Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Deep Learning for Joint Gap Width Classification and Tack Weld Detection in Laser Beam Welding
University West, Department of Engineering Science, Division of industrial automation. (KAMPT)ORCID iD: 0000-0002-8018-6145
University West, Department of Engineering Science, Division of industrial automation. (KAMPT)ORCID iD: 0000-0001-5734-294X
University West, Department of Engineering Science, Division of industrial automation. (KAMPT)ORCID iD: 0000-0002-8771-7404
University West, Department of Engineering Science, Division of computer engineering and computer science. (KAMPT)
Show others and affiliations
2024 (English)In: New Trends in Signal Processing (NTSP), IEEE, 2024, p. 1-6Conference paper, Published paper (Refereed)
Abstract [en]

Laser Beam Welding (LBW) requires precise control to ensure high-quality welds. Accurate classification of joint gap widths and detection of tack welds are crucial for optimizing the process and enhancing product reliability.

This study investigates the application of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to classify instant joint gap widths and detect the presence of tack welds during welding. The goal is to facilitate adaptive joint gap bridging in robotized and autogenous butt joint welding. Sequences of images resembling a time series were captured during welding of prepared workpieces with varying joint gap widths along the joint line.

The results demonstrate that CNNs significantly outperform RNNs, achieving over 99 percent classification accuracy in both validation and test datasets, and 96 percent accuracy under conditions of substantial noise. These findings underscore the potential of CNNs in enhancing the precision and adaptability of welding automation. However, challenges remain in generalizing the CNN model to diverse and noisy operational environments.

Place, publisher, year, edition, pages
IEEE, 2024. p. 1-6
Keywords [en]
deep learning, convolutional neural network, CNN, recurrent neural networks, RNN, machine vision, image classification, robotized welding, laser beam welding, butt joints
National Category
Manufacturing, Surface and Joining Technology
Research subject
Production Technology
Identifiers
URN: urn:nbn:se:hv:diva-22569DOI: 10.23919/ntsp61680.2024.10726306ISI: 001359397200018Scopus ID: 2-s2.0-85210006086OAI: oai:DiVA.org:hv-22569DiVA, id: diva2:1908896
Conference
2024 New Trends in Signal Processing (NTSP), 16-18 Oct.2024
Funder
Knowledge FoundationAvailable from: 2024-10-29 Created: 2024-10-29 Last updated: 2025-03-26
In thesis
1. Dynamic beam shaping with a deformable mirror for control of high power laser processes
Open this publication in new window or tab >>Dynamic beam shaping with a deformable mirror for control of high power laser processes
2024 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

Laser metal fusion is widely used in production technology to manufacture parts, as in welding, cladding, and additive manufacturing. One development of interest to laser metal fusion is the recent advancements in deformable mirror technologies that allow them to withstand multi-kilowatt laser power. This enables the generation of various beam shapes and rapid switching between these within 10 milliseconds.

This thesis investigates the application of deformable mirror-based beamshaping and computer vision techniques to improve laser beam welding and directed energy deposition using laser and wire. Additionally, methods including analysis of melt pool images, cross-cut micrographs, surface topology measurements has been used to evaluate and characterize the results.

An industrially relevant challenge addressed is variation in the joint gap width during autogenous butt welding, which often results in defects such as incomplete groove filling and root concavities. Two near-elliptical beam shapes, with major axes aligned either along or transverse to the welding direction, were studied for gap bridging. Both beam shapes were capable of bridging gaps up to 0.6 mm wide during welding of 2.0 mm thick steel plates. It was shown that the beam shaping made it possible to meet higher quality levels in the ISO13919-1 standard compared to a ordinary circular static beam. The beam with the major axis along the welding direction performed even better in reducing defects compared to the one aligned transversely.

When adaptive beam shaping was used for gap bridging in butt joints it resulted in an 80% reduction in workpiece distortion, demonstrating its effectiveness in accommodating varying process conditions. Additionally, computer vision algorithms, including adaptive Kalman filters and convolutional neural networks, proved successful in gap width estimation, classification, and tack weld detection under varying conditions during welding.

In directed energy deposition using laser and wire, three near-elliptical beamshapes with different orientations relative to the deposition and wire feeding directions were examined. The beam with the major axis transverse to the deposition direction produced a narrower, higher bead, while the beam aligned with the feedstock wire direction (about 35◦ relative to deposition direction) yielded a wider, lower bead with better melting efficiency.

These findings presented in this thesis underscore the potential of dynamic beam shaping, especially when integrated with robust computer vision techniques, to enhance process control, making high power laser processing more reliable and efficient.

Abstract [sv]

Lasersmältning av metall är en produktionsteknik som används for att tillverka komponenter, till exempel genom svetsning, påläggssvetsning eller additiv tillverkning. En utveckling som är intressant för lasersmältning är de senaste framstegen inom formbar spegelteknik gör att denna typ av spegeloptik kan klara av en lasereffekt på flera kilowatt, samt frambringa olika strålformer (beam shapes) och snabbt växla mellan dessa inom 10 millisekunder.

Denna avhandling undersöker ett system med en deformerbar spegel för laserstrålformning i kombination med datorseendeteknik (computer vision) för att effektivisera lasersvetsning och additiv tillverkning. Dessutom har metoder som analyserar bilder på smältan i processen, bilder på den resulterande mikrostrukturen i tvärsnitt samt topologimätningar använts för att utvärdera och karakterisera resultaten.

En industriellt relevant utmaning som undersökts är variationen i fogspaltbredd under stumfogsvetsning utan tillsatstråd, vilket ofta resulterar i defekter såsom svetsdiken. Två elliptiska laserstrålformer, med huvudaxlar riktade antingen längs med eller tvärs mot svetsriktningen, studerades med avseende

på förmågan att överbrygga fogspalten.Båda strålformerna kunde överbrygga en fogspalt på upp till 0,6 mm vidsvetsning av 2,0 mm tjocka stålplåtar. Det visade sig att strålformningen gjorde det möjligt att uppfylla högre kvalitetsnivåer enligt ISO13919-1-standarden jämfört med en vanlig cirkulär statisk laserstråle. Laserstrålen med huvudaxelnl ängs svetsriktningen presterade ännu bättre när det gällde att minska defekterj ämfört med den som var på tvären.

Den adaptiva strålformningen som används för överbryggning av spalter i stumfogar resulterade i en 80-procentig minskning av arbetsstyckets deformation, vilket demonstrerar dess effektivitet när det gäller att hantera varierande processförhållanden. Dessutom visade algoritmer för datorseende bland annat med ett adaptivt Kalman-filter eller ett neuralt nätverk, vara framgångsrik när det gäller att mäta, klassificera och detektera häftsvetsar och varierationer under svetsning.

I samband med laserbaserad additiv tillverkning undersöktes tre elliptiskal aserstrålformer med olika orienteringar i förhållande till deponerings- och trådmatningsriktningarna. Strålen med huvudaxeln tvärs mot deponeringsriktningen gav en smalare, högre deponering, medan laserstrålen i linje med tillsatstrådriktningen (cirka 35◦ i förhållande till deponeringsriktningen) gav en bredare, lägre deponering med bättre smälteffektivitet.

Resultat påvisar tydligt potentialen hos dynamisk laserstrålformning moten bättre processtyrningen, särskilt när den integreras med robust datorseende i ett återkopplat system, något som bidrar till att göra högeffekt-laserprocesser mera tillförlitliga och effektiva.

Place, publisher, year, edition, pages
Trollhättan: University West, 2024. p. 72
Series
PhD Thesis: University West ; 69
Keywords
Laser beam welding; Laser beam shaping; Deformable mirror; Process monitoring; Adaptive beam shaping; Adaptive Kalman filter; Computer vision (CV), Machine learning (ML); Convolutional neural networks (CNNs); Butt joint; Gap bridging; Directed energy deposition laser wire (DED-LB/w), Lasersvetsning; Laserstrålformning; Deformerbar spegel; Processkontroll; Processövervakning; Adaptivt Kalman-filter; Adaptiv strålformning; Datorsyn; Datorseende (CV); Maskininlärning (ML); Konvolutionella neurala nätverk (CNN); Butt joint; Överbryggning av fogspalt
National Category
Manufacturing, Surface and Joining Technology
Research subject
Production Technology
Identifiers
urn:nbn:se:hv:diva-22552 (URN)978-91-89969-05-6 (ISBN)978-91-89969-04-9 (ISBN)
Public defence
2024-12-09, F314, Gustava Melins gata, Trollhättan, 10:00 (English)
Opponent
Supervisors
Note

Following papers are not included in the electronic thesis:

Paper 3 and 4 are to be submitted.

Paper 5  has no permission to be included in the electronic thesis

Paper 6 is under review.

Available from: 2024-11-11 Created: 2024-10-31 Last updated: 2024-11-13

Open Access in DiVA

No full text in DiVA

Other links

Publisher's full textScopus

Authority records

Mi, YongcuiSikström, FredrikNilsen, MorganAncona, Antonio

Search in DiVA

By author/editor
Mi, YongcuiSikström, FredrikNilsen, MorganAncona, Antonio
By organisation
Division of industrial automationDivision of computer engineering and computer science
Manufacturing, Surface and Joining Technology

Search outside of DiVA

GoogleGoogle Scholar

doi
urn-nbn

Altmetric score

doi
urn-nbn
Total: 130 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf