Multimodal Deep Learning Framework for Crack Segmentation in Complex Components Using Robot-Assisted Laser Thermography

  • Crack detection and segmentation in complex components are critical for maintaining the structural integrity and reliability of systems operating under extreme conditions, such as turbine blades in energy and aerospace applications. The integration of automated multimodal imaging-based non-destructive testing (NDT) with deep learning provides a promising path towards precise and automated defect characterization. In this study, a hybrid multimodal deep learning framework is proposed, combining the advantages of an unsupervised generative adversarial network (GAN) and a supervised U-Net segmentation model for comprehensive crack detection and quantification. The unsupervised multimodal GAN performs data fusion by integrating complementary features from high-resolution thermal and RGB images acquired using a robot-assisted flying laser-line thermography system. This data fusion improves the contrast and representation of surface and sub-surface cracks by leveraging spectral featuresCrack detection and segmentation in complex components are critical for maintaining the structural integrity and reliability of systems operating under extreme conditions, such as turbine blades in energy and aerospace applications. The integration of automated multimodal imaging-based non-destructive testing (NDT) with deep learning provides a promising path towards precise and automated defect characterization. In this study, a hybrid multimodal deep learning framework is proposed, combining the advantages of an unsupervised generative adversarial network (GAN) and a supervised U-Net segmentation model for comprehensive crack detection and quantification. The unsupervised multimodal GAN performs data fusion by integrating complementary features from high-resolution thermal and RGB images acquired using a robot-assisted flying laser-line thermography system. This data fusion improves the contrast and representation of surface and sub-surface cracks by leveraging spectral features across multiple imaging modalities. The GAN is trained to reconstruct crack free images and difference between the generated image and real crack image generates an error map that highlights the cracks. The unsupervised approach helps in reducing the need for manual labeled data and generalizes well across different surface conditions. The error maps from GAN are subsequently processed by a U-Net-based segmentation model trained on labeled datasets to achieve precise pixel-level crack localization and morphological estimation. The use of laser thermography induces localized heating on the component surface, providing transient thermal responses that make subtle cracks and defects visible beyond the limits of visual imaging. Experimental validation demonstrates that the proposed hybrid GAN–U-Net framework achieves significantly improved crack detection accuracy and segmentation performance compared to single modal NDE imaging, and data processing based on traditional threshold-based methods. This work underscores the potential of combining unsupervised multimodal fusion with supervised image segmentation to establish a new framework that helps in building automated, data-driven, and robot-assisted NDT systems for intelligent inspection and structural health monitoring of industrial components.zeige mehrzeige weniger

Volltext Dateien herunterladen

  • 1_NDE2025_KP_talk.pdf
    eng

Metadaten exportieren

Weitere Dienste

Suche bei Google Scholar Anzahl der Zugriffe auf dieses Dokument
Metadaten
Autor*innen:Sruthi Krishna Kunji PurayilORCiD
Koautor*innen:Philipp Daniel HirschORCiD, Julien Lecompagnon, Mathias Ziegler
Dokumenttyp:Vortrag
Veröffentlichungsform:Präsentation
Sprache:Englisch
Jahr der Erstveröffentlichung:2025
Organisationseinheit der BAM:8 Zerstörungsfreie Prüfung
8 Zerstörungsfreie Prüfung / 8.3 Thermografische Verfahren
DDC-Klassifikation:Technik, Medizin, angewandte Wissenschaften / Ingenieurwissenschaften / Ingenieurbau
Freie Schlagwörter:Data Fusion; Infrared Thermography; Laser; Multimodal Imaging; NDE 4.0; Non-destructive Testing
Themenfelder/Aktivitätsfelder der BAM:Infrastruktur
Infrastruktur / Leichtbau
Veranstaltung:NDE 2025
Veranstaltungsort:Mumbai, India
Beginndatum der Veranstaltung:11.12.2025
Enddatum der Veranstaltung:13.12.2025
Verfügbarkeit des Dokuments:Datei im Netzwerk der BAM verfügbar ("Closed Access")
Datum der Freischaltung:06.01.2026
Referierte Publikation:Nein
Eingeladener Vortrag (wissenschaftliche Konferenzen):Nein
Einverstanden
Diese Webseite verwendet technisch erforderliche Session-Cookies. Durch die weitere Nutzung der Webseite stimmen Sie diesem zu. Unsere Datenschutzerklärung finden Sie hier.