Tempnet: A graph convolutional network for temperature field prediction of fire-damaged concrete

Abstract

Determining the damage level of the fire-damaged concrete structure is critical for the structural assessment and repair of buildings after fire. Existing methods assess the damage levels of concrete by measuring the remaining mechanical performance in a traditional manner, where they either have limited accuracy or efficiency due to the need of heavy machines and experienced laborers. In contrast with these methods, we propose a deep learning based approach called Tempnet to promote the efficiency and effectiveness of damage level assessment for concrete after fire. Tempnet incorporates a graph convolutional layer and a conventional convolutional layer to encode the temperature interdependency between neighboring areas in the images of fire-affected concrete to capture the exposed temperature fields of the concrete. Three closely related application scenarios together with their corresponding datasets have been proposed to evaluate the performance of Tempnet. Extensive comparative experiments and ablation studies have validated the model design, the high efficiency, and the robust performance of Tempnet, with a performance metric F1 value higher than 0.97 in all applications. Case studies were conducted further to provide insightful illustration of Tempnet’s impressive performance. It is envisioned that the Tempnet can contribute to the efficient maintenance of concrete structures following fire accidents for construction applications.