Multistage Early Fire Detection With Convolutional Neural Networks And Metaheuristic Optimization Algorithm

Abstract

Fire is a frequently occurring disaster that causes significant personal injury and extensive property damage. Effective fire detection is crucial for protecting lives and property. Current trends in fire detection involve using real-time models such as YOLO. However, these models require substantial computational resources, making them difficult to implement on low-resource devices. Moreover, many YOLO-based works are trained on small datasets, failing to capture the dynamic characteristics of fire. To address these challenges, we propose an innovative fire detection approach that employs convolutional neural networks (CNNs) combined with a metaheuristic optimization algorithm for feature selection. This optimization process refines the feature set, ensuring that only the most relevant features are retained for further analysis. Once the optimal features are selected, they are used to train a variety of classifiers to assess performance. These classifiers include advanced models such as XGBoost, LightGBM, CatBoost, KNN, Decision Tree, Random Forest, and SVM, each known for their ability to manage large datasets and complex feature interactions effectively. By employing this diverse set of classifiers, we can conduct a comprehensive evaluation of the feature set, ensuring that the selected features contribute significantly to the performance of the proposed fire detection model. The training process is conducted using an extensive dataset comprising over 100,000 fire images. This large and diverse dataset includes various fire scenarios and ii environmental conditions, enabling the model to learn and recognize fires across different contexts. In addition to leveraging advanced technology, our approach emphasizes practical validation through controlled experiments. Controlled experiments involve manipulating one or more independent variables such as the number of epochs, metrics, and optimizers while keeping other variables constant to observe the effect on dependent variables. By systematically testing the model under controlled conditions, we gather invaluable insights into fire behavior and detection patterns. Our proposed framework addresses the inherent limitations of existing fire detection systems such as late detection, high false alarm rates, and inefficiencies in computational resource usage by combining state-of-the-art technology with rigorous empirical research. This comprehensive approach prioritizes accuracy, efficiency, and early intervention, significantly advancing fire safety measures. The proposed method achieved high performance with 99.17% accuracy with the tested dataset, 99% recall, and 98% F1-score. It achieves a balance between combining performance and speed, resulting in a more dependable and accurate solution for the detection of fire, and is suitable for real-world applications.