A structural adhesive bead is a part you can't inspect once the joint closes. This is how we check beads while they can still be fixed: what to measure, when to use 2D or 3D, why sensors around the nozzle changed the job, and how a robot repairs a gap before the part moves on.
Inspect every millimeter of the bead against its path, while it's still wet. 2D (ideally with a UV-fluorescent adhesive) covers presence, gaps, width and position; 3D laser profiling adds height, cross-section area and volume, and works on black-on-black. Sensors mounted around the nozzle inspect during dispensing and locate defects in path coordinates.
That location is what makes the result useful: the robot can go back and re-dispense a skip inside the open time, and the width and area trends tell you when flow, temperature or speed is drifting. The parts that should be rejected go to a lane, not back into the flow.
Top views, or a section where only 3D can see the problem. Most have a process cause you can trend.
What it is: No material: an air pocket, a partly blocked nozzle, a starved pump or a dropped dispense signal.
How vision catches it: Any method sees it. What matters is the minimum gap length, which sets resolution along the path.
What it is: The bead narrows without breaking: flow dipped, the robot sped up, or viscosity rose as the material cooled. A neck-down is often the warning before a skip.
How vision catches it: Width is measured at every slice and compared to a minimum. Trend it, because a slow drift in width usually means a flow or temperature problem, not a random event.
What it is: Too much adhesive in one place from a slowdown, a pressure spike or a dwell. It squeezes out onto visible or functional surfaces.
How vision catches it: A maximum width or area limit along the path. Squeeze-out after assembly is a separate check.
What it is: Right size, wrong place: part mislocated, bent nozzle, drifted TCP or worn fixture.
How vision catches it: Measure the centerline against the taught path, or better against a part edge the same sensor sees.
What it is: The bead is the right width from above but flat: wrong nozzle, nozzle too high, low viscosity, or the material slumped. Fine in 2D, short on adhesive in the joint.
How vision catches it: Only 3D sees it. A laser profile gives height and cross-section area at each slice, and area times spacing gives volume. 2D width checks will pass this part.
What it is: Air trapped in the material breaks the surface as craters and pinholes, or hides inside the bead. Bad pump priming, cartridges loaded with air, and two-part mixing are the usual causes.
How vision catches it: 2D sees surface craters if the light rakes the bead; 3D sees them as local dips. Neither sees voids fully inside. Fix the material handling and validate with sectioned samples.
What it is: The pump and the robot don't start and stop at exactly the same moment. Material piles up at the start while the robot accelerates and thins or strings at the end.
How vision catches it: Inspect the first and last 10 to 20 mm with their own limits, and inspect the overlap where a closed loop meets its own start. Tuning the dispense on and off timing usually fixes it.
What it is: Strands pulled from the nozzle, or spatter, landing on sealing faces, threads or visible surfaces.
How vision catches it: Look at the area outside the corridor, not just the bead. A keep-out zone check around the bead catches strands that a bead-only tool ignores.
What it is: Robots slow at corners while flow often stays constant, so material piles up. Speed-tracking flow can thin out instead.
How vision catches it: Look at width and area by path segment, not as one average. Corners usually need their own limits or a flow table tied to robot speed.
Start from what the joint needs, not from what the sensor can report. A structural bond needs enough adhesive in the right place to wet both surfaces after the parts close, so volume and position matter most. A formed-in-place gasket needs a continuous bead of the right height to compress correctly. A sealer on a seam needs coverage over the seam with no pinholes. A potting or tacking dot needs presence and size.
Then write the tolerances as limits along the path: minimum and maximum width, minimum area (or height), maximum offset from the path or from a part edge, and maximum gap length (often zero). Give starts, stops, corners and the closing overlap their own limits, because they behave differently from the straight runs.
Which of these you need depends on the joint, not the sensor.
A 2D camera sees the bead as a band of different brightness from the part. When that contrast is good (a white sealer on gray e-coat, a black urethane on glass frit, a colored adhesive on bare aluminum), 2D handles presence, gaps, width and position at high speed, and a single area camera can inspect a whole small part in one image.
When contrast is poor, the usual fix isn't a better camera. It's a fluorescing adhesive. Many adhesive and sealant makers offer grades with a dye that glows in visible light under UV, and a 365 nm black-light LED ring with a long-pass filter on the lens turns a clear bead on shiny metal into a bright stripe on black. It's cheap, it's fast, and it removes most lighting arguments.
What 2D can't do is see height. A slumped bead and a full one look the same from above. If the joint depends on volume, or the bead and part are the same color with no fluorescing option (black sealer on black plastic), go to 3D.
On most bead jobs we've seen go badly in 2D, the failure was glare: a wet, glossy bead on a machined surface throws specular highlights that move with the bead's curvature, and the edge finder locks onto the highlight instead of the bead edge. Diffuse light, a polarizer pair, or a fluorescing grade fixes it. More exposure doesn't.
A laser line profiler projects a line across the bead and views it from an angle. Height shifts the line sideways in the image, and triangulation turns that shift into a height profile. Each profile gives width, height and cross-section area at one point on the path; a sequence of profiles gives the whole bead. Profilers such as Keyence's LJ-X series and Cognex's In-Sight 3D sensors are often used post-dispense, moved along the bead by the robot or by a gantry.
Because 3D measures shape, it doesn't care about color: black on black, clear on clear and glossy on matte all work, within the limits of how the laser scatters off each surface. That's the main reason 3D has taken over structural adhesive and body-in-white sealing.
Two numbers set what you can see. Profile spacing along the path is robot speed divided by profile rate, and a gap needs to span at least two or three profiles to be detected reliably. Height resolution depends on the pixel footprint along the laser line and the triangulation angle.
The flow cross-check is useful: if the pump says it delivered 1.26 mL/s and the profiler measures 1.0, either material is going somewhere else or the substrate fit is wrong.
Post-dispense inspection tells you a part is bad. In-process inspection tells you where, within milliseconds, while the adhesive is still inside its open time.
The obvious way to inspect a bead is to lay it, then scan it with a camera or profiler. That works and is often the right first station. It adds cycle time (a second pass along the path), needs its own path program, and finds defects after the adhesive has started to skin over.
The alternative mounts the sensors on the dispense head, around the nozzle. A single trailing sensor works only if the nozzle always moves in the same direction relative to it, which a 3D robot path almost never does. So commercial systems use several. Coherix's Predator3D puts four 3D laser sensors around the nozzle for 360° coverage and reports width, height, volume and location in real time; Coherix says it samples 400 profiles per second per laser. ISRA VISION's QUISS line (now part of the Atlas Copco group) offers RTVision.t for 2D in-process monitoring of width, position and continuity, with an automatic gap repair function, and RTVision.3d for laser-triangulation 3D inspection.
| Method | Measures | Strengths | Limits |
|---|---|---|---|
| 2D camera, post-dispense | Presence, gaps, width, position | Fast, simple, one image for small parts | No height or volume; needs contrast; glare on glossy beads |
| 2D with UV-fluorescent adhesive | Same as 2D, much higher contrast | Cheap fix for clear or low-contrast beads; tolerant of shiny parts | Needs a fluorescing grade; filter and UV light design; still no height |
| 3D laser profile, post-dispense | Width, height, area, volume, position | Color-independent; true volume; good for structural beads | Extra pass adds cycle time; occlusion on deep flanges; reflective parts need care |
| In-process at the nozzle (2D or 3D) | As above, while dispensing | No added cycle time; defect location in path coordinates; enables repair | Larger dispense head; nozzle-change calibration; higher system cost |
| Flow monitoring / weighing (non-vision) | Total material dispensed | Catches pump and supply problems; simple | Says nothing about where the material went; misses skips that average out |
Table scrolls horizontally on small screens →
Deep learning has a smaller role here than in cosmetic inspection. Beads are geometric, and the tolerances are numbers in a drawing, so measurement tools fit the job. Where AI helps is classifying surface conditions (contamination, strings, squeeze-out after assembly) and coping with busy backgrounds in 2D. For the general trade-off, see deep learning vs rule-based vision; for the 3D sensors themselves, see 3D vision cameras.
The substrate isn't flat. Volume is area above the substrate, and the substrate on a stamped flange is curved, stepped or wavy. If the software fits the substrate badly, a perfect bead reads as low volume on one side of a step. Ask how the substrate line is fit, and test on the worst part geometry.
Shiny parts confuse lasers. Machined aluminum and glass reflect the laser line in ways that create double lines and dropouts. Sensor angle, laser power and exposure settings matter more than the brand.
Occlusion. Deep channels, flange walls and clamps hide one side of the bead from a sensor looking at an angle. Check every segment of the path for line of sight before committing to a mounting.
Material changes. Viscosity moves with temperature and age, so a bead that's perfect on a cold Monday is wide by Wednesday afternoon. The inspection will catch it; the fix is temperature control on the material and nozzle, and a flow loop.
Tolerances nobody owns. Engineering specified "continuous bead" and nothing else. Without numeric limits per segment, every alarm becomes an argument. Agree on the limits with design and quality before tuning.
The most common gap we see isn't a gap in the bead. It's a gap in the data: the inspection system flags a skip, an operator acknowledges it, the part goes on, and nobody can say afterward which serial it was. Tie every bead result and every repair to the part serial, or the inspection is only half done.
Starts and stops cause a large share of rejects on most lines we've looked at. Tuning the dispense on and off delays against the robot's acceleration, and planning the closing overlap on loops, often removes more defects than any change to the inspection.
In a robotic cell the inspection result is an instruction, not a light. With defect positions in path coordinates, the robot can return to a skip, back up by a few millimeters, re-dispense to a few millimeters past the end, and re-inspect that segment, all inside the open time and usually inside the cycle. Repair limits come from the adhesive data sheet and joint testing: how long after dispensing a repair still bonds, the longest gap you'll repair, and how many repairs a part may have.
The second loop is slower. Width and area trends feed back to the dispense parameters (pump speed, pressure, or robot speed) in small, logged steps, and alarm when the correction hits a limit. A rising repair count is a maintenance signal: a nozzle wearing, a filter clogging, a heater failing.
Everything that can't be repaired goes to a reject or rework lane, with its serial and its bead data attached. For structural parts, the bead record (including any repairs) belongs in the part's traceability history.
Relling builds dispensing cells this way: the same robot dispenses, the vision system inspects, and the cell decides whether to repair, reject or adjust. See dispensing and sealing cells and the machine-vision workcell.
Relling builds robotic workcells with machine vision built in. On dispensing jobs that means a robot laying the bead, 2D or 3D inspection matched to the joint, and a cell that repairs what it can, rejects what it can't and trends the rest. We integrate vision from several vendors and aren't paid by any of them, so the sensor is picked on your parts, your material and your cycle time. Send us a drawing of the joint and a few good and bad parts.
See Relling dispensing and sealing cells →A camera or laser profiler measures the bead along its whole path against tolerances: continuity, width, position relative to the path or a part edge, and with 3D sensors also height, cross-section area and volume. It can run after dispensing or during it, with sensors mounted around the nozzle.
Use 2D for presence, gaps, width and position when the bead contrasts with the part, or can be made to with a UV-fluorescent adhesive. Use 3D laser profiling when height or volume matters, as with structural adhesives and gaskets, or when bead and part are the same color, such as black on black.
In-process inspection measures the bead while it is being laid, with sensors mounted around the dispensing nozzle so that one of them always looks at the fresh bead behind the nozzle whichever way the robot moves. Coherix, for example, offers Predator3D with four 3D laser sensors around the nozzle, and ISRA VISION's QUISS RTVision.t does 2D in-process monitoring. The defect is located in path coordinates while the adhesive is still wet enough to repair.
Yes, for many applications. The inspection reports where along the path a skip occurred, the robot re-dispenses that segment with a short overlap at each end, and the repair is re-inspected. It only makes sense inside the adhesive's open time, below a size limit and for a limited number of repairs per part.
A 3D laser profiler captures cross-section profiles across the bead at a fixed spacing along the path. The area under each profile above the substrate is the bead cross-section area, and summing area times profile spacing along the path gives volume. For example, a 6.28 square millimeter cross-section measured every 0.5 mm over a 500 mm bead gives about 3,140 cubic millimeters, or 3.14 milliliters.
Many clear adhesives are almost invisible to a camera on metal or plastic. Fluorescing grades contain a dye that glows in visible light under UV, typically 365 nm black-light LEDs, and a filter on the lens blocks the reflected UV, so the bead shows bright on a dark background.
Only bubbles that break or deform the surface: 2D sees craters and pinholes, 3D sees dips in height or area. Voids fully inside the bead are invisible to both. If they matter, fix the material handling and validate with sectioned samples, X-ray or ultrasonic testing.
Technical guide by Relling, a robotics integrator that works with several machine-vision vendors and is paid by none of them. Vendor products and figures (Coherix, ISRA VISION / QUISS, Keyence, Cognex, Dymax) are from vendor pages and trade publications as of October 2026 and are vendor claims, not our measurements; confirm current models and specifications with each vendor. Worked numbers are examples. 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.
