Defect detection in orbital GTAW joints of Al5083-5 wt% SiC composite using texture-oriented machine learning
High Temperature Materials and Processes · 23.01.2026 · 45 (1)
DOI: 10.1515/htmp-2025-0138 ↗Kurz gefasst
Fachartikel in der Zeitschrift High Temperature Materials and Processes, online erschienen am 16.11.2026. Verfasst von Senthilkumar Thamilkolunthu und 3 weiteren Autor:innen. Die Arbeit nennt 115 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
This study investigates defect detection in orbital GTAW-welded joints fabricated from particle-reinforced Al5083-5 wt% SiC metal matrix composites under specific welding conditions (120 A, 18 V, 100 mm/min). The inherent fusion-induced microstructural heterogeneity is attributed to thermal–mechanical mismatch between the ductile aluminum matrix and thermally stable SiC particles, which leads to spatial variations in the fusion zone that challenge conventional visual inspection. To address this, machine learning techniques are applied to systematically identify and classify welding defects across four joint configurations: flat, vertical, inclined (45°), and curved. A total of 40 welded specimens (10 per geometry), 120 metallographic sections, 360 original SEM micrographs, and 1,200 segmented image patches (300 per geometry, ∼400 per defect class) were collected to capture microstructural details for defect characterization. Texture-based feature extraction methods – Non-Local Binary Patterns (NLBP) and Neighborhood Gray-Tone Difference Matrix (NGTDM) – quantify spatial intensity variations, while Convolutional Neural Networks (CNNs) enable automated feature learning. Classification using Random Forest and CNN models demonstrates that the CNN achieves an overall accuracy of 97.9 % on the independent test set (macro-averaged per-class accuracy: 97.97 %), with strong robustness across defect types and joint geometries. Notably, CNN performance for MTD detection yields an AUC of 0.996. Five-fold cross-validation yields a mean accuracy of 97.4 ± 0.6 %, confirming model stability. NGTDM-derived texture features demonstrate high sensitivity to localized disruptions, with busyness showing a normalized influence value of 0.892, indicating its relevance in identifying irregular fusion behavior. Grad-CAM and confusion matrix analyses further validate model discriminative capability. The novelty of this study lies in applying hierarchical defect classification to orbital-welded Al5083-SiC composite materials, building upon established taxonomies (ISO 6520, AWS D17.2). This includes SEM-level comparison across different joint geometries and quantitative correlation of texture features with microstructural instability induced by fusion in particle-reinforced composite welds. Additionally, NGTDM features were correlated with crack density, porosity, and particle clustering, linking image descriptors to microstructure and improving defect characterization under varying SEM conditions. This offline methodology provides a foundation for future development of non-destructive inspection approaches.
Zitieren
APA
Thamilkolunthu, S., Rengasamy, H., Govindasamy, M., & Muthukumarasamy, A. (2026). Defect detection in orbital GTAW joints of Al5083-5 wt% SiC composite using texture-oriented machine learning. High Temperature Materials and Processes, 45(1). https://doi.org/10.1515/htmp-2025-0138
Vancouver
Thamilkolunthu S, Rengasamy H, Govindasamy M, Muthukumarasamy A. Defect detection in orbital GTAW joints of Al5083-5 wt% SiC composite using texture-oriented machine learning. High Temperature Materials and Processes. 2026;45(1). doi:10.1515/htmp-2025-0138
BibTeX
@article{Thamilkolunthu2026,
author = {Thamilkolunthu, Senthilkumar and Rengasamy, Hariharan and Govindasamy, Mahadevan and Muthukumarasamy, Arulmozhi},
title = {Defect detection in orbital GTAW joints of Al5083-5 wt% SiC composite using texture-oriented machine learning},
journal = {High Temperature Materials and Processes},
year = {2026},
volume = {45},
number = {1},
doi = {10.1515/htmp-2025-0138}
}
EUVIMED