Authors: Xiaofeng Zheng , Guo-Qiang Li , Wei Ji , Shaojun Zhu
Abstract: In emergency scenarios such as building fires, swift and efficient prediction of structural collapses is paramount for effective rescue operations. This paper proposes an innovative edge computing-based framework designed to optimize models to capture the key physical parameters in real time for early warning of fire-induced building collapses, aiming to reconcile the challenge of minimizing model volume and operational costs without compromising accuracy. Our approach focuses on refining the model through a comprehensive optimization strategy. This includes an advanced sampling method that enhances accuracy by comparing datasets derived from normal and uniform distributions. Besides, thermocouples are strategically deployed, guided by explainable artificial intelligence, to significantly reduce unnecessary input features and associated costs. Additionally, the genetic algorithm is used to streamline the deep neural network by minimizing the number of trainable parameters. The culmination of these efforts results in a significantly more compact model. For predicting displacements at vertices, the model achieves a goodness of fit (R-squared value) ranging from 0.9 to 1 for 84 % of the test set, while reducing the total number of parameters by 78 %. Additionally, the model lowers hardware deployment costs and addresses the computational challenges of processing extensive feature sets. This paves the way for the deployment of lightweight, accurate, and cost-effective early-warning systems in emergency response contexts.