Noise influence on defect depth estimation in CFRP by One-Dimensional convolution neural network in Non-Linear frequency modulated thermal wave Imaging

Abstract

Quantitative depth estimation of defects in composite structures using active infrared thermography is a challenging task and recent advancements promote deep learning (DL) techniques to achieve this in a quick and automated manner. However, noise present in thermal data at extreme inspection conditions degrades the performance of the DL models. This paper introduces a one-dimensional convolution neural network (1D-CNN) and analyzes its performance in different levels of noise for automatic depth estimation in Logarithmic Frequency Modulated Thermal Wave Imaging. The experimentation is conducted over two carbon fiber reinforced polymer (CFRP) specimens of different thickness and varying sizes and depths of flat-bottom hole defects. Including the original thermal data of one CFRP specimen, five datasets are prepared by adding additive white Gaussian noise of four levels: 5 dB, 10 dB, 15 dB and 20 dB and the proposed 1D-CNN model, named as M1 to M5 for each case, is individually trained and tested. Further, transfer learning is applied for identifying depths in the second CFRP structure. The results, along with performance metrics, indicate that the 1D-CNN presents more than 98 % and 95 % accuracy for training from scratch and transfer learning cases over original data, whereas it degrades as the noise level increases. Performance metrics such as accuracy, F-score and mean intersection of union demonstrate the defect depth estimation performance, including the defect boundary preservation under various noise conditions that are consistent with the accuracy of the model. In addition, the proposed 1D-CNN achieves the best quantitative performance, with consistently higher accuracy (≈95–98 %), F-score (≈0.85–0.95), and mean IoU (≈0.45–0.75) across all defect depths and noise levels, outperforming conventional DT, SVM, ANN, and 1D-ResNet, whose metrics drop sharply, for deeper defects and higher noise conditions.