CHALLENGE TO EVALUATION OF ROUGHENED CONCRETE AREA RATIO BY DEEP LEARNING

Abstract

 When a new member is connected to an existing concrete member, roughened concrete and post-installed anchors are often applied to the joint. An electric hammer is used to create the uneven roughened concrete surface. There appears to be no detailed and general design code with regard to the shape and strength of the concrete in existing design guidelines. However, according to a series of studies conducted by us, setting the area ratio to 0.3 or more is important for shear resistance in practical constructions. However, there is a problem of how to manage the uneven shape that differs considerably in the consciousness and technique of contractors. For this purpose, a tool that can easily calculate the roughened area ratio quantitatively at construction sites is required. In recent years, deep learning has been applied to various studies and applications. Deep learning does not require new advanced algorithms to be designed and can make complex decisions quantitatively. Therefore, in this study, we investigated whether the roughened area ratio could be measured from photographs using deep learning. Thus, we investigated whether the roughened area can be classified using deep learning and what learning conditions ware suitable.

 First, specimens with roughened surfaces were prepared and trained via deep learning. Hence, it is possible to classify the roughened surface of the test specimen in a satisfactory manner; however, classifying the roughened surface of the practical concrete members is not possible. Because the roughness of the specimen is different from that of the practical frame, it is assumed that training with a practical frame is required to apply it to a practical frame.

 Subsequently, the images of the practical frame were trained. Here, photographs of the two buildings were prepared, and the combination was changed to verify the performance of deep learning based on the difference in the tendency of the training data. As a result, training that involved images with various roughened shapes and smooth surface conditions could cope with various types of roughening. On the contrary, in training using images with uniform roughened shapes and uniform smooth surface conditions, the classification performance toward other types of images significantly reduces, although the classification performance toward similar images is high. For this reason, the performance of deep learning largely depends on the training data. Therefore, performing training using data suitable for the required performance is preferable.

 Subsequently, we investigated the conditions of deep learning to increase the precision such that the measured roughened area ratio was as safe as possible. As a result, the highest precision was obtained by applying DeepLab v3+, Inception-ResNet-v2, and SGDM for stochastic gradient descent. On the contrary, under some bad conditions, the roughened surface could not be evaluated. As mentioned above, evaluating the roughened surface by appropriately training roughening using deep learning is possible, and we believe that it can contribute to site management.