Spiral bevel gears transmit motion between intersecting shafts through curved teeth that engage gradually and can carry high torque with relatively low noise. In helicopter tail-rotor drive trains, this combination of high speed, heavy load, and strict acoustic requirements pushes tooth-flank accuracy into the micrometre range. A representative aerospace case is a 31/38-tooth pair with a 52° shaft angle, a right-hand pinion, and a left-hand wheel. Because the shaft angle is not 90°, the pair is non-orthogonal, and many simplifying assumptions used for orthogonal bevel gears no longer hold.
The manufacturing difficulty is not simply producing a nominal tooth form. It is making the ground flank match a measured target while keeping every machine adjustment within a physically usable range. Wheel-profile error, kinematic error of the machine axes, thermal distortion, and fixture effects all leave a residual normal deviation between the actual and theoretical flanks. The control problem is therefore an inverse problem: given a measured error field, find a small set of machine-setting increments that reduces that field without destabilizing the grinding process.
The Non-Orthogonal Grinding Problem
Non-orthogonal spiral bevel gears are used where the layout of a transmission requires an axis angle other than 90°, as in the 52° helicopter tail-drive configuration. The tooth surface is a complex envelope generated by a rotating cutter and a timed workpiece motion. In free-form, or full-process, grinding, the entire flank is produced in one setup, which improves efficiency and batch consistency but makes the accuracy of the machine-setting model critical.
The duplex helical method developed by Gleason adds a controlled helical translation along the bed direction while the cutter rotates with the cradle. This additional motion changes the relative velocity between cutter and workpiece and introduces a helical correction coefficient into the machine-setting vector. When that coefficient is zero, the kinematic model reduces to the conventional case; when it is nonzero, the generated flank is modified in a way that cannot be represented by the usual orthogonal assumptions alone.
The correction problem is overdetermined in principle: a measured grid contains many more points than the number of available machine settings. The difficulty is that the relationship between settings and normal error is strongly nonlinear, and a direct node-by-node inversion can produce an ill-conditioned or singular system. A practical method must therefore combine an accurate geometric model, a compact error representation, and a numerically stable update rule.
Envelope Model and Machine-Setting Vector
The grinding simulation treats the cutter as a parameterized body and the machine as a chain of coordinate transformations. For the duplex helical process, the machine-setting vector can be written as
$$ \boldsymbol{\xi} = [q, S_r, E_M, X_D, X_B, \gamma_m, \sigma, \zeta, m, H_{\mathrm{eli}}]^{T} $$
where \(q\) is the angular cutter position, \(S_r\) the radial cutter position, \(E_M\) the vertical wheel offset, \(X_D\) the bed position, \(X_B\) the horizontal wheel position, \(\gamma_m\) the installation angle, \(\sigma\) the cutter tilt angle, \(\zeta\) the cutter swivel angle, \(m\) the roll ratio, and \(H_{\mathrm{eli}}\) the helical correction coefficient.
The generated surface family is obtained by mapping the cutter point \(\mathbf{r}_c(\mu,\theta)\) into the workpiece frame:
$$ \mathbf{r}(\mu,\theta,\phi;\boldsymbol{\xi}) = \mathbf{M}_{bt}(\boldsymbol{\xi};\phi)\,\mathbf{r}_c(\mu,\theta) $$
Here \(\phi\) is the cradle angle and \(\mathbf{M}_{bt}\) collects the cutter-to-workpiece transformations. The actual tooth surface is the envelope of this family and must satisfy the meshing condition
$$ \mathbf{n}\cdot\mathbf{v}=0 $$
where \(\mathbf{n}\) is the surface normal and \(\mathbf{v}\) is the relative velocity between cutter and workpiece. The resulting surface equation is implicit and highly nonlinear, so the flank is discretized on a grid bounded by the toe, heel, tip, and root. The grid is set back from the tooth edges by 5% of tooth height and 10% of face width to reduce edge effects.
This model is geometric rather than empirical. It does not assume that a particular machine setting produces a fixed error; instead, it recomputes the envelope for each candidate setting vector. That property is what allows the same correction procedure to be applied to different initial errors and different gear sizes.
Measurement-Based Error Objective
The actual flank is measured on a coordinate measuring machine using a touch-trigger probe and a predefined grid. Each measured point is compared with the corresponding theoretical point along the unit normal. The signed normal deviation at point \(i\) is
$$ h_i = \left[\mathbf{p}_i^{\mathrm{CMM}} – \mathbf{p}((\mu,\theta,\phi)_i,\boldsymbol{\xi})\right]\cdot \mathbf{n}_i $$
where \(\mathbf{p}_i^{\mathrm{CMM}}\) is the measured point, \(\mathbf{p}((\mu,\theta,\phi)_i,\boldsymbol{\xi})\) is the theoretical point on the model flank, and \(\mathbf{n}_i\) is the unit normal at that point. A positive value means the measured point lies outside the theoretical flank along the normal; a negative value means it lies inside.
The correction target is not to force every point to zero independently. The error field is represented as a difference surface, which reduces the number of independent error modes and avoids an ill-conditioned node-by-node inverse. The machine-setting vector is then chosen to minimize the least-squares norm of the normal deviations:
$$ \boldsymbol{\xi}^{*} = \arg\min_{\boldsymbol{\xi}} \frac{1}{2}\mathbf{h}^{T}\mathbf{h} $$
This formulation preserves the physically meaningful quantities: the measured coordinates, the theoretical surface, the normal direction, and the machine settings. It also makes clear that the optimization is local to the current grinding setup. The result is a corrected setting vector, not a universal recipe for every spiral bevel gear.
Sensitivity-Guided Levenberg–Marquardt Correction
Not all machine settings affect the flank in the same way. To quantify their influence, the error model is differentiated with respect to the setting vector. The Jacobian matrix is
$$ \mathbf{J} = \frac{\partial \mathbf{h}}{\partial \boldsymbol{\xi}}, \qquad \mathbf{h} \approx \mathbf{J}\,\Delta\boldsymbol{\xi} $$
Each entry \(J_{ij}\) gives the change in normal error at grid point \(i\) produced by a unit change in setting \(j\). The Jacobian therefore acts as a sensitivity map: it identifies which settings control which error modes and prevents the solver from applying large corrections to settings that have little leverage.
Because the system is nonlinear and may be rank-deficient or poorly scaled, the correction is computed with an improved Levenberg–Marquardt iteration:
$$ \Delta\boldsymbol{\xi}^{(k+1)} = \Delta\boldsymbol{\xi}^{(k)} – \left(\mathbf{J}_k^{T}\mathbf{J}_k + \lambda_k\mathbf{I}\right)^{-1}\mathbf{J}_k^{T}\mathbf{h}_k, \qquad \lambda_k \ge 0 $$
Here \(\mathbf{h}_k\) is the residual vector at iteration \(k\), \(\mathbf{J}_k\) is the corresponding sensitivity matrix, and \(\lambda_k\) is a nonnegative damping parameter. A large \(\lambda_k\) produces a short, gradient-like step; a small \(\lambda_k\) approaches a Gauss–Newton step. The damping is adjusted within a trust region, so that each trial correction is evaluated before it is accepted.
Two practical constraints are imposed on the update. First, the machine adjustments are screened and bounded at each iteration, because a geometrically valid solution that requires an impossible machine motion is not useful. Second, the trust-region logic considers both the cumulative error and the peak error. A step that lowers the total error but raises the maximum deviation is not automatically accepted; in gear grinding, the peak deviation often matters more because it controls local contact and clearance.
Case Study: 31/38-Tooth Pair at 52° Shaft Angle
The case pair has 31 pinion teeth and 38 wheel teeth, a 52° shaft angle, and a 32.00 mm face width. The pinion is right-hand and the wheel is left-hand. The working flanks are the pinion concave side and the wheel convex side, so both surfaces are corrected. The pinion and wheel were ground with cutters of 305.040 mm and 304.80 mm diameter, respectively.
| Blank parameter | Pinion | Wheel |
|---|---|---|
| Number of teeth | 31 | 38 |
| Face width | 32.00 mm | 32.00 mm |
| Pressure angle | 20° | 20° |
| Shaft angle | 52° | 52° |
| Addendum | 5.90 mm | 3.44 mm |
| Dedendum | 4.65 mm | 6.30 mm |
| Outer cone distance | 208.81 mm | 208.81 mm |
| Pitch cone angle | 23.16667° | 28.83333° |
| Face cone angle | 24.26667° | 29.60000° |
| Root cone angle | 22.40° | 27.73333° |
| Hand | RH | LH |
The initial and corrected machine settings for the two working flanks are summarized below. The corrected values are one solution set that satisfies the geometric objective; they should not be read as a fixed recipe for other spiral bevel gear pairs.
| Setting | Pinion initial | Pinion corrected | Wheel initial | Wheel corrected |
|---|---|---|---|---|
| Radial cutter position \(S_r\) / mm | 163.3098 | 161.3222 | 163.52611 | 161.653 |
| Angular cutter position \(q\) / ° | 50.4968 | 50.50201 | 49.8142 | 49.8051 |
| Cutter tilt angle \(\sigma\) / ° | 3.44 | 3.440614 | 0.5985 | 0.6173 |
| Cutter swivel angle \(\zeta\) / ° | 2.5969 | 2.596959 | 14.88 | 14.88 |
| Vertical wheel offset \(E_M\) / mm | 0.12558 | 0.05488 | 0.0000 | 0.0512 |
| Installation angle \(\gamma_m\) / ° | 20.13 | 20.1016 | 23.73 | 23.754 |
| Horizontal wheel position \(X_B\) / mm | 0.43873 | 0.438954 | 0.00000 | −1.3304 |
| Bed position \(X_D\) / mm | −3.1264 | −3.1439 | −3.86914 | −7.07774 |
| Roll ratio \(m\) | 2.5258999 | 2.512648 | 2.075437 | 2.080307 |
| Helical correction \(H_{\mathrm{eli}}\) | 6.68482 | 6.684794 | 0.00000 | 0.00000 |
After four plotted correction steps, the error metrics fell substantially. The cumulative error decreased to 46.765% of its initial value on the pinion concave flank and to 14.8% on the wheel convex flank. The maximum error decreased to 3.77% and 10.62% of the respective initial values. The final maximum residual was no more than 15 µm. These percentages are relative to different initial error levels, so they should not be compared as absolute magnitudes.
| Working flank | Cumulative error, final as % of initial | Maximum error, final as % of initial |
|---|---|---|
| Pinion concave | 46.765% | 3.77% |
| Wheel convex | 14.8% | 10.62% |
The residual surfaces are relatively uniform rather than scattered, which indicates that the correction did not introduce high-order geometric noise. On the pinion concave flank, the larger residuals remain near the four corners of the grid. On the wheel convex flank, negative residuals dominate. Both examples reached acceptable settings within ten correction cycles, while the plotted histories show four steps.
Engineering Interpretation and Limits
The method’s main contribution is procedural rather than numerical. It treats machine settings as a compensation vector that can be updated from measured flank error, instead of treating the setup as a fixed recipe. Sensitivity analysis concentrates the correction on influential settings, and the trust-region Levenberg–Marquardt step keeps the inverse problem stable when the Jacobian is poorly conditioned. The bounded, screened updates also keep the result inside the practical adjustment range of the machine.
The scope is deliberately narrow. The optimization controls geometric normal deviation only. It does not include loaded tooth contact, transmission error, noise, or durability. Machine settings that improve geometric conformity can still shift the contact pattern, so a complete aerospace qualification would need loaded contact analysis and, ultimately, rig or flight-representative testing. The method is also an offline, between-pass correction loop, not a real-time servo control within a single grinding pass.
Measurement uncertainty, thermal drift, and machine repeatability remain outside the model. The correction is local to the initial setting and to the chosen measurement grid, so different initial errors may lead to different corrected settings. That is not a defect unique to this approach; it is a property of nonlinear inverse problems in flank grinding. The practical value lies in reducing the number of trial-and-error iterations and in making each iteration traceable to measured normal deviations.
Conclusion
For non-orthogonal aerospace spiral bevel gears, geometric flank accuracy can be improved by a measurement-driven correction of the machine-setting vector. The duplex helical grinding model supplies the theoretical envelope, the coordinate measurement supplies the normal error field, and the sensitivity-weighted Levenberg–Marquardt update supplies a bounded correction. In the 31/38-tooth, 52° case, this procedure reduced cumulative and peak errors substantially on both working flanks and left a residual maximum of no more than 15 µm.
The result supports a narrower engineering conclusion: machine settings should be treated as an adaptive compensation layer in spiral bevel gear grinding, not as a one-time setup. The same logic can be extended to other non-orthogonal pairs, provided that the geometric model, measurement grid, and machine constraints are updated for the specific application.