Defect Detection via Through-Transmission Ultrasound Using Neural Networks and Domain-Specific Feature Extraction

Defect detection in acoustically matched media remains a significant challenge, particularly when defects, such as fiberglass and polyamide residues, exhibit properties that match those of fiber-reinforced composite laminates as the base material. Techniques, such as through-transmission ultrasound (TTU), often miss subtle residues as defects with the use of conventional amplitude-based TTU detection alone. There is a noticeable research gap in properly identifying such subtle residues in composites using TTU inspection. This study investigated the use of neural networks (NNs) to identify subtle defects in composites based on domain-specific feature extraction from TTU signals. Each signal waveform of each spatial TTU inspection is used as a discrete sample to obtain a larger dataset for each specimen. Domain-specific features were extracted separately from the time, frequency, and wavelet domains, resulting in independent feature vectors to emphasize the signal characteristics. The NN classification used 70% of the overall dataset for training and 30% for testing. Results reveal the features of the time- and frequency domains perform well, achieving macro-F1 scores of 0.96 and 0.97, respectively, while wavelet domain features perform lower with a macro-F1 score of 0.62. Wavelet-domain features perhaps need machine learning methods like recurrent NNs to correctly recognize subtle time-dependent signal variations.