Computer Vision
How a PVC profile manufacturer started detecting defects in 30 seconds instead of 45 minutes
PVC profile and pipe manufacturing
TLDR
The extrusion line runs continuously. The operator measured profile geometry with a caliper once every 45 minutes. At 8 m/min, 360 meters of product are produced in 45 minutes. If a parameter started drifting right after the measurement, that entire volume was in question. BackendFlow installed a continuous automatic geometry control system: when deviation appears, the operator receives a signal in 30 seconds.
Result: non-standard output per incident dropped from 360 to 25 meters - a 93% reduction.
Context
Profile parameters - diameter, wall thickness, cross-section shape - drift out of tolerance gradually. It is not an accident or an obvious breakdown. Melt temperature changes slightly, the die wears slightly, traction becomes slightly unstable. The operator does not see the difference; the profile looks normal.
The operator measured geometry with a caliper every 45 minutes. At 8 m/min, 360 meters were produced during that time. With 2-3 incidents per month, direct raw material losses alone reached up to about $1.4K per month, plus line readjustment time and labor for reprocessing non-standard product.
For production: each incident meant 360 meters of product in question. Part of it went to reprocessing, part reached customers with tolerance violations. Geometry claims are difficult to resolve: the defect is found during installation, after the profile has already been cut.
For the operator: manual measurement is stress-control: 45 minutes of complete uncertainty, then a few seconds of measurement. A human physically cannot monitor the parameter continuously while also running the line.
Task
Detect profile geometry deviation in real time and immediately signal the operator - not after 45 minutes, but within 30-60 seconds after the problem appears.
Solution
The key point: accurate non-contact measurement of moving profile geometry requires not industrial video monitoring, but specialized metrology equipment.
Standard cameras do not provide the required accuracy here. Laser profilometers and calibrated optical triangulation systems measure geometry with accuracy down to tenths of a millimeter in real time. The cost of a profilometer starts at about $3.9K. This needs to be understood before the project so the budget does not diverge from expectations.
Workflow:
- The profilometer continuously measures profile geometry at the line output
- The system compares current parameters with specified tolerances
- If a parameter goes out of tolerance, the operator receives an immediate signal
- The operator reacts in 2-4 minutes: during that time 16-32 meters of non-standard product accumulate instead of 360
- All measurements are logged for line stability analysis
Current status: pilot operation in parallel with manual control. The system is accumulating data on real incidents to validate the financial model.
Insights
A camera does not work here - a profilometer is needed.
The customer's first intuition was to install a camera. But an industrial camera provides visual control, not metrology. Geometry deviations of 0.1-0.3 mm at flow speed cannot be reliably measured with standard optics. The right tool is a laser profilometer. Its cost ($3.9K+) is higher than a camera, so TCO must be calculated honestly from the start.
System data makes die maintenance plannable.
Measurement logging revealed a non-obvious pattern: parameter drift increased predictably as the die wore down. This allowed the plant to move from reactive maintenance, when defects had already appeared, to planned maintenance before defects occurred. This side effect is worth more than defect control itself.
Pilot operation is not a pause; it is data collection.
A financial model with a wide payback range (5-14 months) reflects uncertainty in incident frequency. Pilot operation closes that uncertainty: real quarterly data will narrow the model. This is standard practice for manufacturing CV projects.
Results
| Metric | Before | After |
|---|---|---|
| Non-standard output per incident | ~360 m | ~25 m |
| Time to detect deviation | up to 45 min | 30-60 sec |
| Direct raw material losses | ~$1.4K/month | ~$100/month |
Assumptions: line speed 8 m/min, operator reaction 3-5 min, PVC about $1.30/kg, consumption 1.8 kg/m, 2-3 incidents/month. Payback range is refined during pilot operation based on real incident frequency.
Want the same kind of calculation for your production?
Book an audit ->