Smart ensemble machine learner with hyperparameter-free for predicting bond capacity of FRP-to-concrete interface: Multi-national data. (22nd August 2022)

• Record Type: Journal Article
• Authors: Wang, Wei-Chih; Nguyen, Ngoc-Mai; Cao, Minh-Tu
• Is Part Of: Construction & building materials. Volume 345(2022)
• Publication Date: 2022-08-22
• Subjects: Fiber-reinforced polymer -- Bonding strength -- Structure repair/rehabilitation ensemble model -- Neural network -- Metaheuristic optimization -- Support vector regression; Building materials -- Periodicals; 624.18
• DOI: 10.1016/j.conbuildmat.2022.128158
• Languages: English
• ISSNs: 0950-0618

Abstract

Highlights: The study established a hyperparameter-free ensemble model, abbreviated as MNVIM. 855 data samples of FRP-to-concrete bond capacity were collected to train models. The results showed the MNVIM to achieve the greatest accuracy. A t -test confirmed the MNVIM to be superior than other AI and mathematical approaches. Abstract: Determining the bond capacity of a (fiber-reinforced polymer) FRP-to-concrete interface is important for the design task in repairing or rehabilitating concrete structures in infrastructure. This paper presents a novel framework for creating a hyperparameter-free ensemble model and, using an equilibrium optimizer (EO), demonstrates a radial basis function neural network (RBFNN) and least square support vector regression (LSVR). A new model, the metaheuristic-optimized neuron-vector machine inference model (MNVIM), was used for the accurate estimation of FRP-to-concrete interfacial bond capacity. The MNVIM was constructed using the EO algorithm to blend RBFNN and LSVR adaptively. In the MNVIM framework, the EO is involved in the learning phases of the RBFNN and the LSVR by giving an appropriate set of hyperparameters for the two constituent models, including the number of hidden neurons and the Gaussian impact value of RBFNN, as well as the regularization coefficient and kernel parameter. Simultaneously, EO adaptively adjusts linear combination weight values to blend the prediction values of RBFNN and LSVR. Statistical results from ten runs show that the newly developed model achieved the most desired evaluation criteria, namely, root mean square error (RMSE) (3.989 kN), mean absolute (MAE) (2.632 kN), mean absolute percentage error (MAPE) (15.56%), and R-square (0.839). The t -test statistical measurement further confirms that the prediction accuracy of MNVIM is significantly better than other artificial intelligence (AI) models and sixteen common mathematical approaches. These analytical results indicate that MNVIM is a promising tool for estimating bonding strength; thus, it significantly increases construction safety in concrete element repair/rehabilitation-related engineering design using FRP.