Transfer learning of Impact Echo signal classification from laboratory to the field

Abstract

Impact Echo (IE) is a popular nondestructive testing tool for condition assessment of concrete bridge decks. Traditionally, IE signals require to be manually interpreted and classified into discrete classes – good, fair and poor - representing the different conditions at test points. This study aims to develop an automated fuzzy classification approach for IE signals trained on a set of eight reinforced concrete specimens constructed at the FHWA Advanced Sensing Technology (FAST) NDE laboratory with artificial defects that are known apriori. We choose fuzzy classification to obtain condition “goodness” represented on a continuum rather than on discrete classes (good, fair and poor). We use fuzzy classification models informed by physics-based features and automatically extracted features to determine the departure of a bridge deck condition from ‘good’. Next, using transfer learning, we suitably tune the classification model for optimum performance on data from FHWA’s InfoBridge database that has a limited amount of IE data from real bridge decks. We evaluate the transferred model performance against expert-assisted manual signal classification. This study will demonstrate the utility of laboratory studies for developing models applicable to real field data, where little or no ground truth data are available.