AI & ML In Structures (RAC & Structural Health Monitoring)

Abstract

This project aims to signify the use of Artificial Intelligence and Machine Learning for sustainable construction with primary focus on recycled aggregate concrete as an eco-friendly material. Data collection involved corresponding with researchers at Guang University for access to specialized datasets, poring through Research Gate repositories and University of California Irvine Database. The parameters of focus are strength prediction of NAC, RAC and their performance in deleterious environments including sulfate corrosion, carbonation depth & chloride ion erosion. For parameter prediction we decided on 7 machine learning models; namely Decision Trees, Random Forest, AdaBoost, Gradient Boosted Decision Trees, Historical Gradient Boosting Decision Trees, Extreme Gradient Boosting Trees and Extreme Gradient Boosting Random Forest. XGBoost emerged as the most effective model at ignoring noise and identifying the underlying pattern for prediction. The selection process involved GridCV Search for hyper-parameter tuning. After XGBoost turned out to be the best model it was optimized using Bayesian Optimization with T-Parzen Estimater as the surrogate model for 3000 iterations. This process was carried out for all the 5 parameters and the resultant hyper parameters were used with processed training data for fitting the XGBoost Model. To ensure accuracy and model generalisation numerous steps were undertaken. These include 3 step kfold validation, randomized data splitting, numerous split trains, randomized validation batches and experimental lab casting. After achieving high accuracy an application was developed using the latest pyQt5 framework which allows civil engineers to make use of our machine learning models to aid their mix design and construction processes. Additionally, it enables users to simulate concrete performance in deleterious environments permitting specialized construction for hazardous situations. The application also has built-in graphs which give valuable insight into model accuracy and how their prediction stacks against the training data which was used in the training process. An artificial intelligence neural network was trained using multiple datasets which were compiled from multiple repositories, labelled using supervision library and then used for training of YOLOv8 model with transfer learning for 143 epochs. Key challenges in this project included ensuring model accuracy, data pre-processing, iterative training for result improvement, prolonged training durations and hardware limitations.