Session: Research Posters

Paper Number: 165056

Ai-Assisted 3d Modeling and Defect Classification in Welds Using Phased Array Ultrasonic C-Scan Data 

Welding is indispensable in structural steel members due to its ability to create strong, durable, and versatile connections essential for constructing safe, efficient, and resilient infrastructure. Welded parts must be inspected to ensure they meet required quality and integrity standards. Quality assurance measures such as nondestructive testing (NDT) can be employed to verify the integrity of welds, providing confidence in the structural integrity of steel members. A range of nondestructive testing (NDT) techniques may be used for welding inspection, which depends on several factors, including geometry and configuration of the joints, material, thickness of the welded parts, and inspection environment. Among the variety of NDT methods, configurations of ultrasonic testing (UT) are identified as efficient techniques of inspection for welding. Despite the effectiveness of the traditional UT techniques, the accuracy and repeatability can be significantly improved via advanced phased array technology. Further, the ability to identify defect type and sizing in a fast and efficient manner remains an obstacle in advanced inspection. The Integration of artificial intelligence (AI) into nondestructive testing (NDT) presents a transformative opportunity to enhance the evaluation of welded structures, allowing for reduced inspection time, expense, and risk of operational interruptions linked to structural inspection.

This research explores AI-assisted 3D modeling and defect classification in welds using raw Phased Array Ultrasonic (PAUT) C-scan data. Through the combination and application of convolutional neural networks (CNNs) alongside advanced data visualization using MATLAB for 3D deep learning, this project aims to improve the accuracy, efficiency, and interpretability of weld inspections in structural applications.

 PAUT is used to evaluate the quality of the welding in steel samples that were fabricated in various welding processing conditions. Defects are identified using PAUT and exported to OmniPC, an Evident software. From here, the raw C scan data, reflecting time of flight (TOF) information, is processed through normalization, denoising, and contrast enhancement to optimize inspection analysis. This, when applied to annotated data sets that represent common defect types such as porosity, lack of fusion, slag inclusions, and cracking, is capable of comparing the raw data output and initial classification of the defect indication using a lightweight CNN. Classification outputs are used in a deep learning model to identify flaw type and potential size range.

After the classification model outputs are mapped back onto the spatial coordinates of the scan to support defect localization. Using C-scan frames and MATLAB, a 3D volumetric representation of the weld is rendered to visualize both the weld and internal discontinuities. Defect types are color-coded within 3D space to enhance interpretability, allowing the operator to distinguish between critical and non-critical indications.

This integrated workflow supports both automated decision-making and interactive forensic analysis, offering potential applications in real-time inspection, digital twin development, and structural health monitoring (SHM) of weld-critical infrastructure. The combination of AI classification and volumetric modeling not only reduces subjectivity in defect interpretation but also introduces scalable pathways for intelligent NDT reporting and visualization.

The results of this study demonstrate that AI models can accurately identify weld defects in C-scan data with minimal preprocessing and that meaningful 3D visualizations can be generated directly from the classified outputs. This approach offers significant value for civil, mechanical, and aerospace engineers seeking to improve weld reliability assessment using modern digital tools. Future work will focus on incorporating A-scan and S-scan data for enhanced depth estimation, improving model robustness across multiple weld geometries, and developing a modular software package for industry deployment.

Presenting Author: Elsie Lappin Georgia Southern University

Presenting Author Biography: Elsie Lappin is a master's Civil Engineering student at Georgia Southern University. Her research focuses on Nondestructive Testing (NDT) for Structural Health Monitoring of Structural Steel and Concrete Composite Structures. Elsie has gained practical field experience working with the Georgia Department of Transportation’s State Bridge Design and Forensic Research units, blending academic expertise with real-world applications. She is passionate about advancing structural engineering practices and fostering innovation in NDT technologies. As president of the American Society of Nondestructive Testing–Coastal Georgia Section, Elsie leads industry partnerships and student outreach initiatives.

Authors:

Elsie Lappin Georgia Southern University
Hossein Taheri Georgia Southern University