Automated honeycomb detection during Impact Echo inspections in concrete using AI trained by simulation data

Abstract

The Impact Echo method is well established in the civil engineering world of NDT for defect detection and thickness estimation in thick and highly reinforced concrete structures. For most applications of Impact Echo however, only the resonance frequency of the measured time signal is evaluated, meaning that most information is neglected. Here, artificial intelligence (AI) in the form of machine learning can help to classify signals based on multiple input parameters and therefore make use of the additional information stored in the measured signals. As the most powerful classification models need labelled input data, this usually marks a problem since labelled NDT data sets are rarely available for concrete structures. One solution to overcome this problem is the use of numerical simulations. In the past, numerical simulations showed that they are capable to produce realistic synthetic data for Impact Echo testing in concrete specimens. In this study, numerical simulations of Impact Echo measurements were conducted using the Elastodynamic Finite Integration technique (EFIT) to create training data for machine learning models. The measurements were carried out on two concrete specimens (17 cm and 50 cm thickness) containing honeycombs. Using the simulation data, multi-layer perceptron (MLPNN) and convolutional neural networks (CNN) are trained and tested on measured data from each specimen for performance. Results showed that an accurate honeycomb detection using machine learning was only possible in some cases with many false alarms arising near the specimen edges.