Deep Learning for Detection and Characterization of Cracking in UHPC
Jun Wang, Yail J. Kim, and Chao Liu
Appears on pages(s):
crack-induced damage; deep learning; discrete entity; ultra-high-performance concrete (UHPC)
This paper presents a novel deep learning method built upon a convolutional neural network (CNN) combined with agent-based modeling for crack detection and damage characterization in ultra-high-performance concrete (UHPC). A total of 4,400 digital images taken from previous experimental programs are employed to train and validate the proposed vision-based computational approach, including UHPC mixed with steel and synthetic (collated and monofilament polypropylene) fibers. The architecture of CNN is established by four primary layers (transposed convolution, batch normalization, activation, and max pooling layers) and four auxiliary layers (global max pooling, flatten, dropout, and dense layers), associated with the gradient descent algorithm. In accordance with percolation theory, autonomous agents sense chromatic aberration in the spatial domain of UHPC, differentiate between cracked and uncracked regions, and eliminate morphological noises at the surface level. Processing cropped images at 128 128 pixels alongside 64 filters is computationally efficient in the transposed convolution layer. The inclusion of the batch normalization and max pooling layers in conjunction with a rectified linear unit expedites the calculation procedure. The maximum error of the estimate is 1.31% in the validation of the developed framework. The probability of observing cracks in each image is determined through the auxiliary layers with minimal overfitting. Cracking patterns dominate the failure rate of UHPC, and the steel fibers outperform their synthetic counterparts with regard to topological stability.