Identification of non-dimensional density distribution of concrete structures based on self-attention using hammering response data.

The aging of concrete structures has become a serious problem in Japan, and periodic maintenance is essential for preventing accidents caused by structural aging. In this study, a method for estimating defects in concrete using data from hammering tests on a concrete plate using machine learning was developed. A neural network based on self-attention (SAN) to estimate the three-dimensional position and size of the defects was constructed. Moreover, a dataset was created from the topology of the internal defects and the acceleration response waveform when a concrete plate was struck. The entire plate was represented as a non-dimensional density matrix. The SAN used scalograms of the acceleration responses as the input. Furthermore, two types of data augmentation ('Flip' and 'Rotate') were proposed. The use of both data augmentation techniques achieved the highest accuracy. By setting an appropriate number of rotations, the model was able to estimate all defects in the dataset. [ABSTRACT FROM AUTHOR]