Ultrasound classification of interacting flaws using finite element simulations and convolutional neural network

• Published in: Engineering with computers Vol. 38; no. 5; pp. 4653 - 4662
• Main Authors: Niu, Sijun, Srivastava, Vikas
• Format: Journal Article
• Language: English
• Published: London Springer London 01.10.2022; Springer Nature B.V
• Subjects: Artificial neural networks; CAE) and Design; Calculus of Variations and Optimal Control; Optimization; Classical Mechanics; Classification; Computer Science; Computer simulation; Computer-Aided Engineering (CAD; Control; Corrosion; Cracks; Finite element method; Flaw detection; Machine learning; Math. Applications in Chemistry; Mathematical and Computational Engineering; Neural networks; Original Article; Stress corrosion cracking; Systems Theory; Three dimensional printing; Time signals; Ultrasonic imaging; Validation studies; Corrosion; Neural networks; Pipelines; Non-destructive evaluation; Inline Inspection; Ultrasound; Finite element method; Flaw classification; NDT; Crack
• ISSN: 0177-0667; 1435-5663
• DOI: 10.1007/s00366-022-01681-y

Interacting flaws refer to conditions when two flaws are in close proximity and have the potential to interact with each other to significantly reduce the integrity of a structure. Accurate detection and classification of non-visible interacting and single flaws using ultrasound time signals continue to be a significant challenge. Machine learning-based flaw detection and classification systems are promising, but have been unable to be implemented as they lack training data that are expected to be obtained from a large set of well-labeled field data or experiments. Cracks and corrosion wall loss are two flaw types of primary concern in metallic structures, and are the focus of this study. We present an approach that utilizes finite element simulation data to train an ultrasound time signal-based convolutional neural network (CNN). No flaw, single crack, single wall loss corrosion, two cracks, and combined crack with corrosion are the five categories comprising single and interacting flaws considered in this work. A dataset containing 2000 numerical ultrasound NDT signals created through finite element simulations was used to train an optimal CNN architecture. A validation study was conducted using 13 3D metal printed steel specimens containing a variety of interacting and single flaws. Twenty-five measurements considering precise and offset transducer placements were used for the validation study. The simulation-trained CNN showed 100% accuracy in classifying all categories of flaws from the independent experimental ultrasound NDT signals. The results are promising as the classification of non-visible interacting flaws that has traditionally been a very difficult problem could be addressed using the methodology presented here.