An Enhanced Backpropagation Neural Network for Fire Alarm Detection

Fires are frequent emergencies, threatening human safety and causing considerable losses. Fire alarm detection systems are essential in preventing fires and minimizing more severe losses. However, current fire alarm detection systems suffer from one main problem: A high false alarm rate, which often results from overfitting of models. This occurs due to imbalanced or noisy data. To address these issues, an Enhanced Backpropagation Neural Network (BPNN) approach is proposed and evaluated in this study. The Enhanced BPNN utilizes various parameter sets to optimize its performance. Artificial Neural Networks (ANNs) with Backpropagation (BP) are known for their self-learning ability, adaptiveness, and relatively fast processing. In this study, the BPNN model is extensively experimented with by varying the number of layers, neurons, activation functions, and learning rates. To assess the reliability of the model, k-fold cross-validation and mean squared error (mse) are employed as evaluation metrics. Among the tested configurations, the BPNN model demonstrates superior performance under different scenarios. When employing sigmoid and tanh activation functions with a 3-layer architecture and a learning rate of 0.1, the average k-fold cross-validation accuracy consistently yields promising results at 98.34%. Additionally, utilizing the relu activation function, a 7-layer architecture, and a learning rate of 0.001 results in an average accuracy of 98.61%. This study highlights the significance of hyperparameter tuning and model optimization in achieving accurate and efficient fire alarm detection systems.