At a typical aluminum extrusion line running 8-12 meters per minute, a die defect discovered at the end-of-line CMM check means roughly 6 meters of scrap -- the full stick length. The same defect caught at the die exit, 30 centimeters of material in, means 30 centimeters of trim. Same defect class, same die condition. The only variable is where the camera was.
That geometry -- scrap volume is linear in distance from the catch point to the camera -- is the entire business case for inline inspection in extrusion. No statistical model, no ROI calculator, no vendor comparison is needed. The math is spatial.
Why extrusion defects are different from discrete-part defects
Most machine-vision applications inspect discrete units: a machined part moves into the field of view, stops, gets inspected, moves on. The defect is either present or absent on that unit. The inspection task is classification.
Extrusion produces a continuous part. The same piece of material runs through the die for seconds, minutes, or the full length of the billet. Defects do not appear on units -- they appear at points along a continuous surface that is still in production when the defect originates.
The sampling-plan logic that works for discrete parts fails for continuous products. Inspecting every 50th unit works if each unit is independent. But an extrusion die that begins wearing at minute 3 of the run produces a defect signature that runs from meter 24 through meter 60 -- and a sampling plan that checks at meters 10, 30, and 50 finds the defect at one check, misses it at the next, and finds it again at the third. The inspection data looks inconsistent. The process is not inconsistent; the inspection coverage is.
Extrusion inspection requires temporal coverage, not sampling. The camera needs to see the profile at the same position in the production line continuously -- not periodically.
Common extrusion defect classes and where they originate
Extrusion defects originate at specific points in the production sequence. Matching the defect class to its origination point determines where the inspection camera should be located.
| Defect class | Origination point | Growth pattern | Catch-point implication |
|---|---|---|---|
| Die line / score mark | Die land condition | Continuous along full length once started | Inspect at die exit; die lines do not stop once begun |
| Surface score | Die land, bearing wear | Starts shallow, deepens over billet run | Inspect at die exit; depth increases over time |
| Orange peel / rough texture | Low die temperature, billet temperature variation | Cyclic with billet temperature; batch-pattern variation | Inspect inline; end-of-line may miss interspersed good sections |
| Twist | Bearing balance failure | Progressive along length | Inspect at die exit; twist deepens with bearing degradation |
| Surface pit / porosity | Billet quality, melt pool defects | Random distribution; not progressive | End-of-line sampling may suffice if billet quality is controlled |
| Dimensional deviation (wall thickness, profile geometry) | Bearing wear, die deflection | Progressive with die fatigue | Inline at die exit for early wear signal |
| Extrusion seam (hollow profiles) | Weld chamber condition | Full-length defect once seam begins | Die exit is the only viable catch point |
| Scratch (post-die) | Handling, runout table contact | Point occurrence, random | Post-runout table inspection |
| Contamination / inclusion | Billet or die contamination | Random occurrence | End-of-line; contamination is not progressive |
The split between "inspect at die exit" and "end-of-line may suffice" is not arbitrary. It follows the defect's growth pattern. A defect that starts and deepens over time -- die wear, bearing degradation, scoring -- produces the maximum scrap if caught late. A defect that occurs randomly and does not propagate -- surface contamination, scratch from handling -- does not change in severity with inspection position.
For most structural aluminum and PVC profile production, the high-cost defect classes are the progressive ones: die line, twist, dimensional deviation. Those are the cases where the 6-meter vs. 30-centimeter catch-point geometry applies.
The catch-point distance table
The business case for inline inspection is geometric. Scrap volume is a function of line speed multiplied by detection delay.
| Inspection position | Detection delay (at 10 m/min) | Scrap per event | Defect classes appropriate |
|---|---|---|---|
| Die exit (inline) | ~2-5 seconds | 30-80 cm | Die line, score, twist, dimensional |
| Mid-runout table | ~30-60 seconds | 5-10 m | Progressive surface defects |
| End-of-line (before cut) | ~45-90 seconds | Full stick (6-12 m) | All classes, but maximum scrap for progressive |
| Post-cut CMM or manual | Minutes to hours | Full batch affected | Dimensional; triggers full-batch rework or scrap |
At 10 meters per minute with a 6-meter stick length, a defect that starts at the die and is not detected until end-of-line means 100% of the stick is affected. The same defect detected at the die exit means less than 10% is affected -- 60 centimeters of trim, not a full stick of scrap.
At typical aluminum extrusion costs, the difference between 30 centimeters of trim and 6 meters of scrap on 1,000 sticks per day is not a small number. The catch-point geometry is the business case. When Hypernology works with extrusion customers to model this geometry against their specific line speed and scrap rate, the calculation usually determines the inspection position before any capability discussion begins -- because the ROI depends entirely on where the camera goes, not on which AI model runs it.
How AI vision handles continuous profiles
The inspection challenge for inline extrusion is not detection sensitivity -- standard industrial cameras at 10-micrometer precision are sufficient for the defect classes that matter. The challenge is inspection logic that handles a continuous, moving, variable-width field of view without the discrete-part reset that simplifies most vision applications.
HyperQ AI Vision handles continuous profiles by treating the inspection stream as a temporal sequence rather than a set of independent frames. The model builds a surface-state representation that tracks anomaly patterns as they develop along the profile -- not as they appear in an isolated frame. A die score that is 0.1 mm deep at meter 2 and 0.3 mm deep at meter 5 is detected as a progressive anomaly, not as two separate isolated events. That temporal tracking is what allows early detection and early alert before the defect reaches reject severity.
At line speeds up to 12 meters per minute, sub-1-second processing per inspection zone keeps up with the profile. The detection latency is not the bottleneck; the alert routing to the line operator is. An alert that takes 30 seconds to reach the operator at the pull end is a 5-meter delay, not a 1-second delay. Alert routing design is part of the deployment architecture.
For hollow profiles with extrusion seams, HyperQ AI Vision inspects both the external profile surface and, where camera geometry allows, the seam region. External profile defects are the primary class, but seam integrity matters for pressure-rated applications in SG/MY building and industrial markets.
Die wear tracking and the continuous learning loop
A die that produces acceptable profiles on Monday may produce borderline profiles by Friday -- not because anything has visibly failed, but because the die land has worn incrementally. The wear is below the reject threshold. It is heading toward it.
Inline AI vision with die-wear trend tracking shows the direction of travel. The model logs the defect-severity trend across successive billet runs. When the trend line crosses a user-defined threshold -- for example, surface score depth increasing by more than 20% per 1,000 meters of production -- the system alerts the tooling team before the defect reaches the reject boundary.
This is the same continuous-learning retraining loop that applies across HyperQ AI Vision deployments: as the production process changes, the model updates to track the new normal and flag deviation from it. For extrusion, the "new normal" is the die's current production state. The model adapts to incremental die wear without requiring manual recalibration, and the trend data provides an objective die-maintenance signal.
The practical outcome is a shift from failure-triggered die changes -- "the die produced rejects, change it" -- to trend-triggered die changes -- "the die is 12% toward the reject threshold at the current wear rate, change it at the next scheduled shutdown." The die change happens at a planned time, not an emergency stop. Line utilization improves.
Profile family switching and recipe management
Production lines running multiple profile families -- a common configuration in contract extrusion for construction and automotive components -- require the inspection system to hold reference models for each active profile. A system that requires manual reconfiguration between profiles stops the line or gets skipped.
The same capability that allows a Tier-1 automotive parts supplier to switch between 8,000+ product models in under 2 seconds applies to extrusion profile families: each profile family has its own inspection recipe, and the switch between recipes is automatic. A barcode scan or PLC trigger at the die change initiates the model switch.
For contract extruders running 30-50 profile families across multiple customers, that automatic switching eliminates the manual-reconfiguration bottleneck that causes inspection programs to be "off" during changeovers. The high-risk window -- the first 60 seconds of a new profile run when the die is settling -- is covered, not bypassed.
False-reject discipline at line speed
High false-reject rates at line speed cause a different problem than they do in discrete-part inspection. With discrete parts, a false reject is a rework queue entry. In continuous extrusion, a false reject triggers an operator intervention on a moving line: the alert sounds, the operator checks the surface, finds nothing, dismisses the alert, and the cognitive load goes up. After three false alerts in 20 minutes, the fourth alert -- the real one -- gets the same dismissive response.
HyperQ AI Vision delivers 60-80% false-positive reduction relative to rule-based AOI in surface-inspection applications. In extrusion, achieving that false-positive discipline requires calibrating the model on the expected surface variation for each profile family -- the micro-texture variation inherent to the die geometry, the expected brightness gradient across a hollow profile, the normal reflection pattern of brushed or anodized surface finishes. A model that is not calibrated to the specific profile's normal variation range generates the alert-dismissal pattern.
The deployment validation sequence includes a false-positive rate verification: running 30 minutes of known-good production through the system at line speed and verifying the alert rate is within the defined threshold before go-live. This step is not optional. An inspection system with a high false-positive rate at line speed is operationally worse than no inspection system.
Hardware-agnostic inline deployment
The camera requirements for inline extrusion inspection at standard production speeds are within the range of standard industrial cameras: 5-10 megapixel line-scan or area-scan cameras, appropriate field-of-view optics for the profile width, and controlled LED illumination. None of these require a hardware-locked vendor ecosystem. HyperQ AI Vision runs on cameras from any manufacturer that meet the resolution and frame-rate specifications for the line speed.
For lines that already have cameras installed -- common in operations that have tried rule-based AOI before -- the existing cameras are used where their specifications are adequate. The 30-50% hardware cost saving versus hardware-locked systems applies here: the cost is edge compute and software, not a new camera ecosystem.
The illumination design matters more in extrusion than in many other vision applications because the primary defect classes -- die line, score, twist -- show best under specific illumination angles. A raking light at a low angle to the surface shows die lines that a diffuse overhead light obscures. A cross-polarized configuration shows surface texture variation that specular illumination washes out. The inline deployment includes illumination geometry optimization for the specific defect classes the customer prioritizes.
Decision guide: inline vs. end-of-line inspection
| Production condition | Recommended inspection position | Rationale |
|---|---|---|
| Progressive defects (die line, score, twist, dimensional drift) | Inline at die exit | Catch-point geometry: minimum scrap, early die-wear signal |
| Random defects only (contamination, handling scratch) | End-of-line | No geometry advantage to moving camera upstream |
| Multiple profile families, automatic switching | Inline with recipe management | Changeover windows are highest-risk periods |
| High-value alloy or specialty profiles | Inline | Cost-per-meter is too high for end-of-line catch |
| Commodity volume with tight scrap budget | Inline at die exit | Scrap economics dominate; geometry is the business case |
| Certification-required audit trail (automotive, aerospace) | Both inline and end-of-line | Dual-verification for full-length documentation |
For more on how AI vision handles complex and irregular defect geometries on non-extruded parts, see Complex and irregular defects in 2D vision inspection. For the plastics side of continuous-profile inspection, see AI vision for plastics and injection-molded parts. HyperQ AI Vision capabilities are at /solutions/hyperq-ai-vision.
Send a 30-centimeter sample of a profile family that has caused escapes or high scrap rates on your line. Within 2 weeks, we run it through an inline detection model, show detection performance against the defect classes present, and return the sample with a catch-point analysis for your line geometry. No contract until the detection spec is met against your profile data.
Send your extrusion sample and book the catch-point analysis