EN
News
Large Parts, Tight Tolerances: Dimensional Inspection From 500 To 2000 mm
Overview
Large parts break the assumptions that small-part inspection is built on. A 500 millimetre bracket and a 2000 millimetre frame have the same drawing discipline but different physics: the field of view grows faster than the resolution, the fixture deflects under the part's own weight, and the room temperature becomes a measurement variable.
Big Bird Industrial (formerly Shimada Big Bird Industrial) has delivered dimensional inspection systems for vehicle structures and powertrain components since 2001. This article covers the four constraints that govern large-part inspection — field of view, stitching, thermal behaviour and datum stability — and how each one changes the hardware choice.
The practical consequence is that large-part systems are never a scaled-up version of a small-part cell. They are a different architecture, usually a hybrid of photogrammetry, structured light and probing.
1. Field Of View Versus Resolution
A single camera that covers two metres has a resolution measured in fractions of a millimetre per pixel at best. If the tolerance is plus or minus 0.2 millimetres, that resolution is marginal; if the tolerance is plus or minus 0.05 millimetres, it is useless.
The arithmetic is unforgiving: pixels per millimetre scales linearly with sensor width and inversely with field width. Covering four times the width with the same sensor quarters the resolution. There is no optical trick that beats this, which is why large-part systems move the camera, move the part, or use multiple cameras.
Moving the camera is the most common solution and also the one that introduces the most error, because every move has to be measured or calibrated. Moving the part is often worse: a two-metre frame is heavy, and every rotation introduces a new setup error.
2. Stitching Is A Measurement, Not A Montage
When several images are combined into one coordinate system, the result is only as good as the alignment between them. Two approaches dominate.
Mechanical referencing
The camera is on a calibrated stage and its position is known from encoders. Accuracy depends on the stage, on its straightness and on its thermal stability. It is predictable and auditable, and it is expensive at large travels.
Photogrammetric referencing
Coded targets are placed on or around the part, and camera pose is solved from them. Accuracy depends on target distribution and on the number of views, and it degrades in the middle of a large target field if the geometry is weak.
Both methods need an artefact to verify against. A two-metre artefact is itself a metrology project: it has to be stable, calibrated, and stored somewhere that does not move it between uses.
3. Temperature Is Part Of The Measurement
Steel expands by roughly eleven parts per million per degree. On a two-metre part, a five-degree change moves a feature by about a tenth of a millimetre — the same order as the tolerance. On aluminium the number is roughly twice as large.
That is why large-part inspection rooms are temperature-controlled and why the part is allowed to soak before measurement. A part measured straight off a machining line carries a thermal gradient, and the gradient distorts the geometry as well as the scale.
The pragmatic approach is to record part temperature at measurement and to apply a compensation model where the material is known. Without that record, a drift in results cannot be separated from a drift in the room.
4. The Fixture Deflects Too
A large part on a fixture loads that fixture. Steel frames bend under their own weight, and a fixture designed for stiffness at the supports may still sag in the middle. The datum moves with the sag, and the measurement reports a part that is in tolerance against a datum that is not where the drawing assumes.
Three techniques help. Support the part at the points the drawing specifies, which is often three points plus two, rather than at convenient locations. Measure the fixture with the part removed at a defined interval, so sag is a known quantity. And prefer a scanning approach that captures the whole surface rather than a set of discrete points, because the whole-surface data shows the sag pattern directly.
5. Choosing The Technology
The three families that matter for large parts have different strengths, and most real systems mix them.
| Technology | Strength | Limit |
|---|---|---|
| Photogrammetry | Very large volumes, portable | Needs targets, sparse data |
| Structured light | Dense surface, good for form | Sensitive to ambient light, view count |
| Laser profile | Robust on reflective metal | Line-by-line, motion required |
| Tactile probe | Reference accuracy | Slow, point-wise |
A common architecture uses photogrammetry to establish the global frame, structured light for the surfaces that need form data, and occasional tactile checks as a reference. The combination is more accurate than any single method and easier to defend in an audit.
6. Cycle Time And Inline Feasibility
Large-part inspection is often offline because the measurement is slow and the parts are few per hour. That is a design choice, not a law: an inline station can work if the number of features is small and the cycle budget is generous.
The decision rule is to compare the inspection time with the takt. If inspection takes longer than takt, the station cannot be inline without parallel lanes, and parallel lanes multiply cost faster than they multiply throughput.
| Scenario | Typical approach | Reason |
|---|---|---|
| Few features, tight tolerance | Inline, fixed cameras | Fast and repeatable |
| Many features, tight tolerance | Offline, scanning cell | Accuracy over speed |
| Full-surface form | Offline, structured light | Data volume |
7. Verification And Drift
A large-part system drifts for reasons a small cell never sees: the room changes temperature, the artefact itself moves, and the stage wears over metres rather than millimetres. Verification has to be scheduled against those time constants, not against a generic monthly rule.
The workable pattern is a short daily check on a stable artefact, a fuller weekly check on a reference part, and an annual recalibration that includes the stage geometry. The daily check catches a bumped camera; the annual check catches a slow loss of squareness that no single-part measurement would reveal.
Recording the check results as a trend is what turns verification into prediction. A repeatability figure that creeps from twenty to forty micrometres over six months is a maintenance signal long before it becomes a rejection.
8. Robot, Gantry Or Fixed Rig
Three mechanisms move a sensor over a large part, and the choice is mostly about coverage pattern. A robot is flexible within its reach and needs a rail to go beyond it. A gantry covers a rectangular volume predictably and is heavy. A fixed rig with many cameras has no moving parts and no dynamic error, at the cost of hardware per view.
For parts that are large in one dimension, a rail-mounted robot is usually the best compromise: the rail provides the travel, the arm provides the angles. For parts that are large in two dimensions, a gantry or a fixed rig scales better.
Whichever is chosen, the calibration has to cover the mechanism, not just the camera. A robot that is accurate to a tenth of a millimetre at its wrist is not accurate to a tenth at the end of a two-metre part unless the kinematic model has been verified against an artefact at that distance.
Frequently Asked Questions
Can a 2000 mm part be measured to 0.1 mm inline?
Yes, for a small number of features with a fixed camera layout and a stable fixture. The accuracy comes from the layout and the calibration, not from the sensor alone.
Do targets have to be attached to the part?
Either the part or the fixture, depending on the method. Targets on the fixture are easier to manage but do not capture part distortion; targets on the part capture distortion but add handling.
How often does a large artefact need recalibration?
On a schedule set by the accuracy requirement and by how often it is moved. A stored artefact that never moves drifts slowly; one that is transported needs verification before each use.
Is thermal compensation reliable?
It is reliable when the material and the gradient are known, which is rarely true on a production floor. Measurement in a controlled room is more dependable than compensation in an uncontrolled one.
Conclusion
Large-part inspection is a metrology problem with an automation wrapper. Field of view sets the resolution, stitching sets the accuracy, temperature sets the repeatability, and the fixture sets whether any of it survives a shift.
The systems that work are hybrids, calibrated against artefacts, and honest about which features are measured inline and which are audited offline. The ones that fail promise a single sensor and a single number, and deliver neither.
