Application of neural networks for light gauge steel fire walls. (1st March 2023)

• Record Type: Journal Article
• Authors: Kesawan, S; Rachmadini, P; Sabesan, S; Janarthanan, B
• Is Part Of: Engineering structures. Volume 278(2023)
• Publication Date: 2023-03-01
• Subjects: Fire-resistance rating -- Cold-formed steel -- Neural network -- Light gauge steel frame wall -- Load ratio; Structural engineering -- Periodicals; Structural analysis (Engineering) -- Periodicals; Construction, Technique de la -- Périodiques; Génie parasismique -- Périodiques; Pression du vent -- Périodiques; Earthquake engineering; Structural engineering; Wind-pressure; Periodicals; 624.105
• DOI: 10.1016/j.engstruct.2022.115445
• Languages: English
• ISSNs: 0141-0296

Abstract

Highlights: Two separate neural networks to predict fire performance of LSF walls made of lipped and LSB sections have been proposed. Different available fire test datasets have been used for training the developed neural network. Effect of different neural network parameters on the prediction accuracy was investigated. Suitable neural networks using two layers, least-square loss- function, keep probability value of 0.9 were proposed for LSB walls with lipped and LSB sections. Abstract: This research study evaluates the applicability of neural networks to determine the fire-resistance rating of Light gauge Steel Frame walls, made of cold-formed steel studs and lined with gypsum plasterboard. Many full-scale tests and numerical studies were conducted to determine the fire-performance of LSF walls, but these studies were expensive and time consuming. Therefore, an alternative option of using machine learning to predict the fire performance was proposed. The actual fire performance data from FEA and full-scale tests of previous research studies were used as inputs in the machine learning. Separate sets of training and test data were used; thus, test data is not used to calibrate the machine learning-based analysis. Here LSF walls made of different wall configurations and different steel profiles were considered. Training and testing of the artificial neural network are performed by combining different parameters such as loss function, keep probability factor, learning rate, the number of layers, and neurons to determine the optimal and accurate solution. The structural fire capacity of LSF walls obtained from machine learning was compared against the test data to evaluate its accuracy. Based on the findings, suitable neural network models with two hidden layers and suitable loss functions were recommended.