Study on Applicability of Detection of Internal Cracks by Impact Echo Using Machine Learning

Abstract

In this study, as a part of the comprehensive deterioration evaluation system for concrete structures by combining several diagnostic methods such as impact echo detection and infrared diagnosis, automatic judgement of concrete defects was attempted by machine learning of impact echo spectrogram. In machine learning, the impact echoes were transformed into spectrograms using the short-time Fourier transform, and supervised learning was performed using a convolutional neural network (CNN). This method can judge defects with the same accuracy as experienced engineers in the specimen with simulated defects buried in it. So, it was investigated the possibility of detecting micro cracks in concrete under conditions closer to real structures by this method. In the experiment, a reinforced concrete specimen with dimensions of 1800 x 1800 x 600 mm and the cover concrete of 70 mm was fabricated. Then, by electric corrosion of the specimen, the steel bars inside were corroded, and micro cracks were generated. Impact echoes were measured before and about every week after the start of electric corrosion. Then, machine learning was performed using the impact echo spectrograms before the start of electric corrosion and after internal cracks were determined to have occurred. The CNN obtained from this process was used to judge the impact echo spectrograms during the progress of electric corrosion. As a result, the area judged as defects became larger after the rapid growth of strain considered to be caused by the internal cracks. In addition, when the impact echo spectrograms with further electric corrosion were used as training data, the area judged as defect became smaller. From these results, this method was found to be capable of judging micro-cracks due to corrosion expansion of steel bars in the concrete.