CNN-based image analysis approach for predicting THR of combustible items in buildings

Abstract

The expansion and regular update of fire load databases is crucial for maintaining and improving the reliability of building evacuation safety and fire resistance design frameworks. To enhance the efficacy of fire load surveys, we developed a two-stage method to predict the total heat release (THR) $q$ of combustible items for fire load surveys in buildings. The first stage uses Convolutional Neural Network (CNN) to predict the weight $w$ from color image data of combustible items. The second stage multiplies this weight by the calorific value $\Delta H$ obtained through regression analysis of burn test results. The weight prediction and THR conversion are independent procedures. For weight prediction, we collected web-published data on nine types of furniture and electrical appliances, which served as a training dataset for estimating CNN parameters. We confirmed that predictions could be made with reasonable accuracy for all categories. However, electronic devices such as "Desktop PC", "Laptop PC", and "TV and Monitor" exhibited lower prediction accuracy. These items, often black and box-shaped, proved difficult to evaluate in terms of absolute spatial dimensions. For THR conversion, we enabled processing for twelve types of combustible items. Due to limited burn test data for some combustible items, we employed hierarchical Bayesian modeling to achieve stable regression. We then applied these procedures to predict the weight and THR of combustible items stored in actual buildings. While THR validation proved challenging, we confirmed that weights could be predicted with reasonable accuracy.