Machine Vision, Calibration & Weld Path Correction
Adding a camera to a laser welding system sounds like it should solve a basic problem:
Find the joint, tell the scanner where it is, and put the weld there.
And in principle, that’s exactly what machine vision can do.
Researchers have demonstrated camera systems that look through the same optical path as a laser scanner, measure the actual location of a weld joint, and use that information to correct the programmed welding trajectory. [1] Commercial scanner systems now use the same basic idea for applications such as hairpin welding, power-electronics connections, and battery busbars. [3][4]
But there is an important catch.
Seeing the joint and knowing exactly where to put the laser are two different problems.
That difference is where things can get interesting for the process engineer or technician standing in front of a machine wondering why a perfectly good camera image is producing an off-center weld.
Fact: The Camera Does Not Control the Laser in Pixels
A vision system starts with an image.
The software may determine that the center of a joint is at a particular X and Y pixel location.
But the scanner does not work in camera pixels.
It has its own coordinate system, optical geometry, and correction data. The system therefore needs a mathematical relationship between:
camera coordinates → scanner coordinates → actual position on the workpiece
That relationship has to be calibrated.
This is not unique to welding. Calibration of galvanometer scanning systems has been studied extensively because optical distortion and scanner geometry can cause the actual beam position to differ from the position predicted by a simple linear relationship. [2][5][6]
So the first troubleshooting question should not be:
“Can the camera see the joint?”
It should be:
“When the camera says the joint is here, does the laser agree?”
Problem 1: The Calibration Works in the Center but Not at the Edges
This is one of the more useful symptoms to recognize.
Suppose you check the system near the center of the scanner field and everything looks good.
Move toward the edge, and the weld gradually shifts away from the joint.
That is probably not a simple X-Y offset.
A constant offset would move everything by about the same amount. An error that changes with field position points more toward scale, rotation, scanner-field correction, optical distortion, or an incomplete camera-to-scanner calibration.
Research on coaxial imaging through laser-focusing optics has shown that the optical system can introduce positioning distortion that must be characterized and corrected. [2] The calibrated-mosaicking work by Braunreuther and colleagues likewise corrects perspectively deformed and distorted images so metric measurements can be made over a scanner working field. [1]
Technician’s clue
Check more than one point.
A calibration that is perfect at the center tells you very little about the corners.
I would want verification points distributed across the area actually used by the process—center, sides, and corners where practical.
Record both the commanded location and measured error.
The error pattern can tell you as much as the error size.
Problem 2: The Camera Found Something—But Was It the Joint?
This is particularly important with copper, aluminum, machined surfaces, and other reflective parts.
Machine vision depends heavily on contrast and illumination.
A shiny edge, machining mark, reflection, contamination, or changing surface finish can become more visually prominent than the feature you actually want to locate.
Commercial welding-vision systems address this with process-specific lighting and configurable image-processing tools. Blackbird, for example, describes bar or ring illumination along with functions such as edge extraction, shape recognition, tolerance windows, and conditional processing in its ScaVis system. [4]
That suggests a simple troubleshooting step that is sometimes overlooked:
Look at the image the algorithm is actually using.
Don’t just look at the resulting X-Y coordinates.
If the detected feature jumps around while the physical part stays stationary, recalibrating the scanner probably won’t solve the problem.
The vision measurement itself may be unstable.
Problem 3: The Part Moved in Z, Not X or Y
Another common trap is assuming every apparent positioning error is an X-Y problem.
The calibration is normally established relative to a defined working plane.
If the part height changes because of fixture variation, part geometry, stacking tolerances, or distortion, the relationship between the camera, optics, scanner, and workpiece can also change.
That concern becomes particularly obvious with an external camera viewing the part at an angle, because a height change can appear as a lateral position change.
Coaxial viewing can reduce some of those parallax issues because the camera views the workpiece through the scanner optical path. SCANLAB, for example, describes using a beam splitter to direct light returning from the working plane toward the camera while allowing the processing laser to continue toward the scan system. [3]
But coaxial does not mean calibration-free.
The optics can still introduce distortion, and the working plane still matters. [1][2]
Technician’s clue
If a flat calibration coupon works but the production part doesn’t, measure the actual part height before changing the program.
The problem may be telling you something about the fixture or component—not the vision software.
Problem 4: Someone Changed Something
Calibration describes a physical relationship.
Move one of the components in that relationship, and you may no longer have the same calibration.
Think about what happens during normal machine life:
A protective window is replaced.
The scanner head is removed for service.
A camera mount is adjusted.
An optical component is changed.
The fixture is repaired.
A robot tool frame is modified.
The processing head takes a minor hit.
None of these changes necessarily produces an obvious machine fault.
The machine may still run beautifully.
It may simply run beautifully in the wrong place.
This is why I believe camera-to-laser verification belongs in preventive maintenance and post-service procedures—not just initial commissioning.
Problem 5: We Corrected the Vision but Never Verified the Scanner
There are really two positioning questions.
Can the camera measure where the part is?
And:
Can the scanner put the laser where it is commanded?
Those are separate questions.
NIST researchers have demonstrated an in-situ method for checking galvanometer positioning using a coaxially aligned camera and a dimensional reference artifact. Their work emphasizes an important point: galvo positioning accuracy itself needs to be measured and calibrated. [5]
So before blaming vision, verify that the scanner coordinate system is correct.
If you command several known positions across the working field, does the beam actually reach them?
A beautiful vision calibration on top of an incorrect scanner calibration will still produce an incorrect weld.
A Practical Way to Troubleshoot It
When I approach problems like this, I prefer to work through the chain rather than immediately changing offsets.
1. Verify the image
Can you clearly see the feature you intend to detect?
Take multiple images without moving the part.
Does the measured feature position stay essentially constant?
If not, start with lighting, exposure, focus, filtering, or the detection algorithm.
2. Verify the scanner
Using the manufacturer’s recommended procedure, check known scanner positions throughout the portion of the field used by the process.
Don’t check only the center.
3. Verify camera-to-scanner registration
Put a known target or reference feature at several locations.
Have the camera locate it.
Then verify where the laser would actually be placed relative to that feature.
This is where we are testing whether:
what the camera sees = where the scanner goes
4. Look at the error pattern
This is where technicians can save a lot of time.
If the error is approximately the same everywhere, think offset or datum.
If the error grows toward the edges, think field calibration, scale, rotation, or optical distortion.
If the detected position jumps around with a stationary part, think vision or illumination.
If the problem follows part height, think working plane or geometry.
If everything changed after maintenance, start with what was touched.
Don’t Let Vision Hide a Bad Part
Automatic path correction has another side.
Just because the camera can move the weld path doesn’t mean it should correct every error it sees.
Blackbird describes using tolerance windows and operator interaction when it detects component misalignment. [4] That is an important concept.
A vision correction should normally have a validated operating range.
For example:
The joint moved 0.15 mm?
That may be normal part variation.
The joint moved 4 mm?
That could indicate a missing component, incorrect assembly, a broken locator, the wrong part, or a fixture problem.
Automatically shifting the weld 4 mm may make the machine continue running—but it could also hide the real problem.
Sometimes the best thing a vision system can do is say:
“Something is wrong. Don’t weld this.”
Machine Vision Is More Than a Camera
I think that is the bigger lesson.
A camera is only one component of a vision-guided laser system.
A successful application combines:
Vision — Where is the feature?
Calibration — How does that camera measurement correspond to the laser coordinate system?
Scanner control — Can the beam actually reach the commanded position?
Process limits — How much correction are we willing to accept?
Diagnostics — If the system fails, can a technician determine which part of that chain failed?
When all of those pieces work together, vision can make a laser process much more tolerant of normal manufacturing variation.
When they don’t, you can end up with something much more frustrating:
A camera that correctly sees the joint—and a laser that consistently welds beside it.
That’s why adding machine vision shouldn’t be the end of the discussion.
It should be the beginning of a better positioning strategy.
References
[1] S. Braunreuther, V. Hammerstingl, M. Schweier, G. Theodossiadis, G. Reinhart, and M. F. Zaeh, “Welding Joint Detection by Calibrated Mosaicking with Laser Scanner Systems,” CIRP Journal of Manufacturing Science and Technology, Vol. 10, 2015, pp. 16–23. DOI: 10.1016/j.cirpj.2015.03.002.
Article / DOI information
[2] J. O. Entzinger, D. Iakovou, R. G. K. M. Aarts, and J. Meijer, “Improving Coaxial Measurements in Laser Welding by Correcting Distortions of a Laser Focus Lens with a Wide Field of View,” Journal of Laser Applications, Vol. 19, No. 3, 2007, pp. 141–148. DOI: 10.2351/1.2567495.
University of Twente publication page
[3] SCANLAB GmbH, Camera Adapter — Process Observation Through a Galvanometer Scan Head.
SCANLAB Camera Adapter
[4] Blackbird Robotersysteme GmbH, ScaVis Camera System — Coaxial Camera System for Laser Welding and Seam-Position Correction.
Blackbird ScaVis Camera System
[5] H. Yeung, B. M. Lane, M. A. Donmez, and S. P. Moylan, “In-situ Calibration of Laser/Galvo Scanning System Using Dimensional Reference Artefacts,” CIRP Annals, Vol. 69, 2020. DOI: 10.1016/j.cirp.2020.03.016.
NIST publication and downloadable paper
[6] S. Cui, X. Zhu, W. Wang, and Y. Xie, “Calibration of a Laser Galvanometric Scanning System by Adapting a Camera Model,” Applied Optics, Vol. 48, No. 14, 2009, pp. 2632–2637. DOI: 10.1364/AO.48.002632.
Optica publication page