Abstract

Conventional fire detection systems primarily rely on smoke and heat sensors, yet environmental parameters like air temperature, humidity, and pressure offer critical supplementary data.1 This study employs a sensor fusion approach using a Genetic Programming Symbolic Classifier (GPSC) to generate a transparent AI model in the form of a symbolic expression. To address significant class imbalance, the investigation evaluates multiple resampling techniques, including Random Sampling, Near Miss-1, ADASYN, SMOTE, and Borderline SMOTE. These datasets were processed using a custom random hyperparameter search and 5-fold cross-validation to ensure model robustness. Results indicate that the SMOTE-balanced dataset yielded superior performance, achieving an accuracy of 0.998  4.79  10^-5 alongside consistently high values for AUC and F1-score. The paper concludes by presenting the optimal symbolic expression and validating its efficacy on the original dataset.