Automated Interpretation of Impact Echo Data using Statistical Modeling and Machine Learning

Abstract

Impact echo (IE) is becoming a standard nondestructive testing (NDT) tool for concrete bridge deck assessment thanks to its simplicity, established reliability and relatively low cost. The recent advances in automated data collection and non-contact measurements [1] [2] are also contributing to IE’s increasing popularity. The new generation of IE test equipment uses robotic platforms allowing faster data collection and larger areal coverage compared to the hand-held devices previously employed. The large volume of collected data presents numerous opportunities but also new challenges. The availability of NDT data facilitates data-driven decision making from IE (and other NDT) data while new approaches are needed to process and interpret ‘big’ NDT data [3] [4]. The former has been addressed for example in our recently published work, where IE data collected during Long Term Bridge Performance (LTBP) program are used to predict condition rating (CR) of concrete bridge decks [5]. This study focuses on the latter, the increasingly urgent need to automatically analyze and interpret IE data using statistical modeling, machine learning (ML) and deep learning (DL). We present results pertaining to the analyses of LTBP data without ground truth as well as those obtained on laboratory slabs with well-defined embedded defects [6]. The performance of different methods in IE signal classification is compared and discussed. Our findings indicate that the performance of different methods greatly depends on the amount and quality of available ‘labeled’ data (i.e., data tagged with the corresponding ground truth information). Creating standard quality labeled datasets is a critical step in exploiting ML and DL for IE (and other NDT) data analysis and interpretation.