A scanner can measure a small patch consistently while placing that patch incorrectly within a large core model. When a scan is assembled from several stations, registration adds another measurement problem: the relationship between the stations.
On a repetitive lamination stack, adjacent surfaces can look almost interchangeable. A registration algorithm may align their local shapes convincingly even when it shifts the model slightly along a uniform edge. The resulting point cloud looks smooth, but the complete window or limb spacing can be wrong.
Give the model an independent reference
Place stable, identifiable reference targets around the assembly where the chosen system can measure them. Their arrangement should constrain all relevant directions and rotations. Targets concentrated in one small flat region provide weak information about distant geometry and out-of-plane rotation.
Use one set of targets to calculate registration and reserve separate targets for verification. If every target participates in the fit, the reported fit residual mainly describes agreement with the data already used. It does not independently test the model.
1. Register the stations
Maintain enough common reference information between stations to establish their relationship. Preserve station identities and transformation records. Avoid replacing them with a single merged cloud before the registration has been assessed.
2. Check overlap directionally
Compare overlapping surfaces along their normals and along identifiable geometric features. A small normal residual on a flat face may coexist with a large movement parallel to that face. Report the component that matters to the core dimension being evaluated.
3. Close a loop
Return from the last station to a region measured near the start. Compare the independently reconstructed position of the same verification targets. A systematic closure difference suggests accumulated registration error even if each adjacent overlap appears acceptable.
4. Test a long independent dimension
Measure a task-relevant separation with another qualified method or a calibrated artifact suited to that distance. Choose a dimension that spans several stations. A short artifact near the first station cannot adequately challenge drift across the whole assembly.
Retain the registration contribution
The uncertainty of a reported core dimension includes point measurement, feature extraction and coordinate registration. Those contributions can interact; adding a manufacturer’s headline point accuracy to a drawing tolerance does not resolve the task.
Identify where registration matters most. A local burr height may depend primarily on local surface measurement. The distance between two limbs scanned from opposite ends may be highly sensitive to their relative station coordinates. They need different validation tasks.
When closure fails, inspect target stability, target identification and station geometry before deleting inconvenient points. Refit with a defensible reference arrangement, then repeat the reserved-target check. A visually polished cloud is useful for communication, but the acceptance dimension must survive an independent coordinate test.
Compare registration closure with stock camber
The catalogue permits stock camber up to 0.9 mm per 2 m [1]. Assume a scan route over the same 2 m length contains four sequential station transfers, each with a 0.20 mm translational error in the same direction. Worst-case closure error is 0.80 mm, nearly the stock camber scale; treating independent errors statistically would instead give 0.40 mm by root-sum-square. Those models answer different questions [2]. With an assumed 0.50 mm registration budget, the correlated case fails while the independent model appears acceptable. Measure reserved-target closure before diagnosing material shape. The stock camber limit is a comparison scale, not an assembled-window acceptance limit.
Manufacturing controls. Long-strip camber, repeated packet geometry and target stability influence registration. Support movement and clamp changes can deform the assembly between stations; a more flexible software fit cannot restore a single unchanged geometry.
References
- Nippon Steel Corporation, Grain-Oriented Electrical Steel Sheets, catalogue D004je_03_202004, 2020, Specifications; Size and Form Tolerances.
- JCGM, JCGM 100:2008, Evaluation of measurement data — Guide to the expression of uncertainty in measurement, 2008.