Using an evolutionary heterogeneous ensemble of artificial neural network and multivariate adaptive regression splines to predict bearing capacity in axial piles. (1st October 2022)

• Record Type: Journal Article
• Authors: Cao, Minh-Tu; Nguyen, Ngoc-Mai; Wang, Wei-Chih
• Is Part Of: Engineering structures. Volume 268(2022)
• Publication Date: 2022-10-01
• Subjects: Pile foundation -- Bearing capacity of axial piles -- Heterogeneous ensemble model -- Hybrid machine learning model -- Hyper-parameter optimization; 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.2022.114769
• Languages: English
• ISSNs: 0141-0296

Abstract

Highlights: The study integrates the equilibrium optimization (EO) with the heterogeneous ensemble of MARS and RBFNN, abbreviated as IMNNIM. The IMNNIM's performance is validated on 472 test reports of the driven pile static load collected on the construction site. An extensive analysis was conducted to remove the redundant attributes. The analytical results demonstrated IMNNIM to be the best model in predicting PBC by achieving the greatest values of MAPE (7.24%), RMSE (90.92kN), MAE (67.98kN) and R2 (0.930). A one-tailed t -test has proven the prediction accuracy of IMNNIM to be significantly superior to that of other approaches with the 95% confidence. Abstract: Accurately estimating the bearing capacity of piles is an onerous task in structural design task that requires a powerful computation model able to elucidate nonlinear impacts of geotechnical factors and the dimension and shape of piles. This study develops a novel heterogeneous ensemble artificial intelligence model, named intelligence multivariate neural network inference model (IMNNIM), to accurately and quickly predict bearing capacity of piles. The IMNNIM was created by integrating the equilibrium optimization algorithm (EO) into a combination of multivariate adaptive regression splines (MARS) and radial basis neural network (RBFNN). The predictive values of the IMNNIM were qualified by dynamically merging prediction information generated by MARS and RBFNN and then adjusting the associated weights and assigned tuning parameter values of the learner members. The performance of the IMNNIM was evaluated using data from 472 driven pile static load test reports. The experimental results using a 10 -fold cross-validation method demonstrated the IMNNIM to be the most reliable model for predicting pile-bearing-capacity by achieving the greatest values of MAPE (7.24%), RMSE (90.92 kN ), MAE (67.98 kN ), and R 2 (0.930) and the lowest standard deviation values. A t -test analysis method confirmed the IMNNIM as a superior tool for pile-bearing-capacity estimation.