Spiral bevel gears transmit motion between intersecting shafts with curved teeth that engage gradually, giving high load capacity, smooth motion, and compact packaging in aerospace, automotive, and industrial drives. Their tooth surfaces are not defined by a simple analytic profile. The flank is the envelope of a cutter family produced by a coordinated machine motion, so the same nominal gear dimensions can be manufactured with different cutter and machine settings, producing different local curvature, contact patterns, and transmission error. The design problem is therefore also a manufacturing problem: tooth geometry and contact behavior are both consequences of machine settings.
Traditional development relies on trial cutting, contact-pattern inspection, and repeated adjustment of machine settings. For a high-performance gear pair available only as hardware, the dimensions can be measured, but the settings that produced the flank may be unavailable. A triangulated surface reconstructed from scan data may look correct while remaining impossible to manufacture or analyze without a setting set. The useful reverse-design problem is to recover a parametric manufacturing model, not merely a point-cloud replica.
Structured-light acquisition of the tooth surface
The measurement chain uses an ATOS Triple Scan structured-light scanner. Fringe patterns are projected onto the workpiece and recorded by stereo CCD cameras; the three-scan arrangement captures three viewing directions in one measurement cycle, and reference targets stitch multiple views into a common point cloud. The scanner used here has a stated overall accuracy of ≤0.01 mm, point spacing of 0.01–0.61 mm, and CCD resolution of 3692×2472 pixels. The gear pair fits within a 150–300 mm envelope, which is compatible with the scanner’s multi-view field.
Before scanning, the tooth flanks and datum features are cleaned and coated with a thin developer to suppress specular reflection. Reference targets are applied so that individual views can be registered. The result is a dense point cloud containing the tooth flanks, the gear shaft, and selected machined surfaces. Because the mounting face and the tooth flanks are often not visible in the same view, a machined circular face visible with the flank is used as a transfer datum; its axial offset to the mounting face is measured separately with a height gauge.
Raw point clouds contain outliers, redundant points, and background data. Statistical outlier removal deletes points whose mean distance to their neighbors exceeds a threshold. Voxel-grid downsampling replaces points in each grid cell with the point nearest the cell centroid, reducing millions of points while preserving the flank shape. Manual segmentation in CloudCompare then isolates the target flanks and datum face. These steps do not improve the physical measurement, but they make the subsequent registration and surface fitting stable.
Registration from the measurement frame to the design frame
Point-cloud coordinates are initially expressed in a measurement coordinate system (MCS) defined by the scanner. Tooth-surface modeling, blank definition, and tooth contact analysis use a design coordinate system (DCS), whose origin and axes are tied to the gear axis and pitch apex. Registration must therefore recover the pitch apex position and align the gear axis with the DCS z-axis.
The pitch apex is not directly visible. It lies on the gear axis at a distance equal to the mounting distance from the mounting face. The registration procedure uses a machined circular boundary on the shaft as a transfer feature. A KD-tree accelerates k-nearest-neighbor queries, and local neighborhood statistics identify boundary points. A least-squares spatial circle fit then gives the circle center and normal: a plane is fitted first, the boundary points are projected into that plane, a two-dimensional circle is fitted, and the center and radius are transformed back to the original frame.
With the feature center and normal known, the cloud is rotated so that the normal is parallel to the DCS z-axis, translated so that the feature center returns to its original position, shifted to the mounting face by the measured offset, and finally translated along the axis by the mounting distance. The registered cloud can then be compared with design data and used for blank-parameter inversion. A coordinate-measuring machine (CMM) check with a 5×9 grid on the flank confirmed that the registered optical cloud and the contact-measured points occupied the same design space.
Recovering tooth blank parameters from the axial projection
Once the cloud is in the DCS, the tooth surface is projected onto the axial section. For a point \((X,Y,Z)\), the projection uses
$$R=\sqrt{X^2+Y^2},\qquad L=Z$$
which converts the three-dimensional flank into a two-dimensional boundary problem. Boundary points are classified into toe, heel, tip, and root sets, and least-squares lines are fitted to each set. The resulting four lines define the projected tooth boundary and provide the geometric inputs for blank inversion.
The pitch cone angles follow from the tooth counts and shaft angle. For a pair with shaft angle \(\Sigma\),
$$\tan\gamma_2=\frac{\sin\Sigma}{z_1/z_2+\cos\Sigma},\qquad \gamma_1=\Sigma-\gamma_2$$
The cone distance is obtained from the projected pitch-line geometry, the pitch diameter from \(d_g=2A_0\sin\gamma_g\), and the module from \(m=d_g/z_g\). Face width, addendum, dedendum, face angle, and root angle follow from the projected boundaries. Because the optical cloud, registration, and boundary fitting all carry small errors, the raw values from the pinion and wheel are not accepted independently. The pair is corrected to a common module, cone distance, face width, and whole depth, with the toe and heel lines parallel and perpendicular to the pitch-cone line, and with the pinion root angle plus wheel face angle equal to 90°.
| Parameter | Pinion | Wheel |
|---|---|---|
| Number of teeth | 25 | 37 |
| Shaft angle | 90° | 90° |
| Pitch cone angle | 34.05° | 55.95° |
| Module | 4 mm | 4 mm |
| Hand | Right | Left |
| Cone distance | 89.30 mm | 89.30 mm |
| Pitch diameter | 100 mm | 148 mm |
| Face width | 25 mm | 25 mm |
| Whole depth | 7.6 mm | 7.6 mm |
| Heel addendum | 4.25 mm | 2.55 mm |
| Heel dedendum | 3.35 mm | 5.05 mm |
| Root cone angle | 31.90° | 52.72° |
| Face cone angle | 37.28° | 58.10° |
| Addendum angle | 3.23° | 2.15° |
| Dedendum angle | 2.15° | 3.23° |
The initial inversion produced small inconsistencies between the two members, including a wheel module of 3.99 mm and cone distances of 89.24 mm and 89.15 mm. These were not treated as independent measurements. The pairing constraints forced a common 4 mm module, a common 89.30 mm cone distance, and matched face widths and depths. This correction preserves the measured geometry while enforcing the kinematic requirement that the two pitch cones share an apex.
Initial machine settings and the UMC surface model
The blank parameters define the boundary, but the flank shape is controlled by cutter geometry and machine settings. The relevant cutter parameters include cutter diameter, pressure angle, and edge radius. The machine settings include radial cutter position, angular cutter position, cutter tilt, cutter swivel, vertical wheel position, horizontal wheel position, sliding base, mounting angle, and roll ratio. In the case pair, the wheel is generated and the pinion is machined with a modified roll.
The surface model uses the universal motion concept (UMC). Eight motion parameters describe the relative position and orientation of cutter and workpiece, while modified roll coefficients and a sliding-base function describe the generating motion. The tooth surface is the envelope of the cutter family that satisfies the meshing equation. For a given setting vector, the model computes discrete flank points on a 5×9 grid: nine points along the face width and five along the tooth height.
The settings calculated directly from the recovered blank are only a starting point. They come from an approximate boundary, a registered point cloud, and a nonlinear machine model; the resulting surface will not coincide with the measured flank. The initial and corrected settings for the pinion convex flank show the scale of the adjustment.
| Setting | Initial | After correction |
|---|---|---|
| Cutter diameter \(D\) (mm) | 149.0970 | 151.8920 |
| Cutter pressure angle \(\alpha\) (°) | 21.00 | 21.25 |
| Radial cutter position \(S_R\) (mm) | 69.6218 | 70.1027 |
| Angular cutter position \(q\) (°) | 59.11 | 59.11 |
| Cutter tilt \(\sigma\) (°) | 0 | 0 |
| Cutter swivel \(\zeta\) (°) | 27.74 | 27.74 |
| Vertical wheel position \(E_M\) (mm) | -0.1016 | -0.1016 |
| Mounting angle \(\gamma_m\) (°) | 31.93 | 31.93 |
| Horizontal wheel position \(X_D\) (mm) | -2.7037 | -1.4947 |
| Sliding base \(X_B\) (mm) | 1.4299 | 0.7926 |
| Roll ratio \(m_C\) | 1.7428 | 1.7428 |
Error-surface parameterization and sensitivity-based correction
The measured and computed flanks are compared on the same 5×9 grid. A small angular mismatch between the two point sets is removed by rotating the theoretical surface about the gear axis until the flank midpoints coincide. This separates pitch error from shape error. The remaining normal deviation is
$$h_{ij}=(p_{ij}-r’_{ij})\cdot n_{ij}$$
where \(p_{ij}\) is the measured point, \(r’_{ij}\) is the rotated theoretical point, and \(n_{ij}\) is the theoretical normal. Instead of treating all 45 grid values as independent optimization targets, the deviation surface is fitted by a second-order polynomial:
$$Z(X,Y)=a_0+a_1X+a_2Y+a_3XY+a_4X^2+a_5Y^2$$
With the pitch error removed, \(a_0\) is zero. The first-order terms \(a_1\) and \(a_2\) describe tilt along the face and tooth-height directions; \(a_3\) describes twist; \(a_4\) and \(a_5\) describe crowning in the two directions. The coefficient vector is a compact feature description of the error surface. A Levenberg–Marquardt fit determines the coefficients from the discrete deviation points. This avoids an ill-conditioned 45-parameter correction problem and gives the optimization a five-variable shape target.
Sensitivity is evaluated by perturbing each machine setting, regenerating the surface, fitting the coefficient vector, and ranking the resulting changes. For the pinion convex flank, the five dominant settings were cutter size, cutter pressure angle, radial cutter position, horizontal wheel position, and sliding base. A linearized 5×5 system then maps desired coefficient changes to setting corrections. The model is regenerated, the residual is refitted, and the process is repeated until the correction converges.
| Stage | Maximum normal deviation (μm) |
|---|---|
| Initial settings | 16.3 |
| After first correction | 5.7 |
| After second correction | 1.5 |
| After third correction | 1.4 |
| After fourth correction | 1.4 |
The first correction reduced the maximum model residual by about 65%, and the second by about 91%. The residual then stabilized at 1.4 μm. The same procedure gave 3.6 μm for the pinion concave flank, 2.3 μm for the wheel concave flank, and 3.7 μm for the wheel convex flank.
Geometric and contact verification
The final check used a CMM independent of the optical scan. A GLOBAL STATUS 575 machine with PC-DMIS and QUINDOS software measured the physical flanks on a 5×9 grid, using 0.5 mm or 1 mm styli. The grid was kept inside the active flank by margins of about 5% in tooth height and 10% in face width. The CMM had a stated resolution of 0.1 μm and a maximum permissible probing error of 2.5 μm. The measured points were compared with points computed from the corrected machine settings.
| Flank | Maximum deviation (μm) |
|---|---|
| Wheel concave | 3.0 |
| Wheel convex | 6.7 |
| Pinion concave | 2.3 |
| Pinion convex | 5.7 |
These values answer a different question from the 1.4 μm model residual. The 1.4 μm is the difference between the corrected model and the registered optical cloud; the 2.3–6.7 μm values are independent CMM comparisons against the physical part. The latter include measurement, registration, and surface-fitting uncertainty, and are therefore the more conservative indication of geometric agreement.
Contact behavior was checked by tooth contact analysis (TCA) and by a rolling inspection. The rolling test was run at a load of 9.6658 N·m. Contact was identified by a dye-transfer threshold corresponding to a gap of 0.00635 mm. The measured contact pattern occupied about 30% of the available flank area, with no edge contact or stress concentration. The wheel concave pattern sat near the middle of the flank with a slight bias toward the toe and tip; the pinion convex pattern sat near the middle with a bias toward the toe. The computed TCA pattern at the same condition showed a similar size, shape, and position.
The agreement supports functional equivalence at the tested condition. It does not by itself establish fatigue life, misalignment robustness, or performance over the full load spectrum. Those require separate durability and sensitivity studies.
Implementation value and open issues
The main value of this reverse-design chain is that it returns a manufacturing-ready setting set. A reconstructed CAD surface can support visualization and limited geometric comparison, but it cannot be used to cut or grind a replacement gear without cutter geometry and machine motions. By coupling point-cloud registration, blank inversion, a UMC surface model, and error-surface correction, the workflow connects the measured flank back to the settings that generate it. That connection can support replacement of imported hardware, process feedback between design and manufacturing, and diagnosis of why a manufactured flank departs from its intended contact pattern.
The current workflow still has clear limits. The transmission of scanner error, registration error, and boundary-fitting error into the final settings has not been quantified as a formal uncertainty model. Flank segmentation and feature selection remain partly manual, which makes the process dependent on operator judgment. The case evidence covers one aviation gear pair, one structured-light scanner, and a standard tooth form; large-module gears, non-standard tooth forms, and other optical instruments have not been demonstrated. The contact check is a single-load comparison, not a durability test.
These limits point to the next technical steps: automated flank and datum-feature extraction, uncertainty propagation from point cloud to machine settings, and validation across a wider range of gear sizes and tooth forms. If those steps are taken, point-cloud reverse design can become less of a one-off reconstruction and more of a repeatable bridge between a measured spiral bevel gear and its manufacturing process.