Non-destructive detection of thin micro-defects in glass reinforced polymer composites using a terahertz electro–magnetic wave based on a convolution neural network

Abstract

This study presents a non-destructive evaluation (NDE) method that utilizes a convolutional neural network (CNN) based terahertz (THz) signal processing to quantitively distinguish overlapping signals from micro-defects in glass fiber reinforced polymer (GFRP) composites. Theoretical electro–magnetic wave propagation models for describing the propagation of the THz signal into the GFRP composite were devised considering the interaction of THz with the micro-defects in the GFRP composite. THz signals were simulated with respect to the thickness of the defects, and transformed into two-dimensional (2D) spectrograms using a short-term Fourier transform (STFT). The converted 2D-STFT images were used to train the CNN deep learning model to classify the size and thickness of the defects. A probability map of the classification of the defect thickness was derived from the trained CNN deep learning model. Using the probability map based on the CNN model, the depth and size of the micro-defects thinner than 20 μm inside the GFRP composites could be successfully differentiated with 95% accuracy. This research proposes a promising approach for analyzing micro-defects in composite materials by quantitatively evaluating the extent of signal overlap, going beyond the traditional approach of training the model based solely on the signal.