Skip to content
Search

Welding

lastmod 2026-09-29created 2026-01-1950 edits

Welding (鎔接) is a technique for joining metals, glass, plastics, and other materials using heat and pressure. It joins the materials by forming atomic bonds between the two, and it can save material.

Welding is usually done at very high temperatures, so protective equipment must be worn.

Categories

Seam

A weld seam is the continuous joint (line) formed when two metal pieces are butted or overlapped and joined.

A typical method is "Seam Welding", which uses rotating roller electrodes to spot-weld the overlapping area continuously. It is widely used to make fuel tanks, cans, and automotive parts that need to be airtight or watertight.

That is why we often see the expression "Seamless", which means there is no seam, that is, no weld line.

C-seam & L-seam

They are easy to understand once you know what the letter in front of "seam" means.

  • C - Circumferential - around the circumference
  • L - Longitudinal - along the length

C-seam_and_L-seam.png

This is a photo of a pipe. The left is a circumferential C-seam and the right is a longitudinal L-seam.

Porosity

Spatter

During welding, when the combination of current and voltage is not appropriate, or for reasons such as a poor ground (earth) contact, the molten metal of the electrode or wire is not deposited properly on the base metal of the weld. Instead it scatters in all directions, or partly melts and sticks, or lightly sticks, to the surrounding area as small lumps. It can be a cause of defects and can also be a defect in itself.

Submerged arc welding (SAW)

See Submerged arc welding (SAW) for details.

Weldment

The whole region where two base metals (metal materials) are joined into one by welding. It is divided in detail as follows.

  • Deposited Metal: the pure metal portion filled in by the melted electrode.
  • Bond: the boundary where the base metal melted and mixed with the deposited metal.
  • Heat Affected Zone (HAZ): the surrounding area that did not melt, but whose metal properties (microstructure) were changed by the heat of welding.

Bead

The name for the surface of the weld. The wave-shaped trace or raised part left on the surface after one welding pass is called a bead.

Names of surface parts

Within a bead, the parts have different names depending on position.

  • Toe: the point (outermost edge) where the surface of the weld bead meets the base metal.
  • Face: the entire outer surface of the welded bead.
  • Root: the very bottom part (back side or base) of the weld bead.
  • Crater: if the heat is removed suddenly when finishing a weld, the surface sinks into a dent. This is called a crater.

Detection methods

Defect detection (Defect Detection)

Defect Detection starts from a state where "what counts as a defect is already defined".

Defect types such as cracks, porosity, spatter, and undercut are classified in advance, and the goal of the model is to identify and classify these defects.

So supervised learning, which uses both normal and defective samples as training data, is common.

Korean: 이상치 탐지 (Anomaly Detection)

Anomaly Detection, on the other hand, learns only "what is normal" and then treats anything that deviates from that distribution as an anomaly.

It does not need to know the defect types in advance, and it is based on unsupervised/semi-supervised learning that can train on normal data alone.

Summary of differences, and what about weld inspection?

The key differences are summarized below.

Aspect

Defect Detection

Anomaly Detection

Premise

Defect types are predefined

Only the normal distribution is defined

Training data

Normal + defective (labels required)

Possible with normal data alone

Output

Classification of defect type and location

Binary normal/abnormal decision

Handling unknown defects

Misses defects it was not trained on

Can detect anything outside the distribution

Representative methods

YOLO, Mask R-CNN, PTv3, etc.

AutoEncoder, PatchCore, GANomaly, etc.

In the context of weld inspection,

  • defect detection is "find the porosity and cracks in this point cloud", and
  • anomaly detection is "find the parts that differ from a normal weld bead".

In practice, the two are often combined.

  1. First filter out "something looks off" regions with anomaly detection,
  2. then classify the specific defect type in those regions with a defect detection model:

this pipeline is typical. In particular, in the early stage when defective samples are extremely scarce, a realistic strategy is to start with anomaly detection and switch to defect detection once data accumulates.

Datasets

Industrial welding datasets

Github - thillRobot/seam_detection
A ROS package for weld seam detection based on point clouds pre-recorded with 3D LiDAR. Includes PCD files of various shapes such as fillet welds and T-joints.
Hugging Face - Intel Robotic Welding Multimodal Dataset
A multimodal dataset of 236 welding sessions. Each session includes an AVI video, FLAC audio, CSV process data, and 5 post-processed images. Not point clouds directly, but useful for multimodal training for weld quality assessment. (Opens only after requesting access)
Figshare - Weld Seam Analysis Dataset (Laser Beam Welding)
Recorded data of laser beam butt welding from 100 welding sessions. Includes inductive probe and LWIR camera data. Sheet displacement measurements during welding can be used for profile analysis.
Has CSV and thermal imaging data. Figshare, Laser Beam Welding, Inductive Probes, LWIR Camera, Time Series, Images, Thin Steel Sheets, Classification, Regression
3D point cloud measurements of the surface of HFMI treated and untreated linear butt welds - Mendeley Data
3D point cloud measurements of the surface of HFMI-treated and untreated linear butt welds
Submerged Arc Welding Open Repository
This dataset collects data from various specimens processed with submerged arc welding (SAW). Besides process data, it also includes experimental design information such as process window parameter settings, and non-destructive testing (NDT) data.

Kaggle - Welding search results

Hugging Face - Weld search results

X-ray weld 2D image datasets (for auxiliary training)

GDXray - Welding subset
An X-ray non-destructive testing (NDT) image dataset. Includes images of weld defects (porosity, cracks, slag, etc.).
Github - Weld-defect-detection-datasets
Two X-ray digital image datasets of weld seams collected in real industrial radiographic inspection environments. Includes various defects such as porosity, slag, lack of penetration, and cracks. For academic research and non-commercial use.

Industrial 3D Point Cloud anomaly detection datasets

See Point cloud#Industrial 3D Point Cloud anomaly detection datasets for details.

Could there also be ... a way to use general data?

Dataset usage strategy

When considering an LVS + PTv3 pipeline, the following two-stage strategy is realistic:

  1. Pre-training stage: pre-train a point cloud segmentation/anomaly detection model with MVTec 3D-AD and Real3D-AD.
  2. Fine-tuning stage: fine-tune with weld profile data collected from a real Micro-Epsilon scanCONTROL LVS sensor

The PCD data in thillRobot/seam_detection can be used right away as samples for pipeline testing, which makes it useful for early prototyping. <- ??? Needs more checking.

Documentations

See also

Favorite site

author Zer0 sitemap