Developing a supervised machine-learning model capable of distinguishing fiber orientation of polymer composite samples nondestructively tested using active ultrasonics

This study evaluated the paired performance of different signal processing techniques and supervised learning models being capable of identifying subtle differences in otherwise similar acoustic signals related to detecting the fiber orientation of a polymer composite. Projection of Latent Structures models demonstrated poor predictive capabilities of the composite structure based on spectral analysis of the acoustic signal. AI based models showed great improvements to the capabilities, with artificial neural network modeling exceeding Convolutional Neural Networks for correct classification accuracies. The continuous wavelet transfer highlighted the greatest degree of differences in the signal response compared with fast Fourier Transformation or short time Fourier transformation. The use of regression-based predictions over classification-based was found to greatly improve the predictive capabilities of the models, especially when multiple fiber orientations were present in a sample. A time-based analysis of spectral data showed the frequencies of the signal changed based on the orientation of the fibers. The acoustic signals for the samples with multiple fiber orientations contained individual artifacts representing components of each individual orientation. Use of the frequency domain was shown as capable of observing the targeted fiber information within the bulk material in real-time. This work shows great promise for composite material predictions using active ultrasonics, with the potential to be implemented into in-line systems.