Self- and semi-supervised learning for evacuation time modeling within fire emergencies in nuclear power plants

Using neural networks for effective emergency response planning is essential to safeguard nuclear power plants and their surroundings swiftly and accurately during fire emergencies. However, achieving precise training of neural networks for fire evacuation modeling necessitates the collection of labeled data encompassing a variety of fire scenarios and their corresponding evacuation times. The acquisition of this labeled training dataset involves direct engagement in experimental simulations of evacuation times for diverse fire scenarios, accomplished through the application of the consolidated fire and smoke transport (CFAST) simulator. However, the endeavor to amass a diverse pool of labeled data imposes significant time and financial costs. To overcome this challenge, we propose using self- and semi-supervised learning to construct a metamodel that approximates the simulator and to improve the ability of neural networks that accurately predicts evacuation times even in situations with limited labeled data. The effectiveness of our proposed framework is demonstrated through the experimental results conducted on CFAST datasets, thus emphasizing its potential to develop nuclear safety guidelines based on neural networks.