Computer vision for more consistent dimensional inspection
Your inspectors are not imprecise. A vernier caliper is twenty-five times more accurate than a ±0.5 mm tolerance needs. The problem was never precision — it is how little of each lot ever gets measured at all.
The same bad part costs about ₹1 to catch at the receiving gate, ₹10 once it has reached the assembly bench, and ₹100 once it has reached your customer. Nothing about the part changed. Only how long it took someone to notice.
A box of 500 parts arrives. Five get measured. The other 495 are accepted because someone checked five — and if 2 parts in every 100 are bad, those five will miss the problem nine times out of ten. Worse: when a supplier's tool wears out halfway through the run, the five you checked came off the top of the box, so the lot passes in full.
None of these are failures of skill. They are the limits of measuring by hand.
Five parts out of five hundred is all a human can measure without becoming the bottleneck. The rest is accepted on faith — and a defect rate of a few percent is invisible at that sample size.
A drawing carries a dozen toleranced dimensions. Under time pressure an inspector checks the three that experience says matter. The other nine are never measured — which is exactly where escapes collect.
As a supplier’s cutting tool gradually wears, part dimensions may move closer to — and eventually beyond — the allowed tolerance. A sample taken off the top of the box never sees it.
Not a screenshot. The station below is running. Change the part, push the belt faster, raise the defect rate, switch the marker off — every number updates from the parts actually passing the camera.
This interactive demonstration uses simulated parts so visitors can explore the QualityOps workflow without inspection hardware. Change the component, line speed, defect rate, sampling rate, drift, marking, and reject handling to see how the inspection results respond.
The browser demonstration uses illustrative part, supplier, batch, and measurement data. The real QualityOps solution uses Zero Zeta’s proprietary computer-vision model with configured camera, lighting, part, drawing, and tolerance requirements.
Four steps. Only the first is a project — and you do it once per material.
Dimensions and tolerance limits are read out of the CAD drawing (DWG or DXF) once per material code. The drawing becomes an inspection table to check against.
A telecentric camera and LED backlight capture the external part profile as components run past. The operator selects the material code at changeover.
Zero Zeta’s proprietary computer-vision model measures the configured dimensions and classifies the part as PASS, MARGINAL (Review), or FAIL.
Failed parts are UV-marked. Quality teams can review parameter failures, drift trends, sampling escapes, and lot history via plain-language queries.
QualityOps Copilot is designed to function as a reusable, configurable Zero Zeta Industrial AI solution for different components, drawing parameters, and tolerance limits.
Measure every suitable component rather than relying only on a small sample. Good parts keep moving; non-conforming ones are marked.
Compare all visible dimensions configured from the approved drawing, not just the three features an inspector has time for.
See exactly which dimension is failing—and whether parts are running oversize or undersize, linking failures to their downstream consequences.
See when dimensions begin moving towards or beyond the tolerance limit. Tooling wear shows as a trend hours before it crosses.
Apply a UV-visible fluorescent ink mark without adding a mechanical diverter to the described inspection cell.
Retain closed inspection results for supplier scorecards and quality reviews, archiving yield, escapes, and parameter statistics.
Detailed technical specifications of the computer-vision station.
Equipped with a 5 MP telecentric camera and high-intensity LED backlight. Captures silhouette profiles in under 40 ms at line speed.
Approximately ±0.02 mm repeatability. Generates a 25× guard band against typical ±0.5 mm tolerance examples to pass gauge studies.
PASS, MARGINAL (Review), and FAIL logic. Air-actuated inkjet marker fires UV-fluorescent ink 165 mm downstream; UV verification lamp checks marking.
Tolerances are extracted directly from vector-based drawing files (CAD/DWG/DXF). Zero Zeta's model computes exact measurements and runs parameter-level failure analysis against these values.
The described silhouette-based inspection does not directly measure: thread pitch, hidden bore dimensions, bore depth, concentricity, surface finish, plating thickness, hardness, material composition, internal defects, or hidden features that the camera cannot see. These characteristics continue to require an appropriate inspection method.
QualityOps does more than count rejected parts. It helps teams ask which dimension is causing the problem, whether the issue is becoming worse, and how the lot compares with previous batches.
The model interprets and explains the question. Measurements, statistics, and disposition results are computed rather than generated.
Configurable for different organizations, part families, material codes, suppliers, drawings, and tolerance limits.
High-volume, repetitive geometry where sampling is the norm and a single wrong size reaches assembly unnoticed.
Flat parts that present a clean, stable silhouette — hole pitch, edge distance and overall profile in one capture.
Parts with several critical dimensions where inspectors reasonably check only the few they trust to matter.
Contract manufacturing and assembly operations receiving the same material code from several sources, where supplier comparison is the real question.
One material family, one camera station, one honest measurement of what your current sampling is letting through.
