Modelling compressive strength of concrete from non-destructive tests using machine learning

• Main Author: Villanueva, Daniel Kenneth
• Format: text
• Language: English
• Published: Animo Repository 2022
• Subjects: Concrete—Compression testing; Strength of materials; Civil Engineering
• Online Access: https://animorepository.dlsu.edu.ph/etdb_civ/5; https://animorepository.dlsu.edu.ph/cgi/viewcontent.cgi?article=1001&context=etdb_civ
• Institution: De La Salle University

Concrete is the most generally used structural material for construction these days. Traditionally, concrete has been created from a few well-defined components: Cement, water, fine aggregate, coarse aggregate, etc. In concrete mix design and quality control, the compressive strength of concrete is regarded as the most needed property. The main objective is to model the compressive strength of concrete obtained from non-destructive test using machine learning. The specific objectives are to narrow down the search to factors that significantly contribute to the compressive strength of concrete, to compare results from destructive methods with non-destructive methods and to create a model using the factors to be considered. Linear regression has successfully been used as a sanity check. The most successful non-NN algorithm, stepwise quadratic regression with interactions was also featured. The repeatability of the results is also a matter of interest since the neural network also changes rapidly even without changing the neural network parameters. The neural network, when tweaked, was also able to give a performance better than the other methods. It was demonstrated that neural networks could give results better than other methods could. Nevertheless, the investigation also made known that more data would be appreciated, especially outside of the 10-45MPa range. A model for the compressive strength of concrete from non-destructive test using machine learning has been obtained. The model is a neural network that contains only the more influential variables. This model could be seen on subsection 5.5.3.2 (page 151). This model (tweak neural network) is the best due to the higher R-squared compared to the other models. The best model contains twenty-one inputs with a hidden layer of size one to produce one output. This model has an R-squared of 0.9773. More data is recommended to be gathered. The analysis could be augmented if there is more data. More factors could be considered. The only non-destructive factor that was used in the analysis was rebound number and ultrasonic pulse velocity. Other factors such as Leeb hardness, electrical resistivity, and point load index would still need a greater foray into the current research.