Machine Vision Defect Detection on Roll Forming Lines

1. Definition

Machine vision defect detection on roll forming lines uses industrial cameras, controlled lighting, and image processing to find surface flaws, missing or mislocated features, and selected geometry deviations while product moves at production speed. It is a sibling discipline to in-line laser profile measurement: vision excels at 2D appearance and feature presence; lasers excel at full cross-section dimensional maps.

On continuous roll forming lines, vision stations are typically placed where the product shape is stable, lighting can be controlled, and—when reject is required—defect location can be mapped to a downstream cutoff or diverter.

2. Scope: Surface vs Geometry vs Features

Define inspection scope in the quality plan: which defects stop the line, which log for SPC, and which are advisory only.

3. Typical Defect Types on Roll-Formed Products

• Handling and tooling marks — roll pickup, guide scratches, pinch roll impressions
• Coating damage — galv white rust, paint scuff on PPGI/PPGL, film wedge defects
• Pierce and punch errors — slug retention, burr protrusion, missing hole, wrong pattern
• Edge condition — wave, trim burr, fish-tail zone anomalies near coil ends
• Repeat profile defects — inconsistent rib depth on roofing/siding panels
• Contamination — oil drip, scale flake, foreign object on visible face

Roll forming creates defects both in the forming stations and in upstream/downstream operations (leveler, prepunch, shear). Vision placement should cover the customer-visible state after all critical operations.

4. System Architecture on a Forming Line

A typical architecture includes:

1. Camera(s) — area scan for feature/surface; line scan for wide uniform surfaces
2. Lighting — LED bar, dome, dark-field, or structured light depending on defect class
3. Encoder / line trigger — synchronize frames to strip motion for repeatable field of view
4. Vision PC or smart camera — runs inspection tools or ML inference
5. PLC interface — pass/fail, defect class, longitudinal position tag
6. Reject actuator — flying shear mark, diverter gate, or stop for cut-length parts
7. HMI / MES logging — images, trends, coil ID association

Multi-camera rigs inspect different faces (top cosmetic, bottom structural, edge view). Cable routing and vibration isolation matter on long mills.

5. Lighting and Optics

Lighting dominates success on reflective metals. Strategies:

5.1 Diffuse dome lighting

Reduces specular hotspots on galvanized and polished stainless; common for general surface defect search on visible faces.

5.2 Dark-field / low-angle grazing

Enhances scratches, dents, and emboss height contrast; sensitive to strip flutter—stabilize guides first.

5.3 Structured light / photometric stereo

Reconstructs small 3D surface variations; useful for dent detection on flat panel zones between ribs.

5.4 Coaxial / telecentric

Supports accurate hole metrology when camera axis is aligned and depth of field is sufficient.

6. Panel Products: Pitch, Depth, and Profile Features

Roofing, siding, and deck panels produced on roll formers repeat a corrugated or trapezoidal profile along length. Vision checks often include:

• Rib pitch — center-to-center repeat versus drawing; detects slip in drive or tooling wear
• Rib depth / height — proxy for forming depth when laser profile is not installed on every line
• Flat pan width — between ribs; sensitive to over-forming or material thickness change
• End pattern — after cut, verify rib phase at panel ends for lap-joint fit

Pitch measurement uses encoder-triggered frames and known camera calibration (pixels per mm). Depth from 2D shadow geometry is approximate compared with laser cross-section; acceptable for many building-product QC tiers when tolerances are coarse and lighting is stable.

For high-stakes architectural panels, pair vision with spot laser height or full profile gauge on a subset line or audit station.

7. Inspection Algorithms and Training

Industrial vision toolchains mix classical and ML methods:

• Golden template comparison — subtract aligned reference image; threshold difference map
• Blob analysis / morphology — scratch length, pit area limits
• Pattern matching — locate holes and measure distances
• Deep learning classifiers — defect type on varied backgrounds; needs labeled image sets
• Anomaly detection — train on “good only” when defect samples are scarce

Roll forming adds non-stationary background (rib moves through FOV). Algorithms must track profile phase or restrict ROIs to stable pan regions.

7.1 Validation discipline

Collect defect libraries from production: confirm vision catches real defects at line speed without destroying throughput. Report detection rate and false reject rate separately—marketing “99%” numbers without denominator are meaningless.

8. False Rejects and False Accepts

False reject (good product flagged bad) wastes material and erodes operator trust. Common causes on roll form lines:

• Oil film or water streak mimicking scratch
• Coil join or weld bump not masked in software
• Lighting drift as LEDs age or ambient sun hits open bays
• Strip flutter changing rib shadow geometry
• Over-tight thresholds copied from lab golden part only
• Color change at coil splice without recipe update

False accept (bad product passes) often traces to insufficient lighting contrast, wrong ROI after tooling change, or ML model trained on obsolete surface finish.

Implement graded responses: caution alarm for borderline defects, hard reject only for confirmed classes. Operators override with logged reason codes to feed algorithm tuning.

9. Integration with Cutoff and Reject Handling

Continuous lines map defect longitudinal position from encoder count to a downstream actuator:

1. Vision detects defect at position s along strip
2. PLC computes delay to cutoff or diverter based on line speed and distance
3. Flying shear cuts out defect zone or marks length for manual sort
4. Alternatively, stacker diverts entire cut-length pack if end-of-panel check fails

Integration requirements:

• Common encoder reference or synchronized time base
• Compensation for acceleration/deceleration during speed changes
• Minimum defect length vs shear capability
• Safe state when vision PC offline (fail-safe stop vs run-with-warning policy)

For coil-to-coil production without inline reject, systems log defect maps for later trimming at slitter or customer—less ideal but common on lower-value commodity profiles.

10. Vision vs Laser Profilometry

Many modern lines deploy both: laser for dimensional CTQs, vision for surface and feature integrity. IIoT historians can correlate a dimensional drift event with sudden surface defect rate (e.g., roll pickup after gap crash).

11. Environment and Maintenance

• Enclosures with air purge reduce oil mist on lenses
• Vibration isolation pads for camera mounts near flying shear
• Scheduled lens cleaning aligned with shift handover
• Temperature management for vision PC cabinets in hot bays
• Version control on inspection recipes when tooling revisions ship

Vision is not install-and-forget. False reject rate creeping up over weeks often means optics contamination, not mysterious process magic.

12. Implementation Checklist

1. Rank defect classes by customer cost and frequency
2. Pick inspection station after shape and punch operations stabilize
3. Run lighting trials on worst-case coils (bright galv, dark paint, oily)
4. Define pass/fail thresholds vs advisory bands
5. Integrate encoder and validate position error budget to cutoff
6. Train operators on override logging and image review
7. Link defect logs to coil ID and shift in MES
8. Plan recipe management when product mix is high
9. Measure false reject rate weekly during ramp-up

13. Boundaries

This page covers machine vision for defect detection on roll forming lines. It does not specify camera brands, guarantee detection percentages, or quote system prices. It does not replace laser-based dimensional inspection where drawing tolerances demand full section measurement.

14. Buyer / Engineer FAQ

Can vision replace laser profile gauges?

Usually no for full section CTQs. Vision complements lasers for surface and 2D features.

How do we reduce false rejects on galvanized?

Diffuse or polarized lighting, stable guides, recipe-specific exposure, and mask non-inspection zones.

Do we need deep learning?

Not always. Classical tools suffice for many hole and print checks. ML helps varied cosmetic defects with enough labeled images.

Where to mount cameras on a long line?

After final forming and critical secondary ops; multiple stations if both top and bottom faces matter.

What happens when vision fails?

Define fail-safe policy with safety and quality stakeholders—stop line vs run with warning.

Can vision measure rib depth on panels?

Approximate from geometry and lighting; for tight depth CTQs add laser or contact audit.

How does vision tie to IIoT?

Export defect counts, images, and positions to historian for coil traceability and Pareto analysis.

15. Related Encyclopedia Topics

• In-line Inspection; IIoT in Roll Forming
• Edge Waviness; Cracking in Roll Forming; Pre-painted Steel (PPGI)
• Closed-loop Control; Digital Twin

16. Summary for Specifiers

Machine vision on roll form lines detects surface defects, feature errors, and selected panel geometry when lighting, motion sync, and algorithms match the product. Manage false rejects through on-line tuning and graded alarms. Integrate with cutoff or diverters using encoder-based defect mapping. Pair with laser profilometry when dimensional CTQs dominate; use vision where appearance and feature integrity drive customer acceptance.