Interpreting Impact Echo Data to Predict Condition Rating of Concrete Bridge Decks: A Machine-Learning Approach

Abstract

Maintaining the structural reliability of highway bridges under a budget constraint necessitates the development of accurate prediction models of bridge deck deterioration to maximize bridge service life while minimizing life-cycle costs. Traditionally, the structural condition of a bridge deck is assessed using ordinal discrete indices, referred to as condition ratings (CRs), assigned based on an assessment of the visible signs of deterioration. Nondestructive evaluation (NDE) is being increasingly utilized to gain objective insights into structural deterioration. The impact echo (IE) test is a common NDE technique that relies on the acoustic resonance response of a bridge deck to detect subsurface delamination that can lead to spalling. However, IE data interpretation is largely done manually and the connection between the IE results and CRs is not fully explored. The aim of this study is to model the spectral characteristics of IE signals to quantify the structural integrity of bridge decks and predict CRs. First, a nearest neighbor clustering of IE signal energy distribution in the frequency domain is conducted to generate condition labels for each IE response (good, fair, poor) automatically. The condition labels are then input to a support vector machine (SVM) classification model to predict the CRs. The models are trained and tested using data from the Long-Term Bridge Performance (LTBP) data set pertaining to 38 tested bridges with recorded NDE data collected over a span of 2 years on average. The findings indicate that the proposed model is capable of automatically predicting CRs for bridge decks given the raw IE test data with an accuracy of 87.5%.