HyperQ AI Vision reaches 99% detection accuracy at micrometer-level precision — including on the reflective, curved metal surfaces that make beverage-end inspection technically demanding. At 2,000 to 2,400 cans per minute on a modern can-filling line, that detection accuracy matters only if the system maintains pace without becoming a throughput bottleneck. Whether it does depends on cycle-time design, not lab demonstrations.
The post covers three things: the defect modes that produce the most customer complaints and field rejections on can lines; why statistical sampling is not a viable quality strategy at this speed; and the disqualifiers for 100% vision inspection — the line configurations where sampling is still the right call.
There is no direct customer deployment in the SG/MY beverage or contract-packing sector to cite here. The architectural pattern described applies across precision manufacturing contexts: the same detection and integration logic operates in the Tier-1 automotive parts deployment, where HyperQ runs 6 production lines against 8,000+ product variants at 11,520 units per day per line. Substrate and defect classes differ; the underlying model architecture and integration requirements translate.
The defect modes that matter on a can line
Can and beverage-end defects sort into three groups by consequence: pressure-integrity defects, seal-integrity defects, and regulatory/traceability defects.
Dome and panel geometry — the can end is a pressurised vessel. An end with insufficient dome height or a pre-stressed panel can reverse under carbonation pressure after sealing, compromising seal integrity and producing visible case bulge. Dome height deviation at the failure threshold is typically 0.3 to 0.5 mm below nominal. Ambient lighting does not generate sufficient contrast on a specular aluminium surface to detect this deviation. Coaxial or structured-light illumination is required to produce the brightness gradient across the dome face that makes sub-millimetre height variation visible to a camera system. Lines that specify the model and camera correctly but skip the illumination engineering will fail dome inspection regardless of inference quality.
Double-seam defects — the mechanical interlock between can body and end is where fill integrity lives. Seam anomalies — insufficient seam height, knocked-down flanges, countersink droop — are among the most serious defects in can manufacturing because they relate directly to fill integrity in transit and on the shelf. The traditional detection method is destructive teardown: three to five cans per shift per seamer, inspected against BCME or Canmakers' Code tear-down criteria. On a four-seamer line running 2,000 cans per minute through a 10-hour shift, this generates 12 to 20 destructive samples from 1.2 million units produced — a coverage rate below 0.002%. A systematic seamer fault that develops between teardown intervals produces defective product across the entire unsampled window.
Dents, score depth, and code legibility — panel dents above approximately 1 mm depth at the chime or body panel create fatigue concentration points that can fail under shelf stacking or shipping compression loads. Pull-tab score depth on easy-open ends carries tolerances as tight as ±0.05 mm; underscored ends require excessive opening force that leads to consumer complaints and potential injury; overscored ends fracture prematurely under normal use. Lot, date, and shift codes — laser-etched or inkjet-printed — are regulatory requirements under both Singapore Food Agency frameworks and the Malaysian Food Act. Code legibility failures are a recall trigger under both regimes, and lot traceability without confirmed code legibility at production time is an audit gap.
Why sampling is not a quality strategy at high speed
The case for statistical sampling rests on three assumptions: defect rates are stable and predictable, defect modes are random across the production run, and field-failure consequences are low enough to absorb the statistical escape rate. None holds reliably on a high-speed can line.
Defect modes on these lines are typically systematic rather than random. A worn seamer produces low seam height on every unit it processes after a certain wear threshold, not randomly distributed across shifts. A misaligned scoring tool produces score-depth failures in runs that track the tooling maintenance cycle. An operator who adjusted a seaming chuck at the start of shift three without sign-off creates a fault that runs from adjustment until the next teardown sample — which may be four hours later.
Sampling across systematic failure modes misses entire production windows between samples. The calculation is direct:
A line at 2,000 cpm with one-per-five-minute sampling captures approximately 120 samples in a 10-hour shift from 1.2 million units. If a seamer fault develops at minute 47 and the teardown at minute 60 catches it, approximately 26,000 units produced in that 13-minute window are unsampled and at risk. Under 100% inspection, every unit in that window has an image record and the fault generates an alert at the moment the first defective unit passes the camera.
| Inspection method | Samples in 10-hour shift at 2,000 cpm | Systematic fault coverage | Dome defect detection | Per-unit audit trail |
|---|---|---|---|---|
| 1-per-5-min operator sampling | ~120 | 0% if fault runs between intervals | Not possible | None |
| Destructive seam teardown | 12–20 per 4-seamer line | 0% between teardown intervals | Not applicable | None |
| 100% vision at line speed | ~1,200,000 | 100% while fault runs | Yes, with correct illumination | Full image record per unit |
The coverage difference is not incremental. It is structural.
What 100% inspection requires on a can line
Detection accuracy is the tractable problem. Integration at line speed is harder and more commonly where projects fail.
Cycle time at 2,000 cpm. At this throughput, the inspection window per unit is 30 milliseconds from the leading edge of one can to the leading edge of the next. A single-station system must complete frame capture, inference, and reject-trigger output within that window. Multi-camera configurations split the inspection task across stations — seam geometry at station 1, dome at station 2, code verification at station 3 — giving each station its own 30-millisecond window on a different defect class. Single-station 100% inspection at 2,000 cpm requires GPU-accelerated inference with sub-20-millisecond model output. That ceiling is reachable, but it must be designed in at the start — not discovered after the hardware is bolted to the line.
Reject handling calibration. A vision system that detects defects but cannot trigger a confirmed mechanical reject is a monitoring system, not a quality system. High-speed can lines use pneumatic pushers or air-blast reject stations downstream of the inspection point. The vision system generates a reject trigger with an output delay calibrated to the physical distance between inspection point and reject station. At 5 metres and 2,000 cpm, the delay is approximately 150 milliseconds. Encoder drift or trigger-timing errors at this speed produce incorrect good-product rejections or, more seriously, defective-product escapes through the reject station. Calibration is not one-time at commissioning; it requires periodic verification as encoder performance changes.
Illumination geometry as the prerequisite, not the afterthought. The most common failure mode in poorly integrated can line deployments is not an underpowered model — it is a camera and model correctly specified for a defect class that the illumination geometry renders invisible. A practitioner familiar with multiple can-line integrations described this pattern directly: the project team specifies the camera resolution and the AI model, procures both, installs them at the line, and then discovers that the dome defects the customer was paying to catch do not produce any detectable signal under the ring light that was already installed for barcode reading. The rework at that point is not a model retrain — it is a complete illumination re-engineering.
Dome height variation on a specular aluminium surface requires coaxial illumination or structured light to produce the contrast gradient needed for sub-millimetre height measurement. A model trained on correctly illuminated reference images will fail to detect dome defects when deployed under diffuse ambient lighting, because the physical signal that the defect produces under structured light does not exist under ambient conditions. Illumination engineering comes before camera selection, not after.
When 100% vision inspection is the wrong answer
Low-speed or short-run lines. Below approximately 400 to 600 cpm, the coverage advantage of 100% vision over a trained inline inspector narrows. On a 200-cpm specialty beverage line running 8-hour shifts, the number of units between sample intervals is low enough that systematic faults produce smaller uninspected windows. Capital investment in a full 100% vision integration at this speed competes directly with staffing and process-control investments, and the business case on throughput-economics alone is harder to close.
Lines with stable, documented low-defect profiles. If a line has run the same can format for several years with documented defect frequency consistently below 50 ppm, no systematic seamer events in recent FMEA history, and stable teardown results across shifts, sampling is a defensible quality strategy. 100% vision adds assurance but does not prevent escape events that the historical record indicates do not occur. The investment case then rests on regulatory traceability and audit requirements, not defect prevention — which changes the evaluation entirely.
Lines where reject handling cannot be integrated at the inspection point. A vision system that cannot mechanically reject detected defectives at the point of detection is a monitoring tool. If the line's mechanical layout does not permit a reject station within the distance that allows correct trigger-delay calibration, 100% detection does not translate to 100% quality control. This is a feasibility check that must happen before capital commitment, not after.
Lines running multiple formats below 30-minute changeover intervals. Model changeover at format change adds time at the line. Where format changes are frequent and changeover time drives OEE directly, the operational overhead of model selection per format needs to be accounted for in the cost model.
Integration with existing fill-check stations and MES
Most high-speed beverage lines already have fill-level inspection — typically X-ray or capacitance-based — installed downstream of the filler. The optical vision station for surface and end inspection integrates to the same line PLC and shares the rejection control loop where the mechanical layout permits. This avoids installing a second independent reject mechanism.
X-ray fill cameras are a different imaging modality and do not serve optical surface inspection. Where existing optical cameras are installed for label or date-code reading, reuse for defect inspection depends on the camera's frame rate and sensor resolution relative to the line speed and target defect size. A site survey is required to confirm reusability — not as a formality, but because a camera installed for label reading at 400 cpm running on a line now operating at 1,800 cpm will not produce usable images for defect detection.
Code verification outputs — lot number, date code, shift code, legibility pass/fail — route via OPC-UA or MQTT to plant MES in real time, creating per-unit traceability records without manual entry. For SFA and Malaysian Food Act compliance, this record chain from production to distribution is the audit trail that supports recall response. When it exists at the unit level, a recall scope can be bounded to a specific lot, a specific shift's production window, and specific line positions. When it does not exist, the recall scope defaults to the full batch period.
For SG/MY contract-packing lines that produce for multiple brand owners, per-unit image records also resolve brand-owner disputes about whether a defect was present at production or caused in downstream handling. This is a contractual risk-management argument that belongs in the capital justification alongside the quality-defect-prevention argument.
HyperQ AI Vision on beverage lines: training and deployment
HyperQ AI Vision uses a patented low-data training architecture that reaches 99% detection rate at micrometer-level precision with as few as 1,000 labeled images per defect class. For a can line with three primary defect-class families — seam geometry, dome geometry, and code legibility — a full model library typically requires 3,000 to 4,000 labeled images, not the 10,000-per-class requirement that makes some deployments into multi-year data-collection projects.
The 8,000+ pre-trained model library covers common defect classes in metal packaging, including dome geometry, countersink irregularities, and seam anomalies. These library entries serve as starting points; client-specific training on the actual production line's defect sample refines performance against the specific surface finishes, alloys, and defect morphologies on each line.
Deployment on existing camera infrastructure is supported where the existing cameras meet frame-rate and resolution specifications. The site survey determines what is reusable and what needs replacement before any capital commitment.
Frequently asked questions
What illumination setup is needed for dome height inspection on aluminium ends? Sub-millimetre dome height detection requires coaxial illumination or structured-light (fringe projection or laser line) to generate a contrast gradient across the dome surface. Diffuse backlighting and ring lights do not produce this gradient on specular aluminium. The illumination choice drives the camera placement geometry: coaxial setups place the light source along the optical axis, requiring the camera to view the end face square-on rather than at an angle.
Can a vision system detect seam defects without destructive teardown? Non-contact vision detects seam geometry anomalies — seam height, width, and body-hook length within certain bounds — from the seam's external profile under directional side-lighting. It does not replace destructive teardown for internal hook engagement, which requires physical section analysis. The correct framing is that vision inspection adds coverage for seam anomalies between teardown intervals; it does not eliminate the teardown requirement.
How does frame rate scale with line speed? A camera requires a frame rate that reliably places each can under inspection regardless of minor line speed variation. At 2,000 cpm on a single-track line, the inter-can pitch at 300 mm means 100 mm travel per 30-millisecond window. A camera with a 5-microsecond exposure time at 60 fps produces acceptable motion blur for most defect classes at this speed. Dome height defects requiring sub-0.5-mm measurement may require higher frame rates or strobe illumination to freeze motion adequately.
What happens when the line changes from 330 ml to 500 ml cans? Model changeover at format change is a configuration step: the operator selects the format-specific model and the inspection parameters update. Training a model for a new format requires a labeled image set from the new format's production; with 1,000 images per defect class and the pre-trained library as a starting point, the training cycle is typically shorter than for an initial deployment. The format database accumulates as lines are commissioned for additional formats.
Does vision inspection data integrate with HACCP and food-safety documentation? Per-unit inspection records — image, timestamp, outcome, operator shift — export to MES and eQMS systems via OPC-UA or MQTT. For HACCP documentation, the inspection station functions as a monitored Critical Control Point, with the inspection record serving as the CCP monitoring log. This is the same data structure required for SFA licensing and Malaysian Food Hygiene Regulations audit submissions, and it is more complete than paper-based or end-of-shift summary logs.
At 2,000 cans per minute, a seamer fault running for 15 minutes produces approximately 30,000 units in an unsampled window under conventional teardown-and-sampling QC. Send us your line speed, seamer count, primary defect modes from your last quality event or customer complaint, and whether your line has an existing reject mechanism. We will specify a 100% inspection configuration and model library against those parameters within 2 weeks — no contract until the configuration runs confirmed on your actual line.
Send your line spec and defect history for a 100% inspection configuration