Concrete crack simulation and its machine learning application in propagation prediction

Abstract

This study introduces an innovative approach for the Automatic Simulation of Concrete Cracks (ASCC), integrating simulation and programming software, as well as its machine learning (ML) application in propagation prediction. The ASCC offers an automated simulation facilitated through a user interface, allowing for seamless adjustment of boundary conditions, including support conditions, geometric sizes, and simulation parameters. The crack propagation data obtained from ASCC are employed to train ML models, the correlation of which with real-world crack is verified on some reputable crack image datasets. Experimental results confirmed the effectiveness of samples generated from the Cantilever simply supported beam in approximating real-world cracks. A comparison with a recent relevant work demonstrated smaller fitting errors on 50% of the examined crack image datasets when approximating real-world samples with the best-fit simulation. Comparative analysis indicates that ASCC is significantly faster than manual intervention, i.e. the time required for a simulation in some boundary support conditions is only 4.5% compared to the processing time of an engineer. This achievement is meaningful in simulation and has potential applications for data-intensive tasks, such as vision-based crack detection using ML or deep learning. This work also highlighted the importance of the number of crack tips, which could lead to overfitting when using ML.