Automatic elimination of invalid impact-echo signals for detecting delamination in concrete bridge decks based on deep learning

Abstract

The impact-echo (IE) method is effective for evaluating invisible defects. However, it might return misleading results when its signals are invalid. This challenge aggravates when the tests are conducted using robotic devices that automatically collect massive data. This study proposes an automatic method to eliminate invalid signals based on the ResNet model. First, the signals are visualized into two-dimensional images as the input for ResNet. The input data can then be classified into valid and invalid data via the ResNet model, which is trained with 11,290 signals and tested with 5664 signals. Finally, defects can be detected using the dominant frequencies of the valid-class data. A case study with IE data from two concrete bridges was employed to validate the feasibility of the proposed approach. The results indicate that the method can achieve an average accuracy of 90.6% for eliminating invalid signals and significantly improve the IE test accuracy.