Unsupervised machine learning for flaw detection in automated ultrasonic testing of carbon fibre reinforced plastic composites

• Published in: Ultrasonics Vol. 140; p. 107313
• Main Authors: Tunukovic, Vedran, McKnight, Shaun, Pyle, Richard, Wang, Zhiming, Mohseni, Ehsan, Gareth Pierce, S., K. W. Vithanage, Randika, Dobie, Gordon, MacLeod, Charles N., Cochran, Sandy, O'Hare, Tom
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.05.2024
• Subjects: Aerospace composites; Automated defect detection; Carbon fibre reinforced polymers; Deep learning in ultrasonic testing; Machine learning; Non-destructive evaluation; Carbon fibre reinforced polymers; Deep learning in ultrasonic testing; Non-destructive evaluation; Aerospace composites; Automated defect detection; Machine learning
• ISSN: 0041-624X; 1874-9968
• DOI: 10.1016/j.ultras.2024.107313

•The automated scanning setups have enhanced the throughput of NDE data.•Unsupervised machine learning models can be used to accelerate data interpretation.•Unsupervised clustering can be used to automatically preprocess ultrasonic data. The use of Carbon Fibre Reinforced Plastic (CFRP) composite materials for critical components has significantly surged within the energy and aerospace industry. With this rapid increase in deployment, reliable post-manufacturing Non-Destructive Evaluation (NDE) is critical for verifying the mechanical integrity of manufactured components. To this end, an automated Ultrasonic Testing (UT) NDE process delivered by an industrial manipulator was developed, greatly increasing the measurement speed, repeatability, and locational precision, while increasing the throughput of data generated by the selected NDE modality. Data interpretation of UT signals presents a current bottleneck, as it is still predominantly performed manually in industrial settings. To reduce the interpretation time and minimise human error, this paper presents a two-stage automated NDE evaluation pipeline consisting of a) an intelligent gating process and b) an autoencoder (AE) defect detector. Both stages are based on an unsupervised method, leveraging density-based spatial clustering of applications with noise clustering method for robust automated gating and undefective UT data for the training of the AE architecture. The AE network trained on ultrasonic B-scan data was tested for performance on a set of reference CFRP samples with embedded and manufactured defects. The developed model is rapid during inference, processing over 2000 ultrasonic B-scans in 1.26 s with the area under the receiver operating characteristic curve of 0.922 in simple and 0.879 in complex geometry samples. The benefits and shortcomings of the presented methods are discussed, and uncertainties associated with the reported results are evaluated.