Delamination damage identification through multi-input data using impact echo and study of model efficiency

This paper describes delamination damage identification using a robust multi-input deep convolutional neural network (MI – DCNN) model. By incorporating various impact echo signals, including time-domain, frequency-domain, and short-time Fourier transform (STFT) data, the proposed model captures a wider range of signal characteristics, thus improving the accuracy of damage detection. The study compares four different damage identification models applied to six artificially created delamination cases with varying depths and sizes. Among the models, the MI – DCNN that leverages frequency-domain and STFT data achieves the highest testing accuracy at 89.7%. Furthermore, the study explores the trade-off between model accuracy and computation time, concluding that optimal performance is achieved using 200 × 200 resolution images and 100 input samples. To assess real-world applicability, the model's performance was validated using previously collected field data from a concrete wall, where known defect areas were identified through core sampling. These findings provide valuable insights for improving damage detection in non-destructive testing techniques, offering practical implications for enhancing structural health monitoring systems, where both accuracy and computational efficiency are critical.