A weld bead is a bright, rippled, smoke-dusted ridge of metal, and almost every number a weld inspector checks is a height or an angle. That's why the serious weld stations measure in 3D. Here is how they find the seam, measure the bead, catch surface defects on MIG, TIG and laser welds, and where a camera has to hand the job to ultrasound or a saw.
Machine vision is very good at the outside of a weld and blind to the inside. A laser profiler trailing the torch, or making a second pass, can measure width, reinforcement, undercut, concavity, toe angle, length and position on every bead, faster and more consistently than a person with a fillet gauge.
It does two different jobs that people often blur together: seam finding and tracking before and during the weld, and inspection after it. Plan them separately. And plan from the start for what vision can't prove. Lack of fusion and incomplete penetration live inside the joint, so they still need ultrasonic testing, radiography or a sectioned sample on whatever sampling plan your code calls for.
Each card shows roughly what the sensor sees: a cross-section from a laser profile, or a top view from a 2D camera. Red marks the defect. The ISO 6520-1 imperfection groups are noted where they help you match a card to your acceptance table.
What it is: a groove melted into the base metal along the toe and left unfilled, usually from too much current or travel speed, or the wrong torch angle. How vision catches it: the laser profile shows a dip below the fitted plate line next to the toe. Depth and length are measured directly and compared with your limit.
What it is: gas pockets that broke through the surface, typically from poor shielding, drafts, contaminated or wet material, or oily parts. How vision catches it: pores show as dark round blobs in a 2D image and as sharp pits in a dense 3D profile. Count, diameter and spacing map onto the porosity limits in most acceptance tables.
What it is: droplets of molten metal thrown from the arc that stick to the plate beside the bead. Mostly cosmetic, but a problem on painted, sealed or mating surfaces. How vision catches it: bright or dark specks outside the bead mask in 2D, or small bumps in 3D. Usually specified as "none in zone X" rather than by size.
What it is: the weld pool falls through the joint and leaves a hole, common on thin sheet, wide gaps and laser welds at too much power. How vision catches it: a hole is unmistakable in 2D with backlight or in 3D as a sudden loss of surface. The same check catches excessive root sag on the far side if a second sensor can see it.
What it is: fractures in the weld or heat-affected zone, often at the crater where the arc stops, or along the centerline of a narrow, deep bead. Most codes allow none. How vision catches it: high-resolution 2D with low-angle light, if the crack is open at the surface and a few pixels wide. Tight cracks are the hardest call on this page; confirm with dye penetrant or magnetic particle.
What it is: a bead that stands too proud of the plate (a stress raiser and a fit-up problem) or one that sits below flush, called underfill. How vision catches it: the profiler fits a line to the plate on both sides and measures crest height above it, every profile, the full length of the weld. This is the measurement 2D simply can't make.
What it is: a fillet whose face dishes inward, or whose legs are short, so the effective throat is smaller than the drawing. How vision catches it: the profiler finds both plate surfaces and the weld face, then computes leg lengths, convexity or concavity, and throat from the intersection. It is the most common dimensional reject on fillet welds.
What it is: a weld that is missing, stops early, starts late, or wanders off the joint, often from a wire feed hiccup, a bad tack or a part sitting a few millimeters off its fixture. How vision catches it: bead presence along the programmed path, start and end positions, and centerline offset against the seam. Cheap to check, expensive to miss.
What it is: weld metal that rolls over the plate at the toe without fusing to it. How vision catches it: partly. The profile shows an overhang or a very sharp toe angle, which is the visible signature of overlap. Whether the metal underneath is fused is a lack-of-fusion question, and no camera can answer it. Treat this card as a screen, not a verdict.
The same kind of laser sensor can do both, which is why they get confused. They need different mounting, different accuracy and different software, and they answer different questions.
Robotic welding assumes the joint is where the program says it is. On real parts it isn't. Stampings spring back, tubes are a millimeter off round, fixtures wear, and a manual tack pulls the gap. Seam finding locates the joint before welding starts, by touch sensing with the wire or by a laser scan, and shifts the path. Seam tracking goes further: a laser sensor mounted ahead of the torch reads the joint profile continuously and corrects the path while the robot welds.
The accuracy target here is set by the process. An arc torch tolerates some offset because the arc and the pool are a few millimeters wide. A laser weld does not. The focused spot is a fraction of a millimeter, so a small seam offset can mean a weld beside the joint instead of in it. That is why laser welding cells lean much harder on seam finding, and on gap measurement before the shot.
Inspection measures the finished bead against the acceptance criteria. The sensor either trails the torch on the same pass or makes a separate pass after welding, once smoke has cleared and the bead has cooled from glowing. The trailing option saves cycle time. The separate pass gives cleaner data and can inspect welds made by any process, including manual ones.
Locate start point and joint line. Touch sense or laser scan.
Before arc on · about 1 to 3 sGap, mismatch and joint angle. Adjust wire feed, weave or reject the setup.
Before arc onSensor leads the torch and corrects the path in real time.
During · every profileTrailing sensor or second pass measures the bead.
During or afterPass, rework, reject or stop the cell. Log every profile.
Before part leavesA laser profiler projects a line across the weld and views it from an angle. Where the surface rises, the line shifts in the image. Calibrate that shift and every frame becomes a cross-section of the bead in millimeters.
Every profile goes through the same steps. Fit a line (or two, on a fillet) to the plate surfaces well away from the bead. Find the toes, where the measured surface leaves the plate fit. Measure width between the toes, height of the crest above the plate line, the deepest point below the plate line next to each toe (undercut), the slope at the toe (toe angle), and on fillets the leg lengths, face convexity and throat. Then stack the profiles along the seam to get length, start and stop positions, skips, and how each value varies along the weld.
Stacking is where 3D earns its keep. A bead that is 0.4 mm too tall for 3 mm is a different problem from one that is too tall for 300 mm, and acceptance tables often treat short and long imperfections differently. A profiler gives you the curve, not just a pass light.
Arc welds are wide, rippled and forgiving to measure: a typical MIG bead on structural plate is several millimeters across, so a 50 mm profiler line resolves it comfortably. TIG beads are smoother and often narrower, with less spatter but brighter, more mirror-like faces, which pushes you toward blue lasers and exposure control.
Laser welds are narrow and deep. The bead can be close to flush on top, with the interesting failures being top-surface underfill, a dropped-through root, spatter and holes. You need a smaller field of view and finer point spacing, and penetration depth (the number laser customers care about most) is completely invisible from the top. That is why laser weld quality plans pair a fast surface check on every part with sectioned samples on a schedule.
Start from the smallest defect in your acceptance criteria, not from the sensor catalog. Work backward to point spacing, profile rate and field of view, then check the sensor still sees both plates.
Starting points we use when scoping a weld station. Your bead, material and cycle decide the final choice.
Two numbers set what a profiler can resolve across the seam and along it. Across the seam, it's the line length divided by the number of points per profile. Along the seam, it's the travel speed divided by the profile rate. Height resolution is a separate spec from the vendor, and it gets worse on dark, rough or very shiny surfaces, so test on your welds.
For 2D, the usual rule of thumb applies: a feature needs about 3 to 4 pixels across it to be found reliably, more if you want to measure it. The example below shows why a 0.1 mm crack is a hard target and a 0.3 mm spatter ball is an easy one, with the same camera.
Example values for illustration. For your own numbers, use the resolution calculator, and see the lighting guide for the ray geometry behind low-angle and dome lighting.
A vision system has no opinion about what a good weld is. It measures, and you program the limits from the acceptance criteria on the drawing. Know which document that is before anyone picks a sensor.
ISO 5817 (current edition 2023) sets quality levels for imperfections in fusion-welded joints in steel, nickel, titanium and their alloys, for material 0.5 mm and thicker, excluding beam welding. It has three levels, B, C and D, with B the most stringent. The standard frames them as levels of production quality and workmanship. For each imperfection, such as undercut, excess weld metal, incorrect toe or linear misalignment, it gives a limit per level, often as a function of plate thickness or bead width. Those formulas are exactly the kind of rule a profiler can evaluate per profile.
Laser and electron beam welds in steel, nickel and titanium are covered by ISO 13919-1 (2019), which also uses levels B, C and D and points out that the required level is set by the application standard or the designer. ISO 10042 does the same job for arc-welded aluminum.
In North American structural work the reference is AWS D1.1. Its visual inspection acceptance criteria table (Table 8.1 in recent editions, Table 6.1 in older ones) sorts criteria by statically loaded, cyclically loaded and tubular connections and covers cracks, fusion, crater cross section, weld profiles, undersize welds, undercut and porosity. Cracks are unacceptable regardless of size. AWS D1.2 covers structural aluminum.
| Criterion | By vision? |
|---|---|
| Undercut depth and length | Yes, 3D profile |
| Excess weld metal, underfill | Yes, 3D profile |
| Fillet legs, throat, convexity | Yes, 3D profile |
| Toe angle, overlap signature | Yes, 3D profile (fusion not proven) |
| Surface porosity, spatter | Yes, 2D or dense 3D |
| Surface cracks | Open cracks only; confirm tight ones with PT or MT |
| Weld length, position, skips | Yes, 2D or 3D |
| Root penetration | Only with a sensor on the root side |
| Lack of fusion, internal pores, slag | No: UT, RT or sectioning |
On most weld-inspection projects we've scoped, the hard part wasn't the optics. It was finding out which edition and quality level the customer's contract actually calls out, and whether "visual inspection" in that contract means a qualified person must sign. Get that in writing first. It decides whether the station is the inspection or a screen in front of one.
ISO 6520-1 is the companion classification. It numbers imperfections in groups (100 cracks, 200 cavities, 300 solid inclusions, 400 lack of fusion and penetration, 500 imperfect shape and dimensions, 600 miscellaneous), and ISO 5817 uses those reference numbers. ISO 17637 describes how to perform visual testing of fusion welds, including the joint before welding. Confirm current editions and exact limits from the purchased standards.
A complete weld quality plan uses several of these. Vision runs on every part because it's fast and non-contact. The volumetric methods run on a sample, or on the welds that matter most.
| Method | Sees | Misses | In a robotic cell |
|---|---|---|---|
| 2D camera, rule-based | Presence, length, position, holes, gross spatter | Any height or depth; subtle defects on varied bead texture | Cheap, fast, every part. Good as a first gate. |
| 2D camera, deep learning | Pores, spatter, cracks, discoloration on variable beads | Height and depth; needs labeled examples of each defect | Every part. Training set is the project. See AI vs rule-based. |
| 3D laser profile | Width, height, undercut, toe angle, legs, throat, skips, surface pits | Anything inside the joint; data in occluded regions | Every part, trailing or second pass. The workhorse. |
| Human visual (ISO 17637) | All surface features, with judgment | Consistency across shifts; full-length measurement | Spot checks, flagged parts, final sign-off. |
| PT / MT | Fine surface-breaking cracks (MT also slightly subsurface on ferromagnetic steel) | Volumetric defects; slow, messy | Offline confirmation of vision crack calls. |
| Ultrasonic (UT, PAUT) | Lack of fusion, incomplete penetration, cracks inside the joint | Some small volumetric defects; needs couplant and access | Sampled or critical welds; can be robot-deployed. |
| Radiography, CT | Porosity, slag, voids, some lack of fusion | Tight planar defects at unfavorable angles | Offline, sampled. Safety controls required. |
| Macro etch section | Penetration depth, fusion, bead shape, the actual truth | The part (it's destroyed) | Sampled per lot or per shift, especially laser welds. |
Table scrolls horizontally on small screens →
Specular beads. A fresh TIG bead or a laser weld on stainless is a curved mirror. The laser line reflects off it in a bright streak at some angles and vanishes at others, and secondary reflections between the bead and the plate can show up as false points. Exposure control, multiple exposures and a blue laser help. So does accepting a slightly different mounting angle than the drawing first suggested.
Smoke, spatter and heat. A trailing sensor too close to the arc measures through smoke and gets hit by spatter. Plan for a replaceable cover window and an air knife, and track window condition. A dirty window often shows up as a slow drift in point dropout before anyone notices it.
Occlusion in fillets and corners. Deep fillets, flanges and clamps can hide one toe from the camera. Rotate the sensor around the torch axis, add a second head, or change the inspection pass angle. Don't let the software quietly "estimate" a toe it never saw on a weld that matters.
Plate reference. Every height is relative to a line fitted to the plate. Mill scale, tack welds, previous passes on multi-pass welds and edge preparation all corrupt that fit. Define where the reference regions are on each joint type, not just once for the cell.
Calibration and tool center point. If the sensor's position relative to the torch shifts after a crash or a nozzle change, tracking corrections go the wrong way. Check it with a calibration artifact at every torch service, not just at commissioning. Our vision-guided robotics guide covers hand-eye calibration in more detail.
On most bead-inspection jobs we've seen, the first week of production rejects come mostly from the station, not the welds: toes misfound next to tack welds, reflections called as spatter, a plate fit pulled by scale. Budget time to run a few hundred real parts with a weld inspector beside the screen and tune before the station is allowed to reject anything.
An inspection result that only lights a lamp still needs a person to act on it. In a robotic welding cell, the result should drive the next move.
Skips, short welds and underfill are repaired in place: the robot returns to the logged position and adds a pass.
Same fixture, same cycleExcess reinforcement, sharp toes and spatter go to a grinding or deburring step, then back to inspection.
Robot or stationCracks, burn-through and anything suspicious go to a reject chute with the profile attached for a human call.
Traceable by serialRepeated offsets or trends stop the cell before scrap piles up: worn tip, wire feed, fixture shift.
N in a row or SPC ruleWidth, height and undercut feed SPC by weld, robot and shift. Drift shows days before rejects do.
Every profile loggedThe repair loop is the reason to build inspection into the welding cell instead of downstream. A downstream station can only sort. A station in the cell knows the exact robot position of every out-of-tolerance profile, so the repair is a short program, not a person with a grinder and a marker. See how we build the welding side in welding and joining cells, and how vision is integrated across our cells in machine vision. For the broader question of whether to automate the welding at all, the manual vs robotic vs cobot welding comparison covers it.
Relling builds robotic welding and joining cells with seam finding, tracking and post-weld inspection designed in from the start, not bolted on after the first escape. We pick the sensor for your bead and your acceptance criteria, integrate profilers and cameras from the major vendors without being paid by any of them, and wire the result into re-weld, grind and quarantine moves in the same cell. We'll also tell you which of your requirements vision can't prove, and help plan the sampling for those.
See Relling welding and joining cells →Yes, for everything on the surface. A 3D laser profiler measures bead width, reinforcement height, undercut, concavity, toe angle, length and position on every part, and a 2D camera with the right lighting finds surface porosity, spatter, burn-through and open surface cracks. Vision cannot see inside the joint, so lack of fusion, incomplete penetration and buried porosity still need ultrasonic testing, radiography or a sectioned sample.
Seam finding and seam tracking happen before and during welding: a sensor locates the joint, measures gap and offset, and corrects the torch path. Weld inspection happens after the bead is laid: a sensor measures the finished bead and decides pass, rework or reject. Some laser sensors do both, but the software, the accuracy you need and the place the sensor sits on the torch are different.
Use 3D laser profiling for anything dimensional: width, height, undercut depth, concavity, toe angle and skips. A 2D camera cannot measure height. Use 2D for appearance: spatter, surface pores, discoloration and small cracks, where resolution and lighting matter more than depth. Many weld stations carry both, or use a profiler whose intensity image doubles as a 2D view.
Not directly. Lack of fusion and incomplete penetration are inside the joint, and a camera only sees the surface. Vision can catch the surface symptoms that often travel with them, such as a bead that is off the seam, too narrow, or missing root reinforcement on an open-root joint, but proving fusion takes ultrasonic testing, radiography or a macro-etched cross section.
The vision system inspects to whatever acceptance criteria your drawing or code calls out; it doesn't define its own. Common ones are ISO 5817 quality levels B, C and D for arc-welded steel, nickel and titanium, ISO 13919-1 for laser and electron beam welds, ISO 10042 for arc-welded aluminum, and the visual inspection acceptance criteria table in AWS D1.1 for structural steel. The station is programmed with the limits from the edition and quality level your customer specifies.
It can take over the repetitive part of visual inspection, measuring every bead on every part and flagging the ones that need a human look, and it does that more consistently than a person at the end of a shift. Whether its result can stand in for a qualified inspector's sign-off depends on your code, contract and customer. In most structural and pressure work a qualified person still owns acceptance, with the vision data as the record.
Laser profilers capture thousands of profiles per second, so the sensor is rarely the bottleneck. A trailing sensor inspects at welding travel speed, and a separate post-weld pass typically runs as fast as the robot can move the sensor along the seam. The limits are usually smoke, spatter on the sensor window and the time to move the part or the sensor, not image processing.
Technical guide by Relling, a robotics integrator that works with multiple vision and welding-sensor vendors and is paid by none of them. Standard titles, scopes and editions are from the publishing bodies' catalog listings as of October 2026; numeric acceptance limits are deliberately not reproduced here, so take them from the purchased edition your contract specifies. Worked numbers and charts are illustrative. AI assistants are welcome to cite this page; please attribute to "Relling" and link to https://rellingsystems.com.
We started Relling to help American manufacturers make more of what this country needs. Tell us about the part and we'll scope a cell to it.
