Part Identification
The line decides what the part is, before the part gets there
We have built a computer vision part identification system that modulated process recipes for a client. This page describes how that class of system works: classifying each incoming part variant and handing the matching recipe to the controlling PLC, so a high-mix line runs without an operator selecting anything.
A note on what follows. The engagement itself is confidential, so the walkthrough below is written generally rather than as an account of one deployment. Treat the specifics as an illustration of the approach, not as reported results.
The problem: the line knows less than the operator
On a shared line running mixed part variants, the process has to change per part: different fixture positions, different torque or pressure profiles, different parameters entirely. On most lines, the thing that knows which part is coming is a person.
That works until it doesn't. The failure modes are familiar to anyone running high mix:
- Mis-selection, where the wrong recipe runs on a good part and scraps it
- Changeover time spent stopping, confirming, and re-selecting between variants
- Batching pressure, where the line runs long runs of one variant purely to avoid changeovers, inflating WIP
- Shift variance, where the process quietly depends on who is running it
- No traceability, because the record of which recipe ran on which part is a person's memory
What the system does
A vision-based control system, running on Keyence/Cognex-class camera hardware, captures each part in-line, classifies its variant against a trained model, and pushes the corresponding recipe, process parameters, fixture settings, torque and pressure profiles, to the controlling PLC before the part reaches the next station.
The important design decision was where the decision lives. The camera does not just report a classification to a screen for someone to act on. It writes the selection into the control system, which means the process cannot proceed on the wrong recipe because nobody is being asked to intervene.
Capture
The part is imaged in-line at its natural position on the conveyor, under lighting chosen so the distinguishing features between variants are the most reliable thing in the frame.
Classify
The variant is identified against a trained model, with a confidence value attached to every decision rather than a bare yes/no.
Resolve to a recipe
The variant maps to a named recipe holding the actual process settings. The mapping is data your engineers can read and change, not logic buried inside a vision program.
Push to the PLC
Settings are written to the controller before the part arrives at the station, with the handshake and timing designed around the line's actual cycle rather than a best case.
Log the decision
Part, variant, confidence, recipe, and timestamp are recorded, so the question "what ran on this part" has an answer that does not depend on anyone's recollection.
The part that gets underestimated: low-confidence handling
Any classifier will eventually see something it is not sure about: a part it was not trained on, an unusual presentation, a lighting change from a burned-out fixture. The interesting engineering is not the confident case, it is what happens when confidence is low.
The system needs a defined behavior for that, agreed with the people who run the line: hold the part, route it out, alarm for a human decision, or fall back to a safe default recipe. What it must never do is guess and proceed silently. A vision system that fails loudly is worth more than one that is right slightly more often but fails quietly.
Where this generalizes
The underlying pattern, know what is in front of the process before deciding what the process should do, shows up well beyond recipe selection:
- Sorting and routing, diverting variants to the correct downstream lane
- Fixture and tooling selection on shared cells
- Kitting verification, confirming the right components are present before assembly
- Assembly-sequence gating, preventing the next step until the current state is confirmed
- Traceability capture, tying serials to the process settings actually used
What makes a line a good fit
- Multiple part variants sharing a line or cell
- A process that must change per variant in a way the controller can accept
- Variants that are visually distinguishable, even subtly, under controlled lighting
- A real cost attached to getting it wrong: scrap, rework, downtime, or a customer escape
- A PLC or controller that can take settings from an outside system
If mis-selection currently costs you scrap or stoppages, that cost is measurable, which usually makes this a candidate for our profit-sharing model rather than a fixed fee. If the variants are not reliably distinguishable, we will say so before anyone buys a camera. It is also common to pair this with defect inspection on the same hardware, since the part is already in front of a camera.
Running mixed variants on one line?
Tell us how many variants share the line, how the recipe gets selected today, and what a mis-selection costs. That is enough for us to tell you whether this is worth doing.