It is also notable that 15 parameters like temperature, humidity, TVOC, CO2, UTC, and others influenced the performance of the models. In particular, the performance of the top three models was most affected by temperature, humidity, and UTC. The Kruskal-Wallis test was employed to analyze the effect of each of these factors to see which one had the greatest effect.