A Waveform Category Neural Network Approach to Air Gap Detection in Reinforced Concrete

In our previous work, plain (non-reinforced) concrete was modeled as a homogeneous medium, and a Neural Network (NN) was trained using scattered waveforms to detect the presence of an air gap. In this study, we address the problem of air gap detection in reinforced concrete. The presence of reinforcing bars significantly increases the diversity of scattered waveform patterns, which in turn requires a large amount of training data and makes the training of NN more challenging. To overcome this issue, the scattered waveforms are classified into multiple categories, and separate Convolutional Neural Networks (CNNs) are trained for each category. The outputs of these CNNs are fused and streamed to another neural network for air gap detection capable of generalizing across different waveform categories; this network is referred to as a waveform category neural network (WCNN). This paper describes the concept of the WCNN and demonstrates its usefulness by applying it to air gap detection in reinforced concrete.