Non-orthogonal spiral bevel gears transmit torque between intersecting axes that are not at 90°. In helicopter tail reduction gears, the combination of high speed, heavy load and strict acoustic requirements pushes the tooth flank into a tight tolerance band. The teeth are curved, the contact pattern is sensitive to small form deviations, and grinding is the operation that determines whether the measured flank can be brought close enough to the design flank. Tool-profile errors, machine-axis motion errors and thermal distortion make perfect matching impossible, so the practical problem is not whether deviation exists but how to adjust the machine settings after measurement so that the next part is closer to the design surface.
Why the grinding problem is difficult
Two features distinguish this case from ordinary flank correction. First, the shaft angle is non-orthogonal. In the example below, the pinion and wheel pitch angles are 23.16667° and 28.833333°, and their sum is the 52° shaft angle. The machine kinematics therefore cannot be reduced to the orthogonal formula. Second, full-process grinding, also called the duplex helical method in Gleason terminology, uses a helical cutter motion to generate the flank in a completing process. The method is efficient and gives good batch consistency, but its production model is not openly published. A kinematic reconstruction is needed before any correction can be computed.
Without such a model, correction becomes trial and error: grind, measure, shift settings, grind again. Each trial consumes a part and machine time, and the direction of the shift is not obvious because several machine settings can produce similar changes in flank form. A model-based correction replaces part of that search with a sensitivity calculation and a constrained least-squares update.
Kinematic model of full-process grinding
The cutter is a parabolic-profile grinding wheel. Its active profile has an arc defined by a pressure angle \(\alpha_g\), a profile-modification radius \(\delta_g\), a blade arc length \(s_g\) and a phase angle \(\theta\). The reference cutter radius is
$$R_g=R_u\pm\frac{P_w}{2},$$
where \(R_u\) is the mean cutter radius, \(P_w\) is the point width, and the sign distinguishes the convex and concave sides. The parabolic term \(\delta_g s_g^2\) makes the local pressure angle vary along the blade, which provides a controllable profile modification. A circular fillet joins the active arc to the tip.
The kinematic chain connects four frames: the cutter frame \(O_t\), the auxiliary frame \(O_a\), the blank frame \(O_b\) and the machine frame \(O_m\). Their overall transformation is
$$M_{bt}=M_{ba}M_{am}M_{mt}.$$
The machine-setting vector contains ten components:
$$\mathbf x=(q,S_r,E_M,X_D,X_B,\gamma_m,\sigma,\zeta,m,H_{eli}),$$
where \(q\) is the angular cutter position, \(S_r\) the radial cutter position, \(E_M\) the vertical offset, \(X_D\) the bed position, \(X_B\) the horizontal offset, \(\gamma_m\) the installation angle, \(\sigma\) the cutter tilt, \(\zeta\) the cutter swivel, \(m\) the roll ratio and \(H_{eli}\) the helical correction coefficient. The cradle rotation \(\phi\) is the motion parameter.
Duplex helical grinding adds a translation along the bed direction proportional to cradle rotation. In the auxiliary-to-blank transformation, the vertical term is therefore modified from \(-X_B\) to \(-X_B-H_{eli}\phi\):
$$M_{am}=
\begin{bmatrix}
1&0&0&0\\
0&1&0&E_M\\
0&0&1&-X_B-H_{eli}\phi\\
0&0&0&1
\end{bmatrix}.$$
This screw motion also changes the relative velocity between cutter and blank. An additional vertical component \(H_{eli}/R_a\) is superimposed on the velocity \(\mathbf v_{c-b}\) that would exist without the helical motion:
$$\mathbf v_{HL}=\mathbf v_{c-b}+\mathbf v_{XB},\qquad
\mathbf v_{XB}=
\begin{bmatrix}0\\0\\H_{eli}/R_a\end{bmatrix}.$$
The tooth surface is the envelope of cutter positions that satisfy the meshing condition. With \(\mathbf F\) denoting the surface generated by the cutter and \(\mathbf n\) its normal, the condition is
$$f(\mathbf x;\phi,s_g,\theta)=\mathbf n\cdot\mathbf v=
\left(\frac{\partial \mathbf F}{\partial \mu}\times
\frac{\partial \mathbf F}{\partial \theta}\right)\cdot
\frac{\partial \mathbf F}{\partial \phi}=0,$$
and the surface itself is
$$\mathbf F(s_g,\theta,\phi;\mathbf x)=\mathbf M_{bt}(\mathbf x;\phi)\,\mathbf r_c(s_g,\theta).$$
The resulting equations are implicit and strongly nonlinear. Closed-form solution is not practical, so the flank is solved point by point on a grid bounded by the toe, heel, tip and root. This grid becomes the reference surface for the error calculation.
From measured deviation to a machine-setting correction
The actual flank is measured on a coordinate measuring machine. For grid point \(i\), the normal deviation is
$$h_i(\mathbf x)=\left[\mathbf p^*_{\mathrm{CMM},i}-\mathbf p_i(\mathbf x)\right]\cdot \mathbf n_i(\mathbf x),$$
where \(\mathbf p^*_{\mathrm{CMM},i}\) is the measured point, \(\mathbf p_i\) is the corresponding point on the design flank generated by the kinematic model, and \(\mathbf n_i\) is the unit normal at that point. The design flank is the reference; the measured points describe the current actual flank. The error vector \(\mathbf h\) has one component per grid point.
The correction target is a least-squares problem:
$$\min_{\mathbf x}\;\frac12\mathbf h^T(\mathbf x)\mathbf h(\mathbf x).$$
The deviation field is represented as a difference surface rather than as a collection of unrelated node values. This keeps the optimization smaller and reduces the ill-conditioning that arises when every measured point is treated as an independent target. The difference surface also makes the systematic part of the flank error visible, which is the part that machine-setting changes can correct.
Different settings have different authority over the flank. A first-order sensitivity matrix is formed by differentiating the normal deviation with respect to each setting at each grid point:
$$J_{ij}=-\frac{\partial \mathbf r_i}{\partial x_j}\cdot \mathbf n_i,
\qquad \mathbf h\approx M\,\Delta\mathbf x.$$
Here \(M=J\in\mathbb R^{m\times 10}\), \(m\) is the number of measured grid points, and the system is overdetermined because there are more points than settings. The columns of \(M\) are not orthogonal: a cutter tilt and a vertical offset, for example, can both change the flank slope. Direct linear least squares can therefore produce a numerically unstable correction, especially when the surface response is strongly nonlinear.
Damped least-squares solution with trust-region control
The setting correction is solved with a damped Gauss-Newton, or Levenberg-Marquardt, update:
$$\Delta\mathbf x_{k+1}=\Delta\mathbf x_k-
\left(J_k^T J_k+\lambda_k I\right)^{-1}J_k^T\mathbf e_k,
\qquad \lambda_k\ge 0,$$
where \(\mathbf e_k\) is the residual vector at iteration \(k\), \(J_k\) is the sensitivity matrix at that iteration, and \(\lambda_k\) is a nonnegative damping parameter. The setting vector is then updated by \(\mathbf x_{k+1}=\mathbf x_k+\Delta\mathbf x_{k+1}\). The damping term suppresses the bad conditioning of \(J_k^T J_k\) and controls the step size.
\(\lambda_k\) acts as an indirect trust-region control. A trial step is accepted when it reduces the residual; if it fails, \(\lambda_k\) is increased and the step is shortened. The implementation also screens each machine-setting change against a practical range so that the optimized result remains usable in production. This screening matters because a mathematically valid correction can still require a machine axis to move beyond its useful adjustment window.
The acceptance rule tracks both the accumulated error and the maximum error. A step can lower the sum of squared deviations while increasing the peak deviation at a few points, and for flank grinding the peak error is often the more important constraint. The algorithm therefore prefers a step that reduces the maximum error, even when the aggregate index does not fall as quickly.
Case study: 31/38 pair at a 52° shaft angle
The example is a 31/38-tooth non-orthogonal pair with a 52° shaft angle, 32 mm face width, 208.81 mm outer cone distance and 20° pressure angle. The pinion is right-hand and the wheel is left-hand. The working flanks are the pinion concave flank and the wheel convex flank, so both were corrected. Table 1 lists the blank parameters.
| Parameter | Pinion | Wheel |
|---|---|---|
| Number of teeth | 31 | 38 |
| Face width (mm) | 32.00 | 32.00 |
| Pressure angle (°) | 20 | 20 |
| Shaft angle (°) | 52 | 52 |
| Addendum (mm) | 5.90 | 3.44 |
| Dedendum (mm) | 4.65 | 6.30 |
| Outer cone distance (mm) | 208.81 | 208.81 |
| Pitch cone angle (°) | 23.16667 | 28.833333 |
| Face cone angle (°) | 24.26667 | 29.600000 |
| Root cone angle (°) | 22.4 | 27.73333 |
| Hand | RH | LH |
Table 2 gives the initial cutter and machine settings and the values obtained after four reverse-correction cycles. The pinion uses a nonzero helical correction, \(H_{eli}=6.68482\); the wheel uses zero. The corrections are small for some settings and larger for others, which is consistent with the sensitivity pattern of a ten-parameter machine model.
| Setting | Symbol | Pinion initial | Pinion corrected | Wheel initial | Wheel corrected |
|---|---|---|---|---|---|
| Cutter diameter (mm) | \(D\) | 305.0400 | — | 304.80 | — |
| Point width (mm) | \(P_w\) | 2.1819 | — | 1.9050 | — |
| Radial cutter position (mm) | \(S_r\) | 163.3098 | 161.3222 | 163.52611 | 161.653 |
| Angular cutter position (°) | \(q\) | 50.4968 | 50.50201 | 49.8142 | 49.8051 |
| Cutter tilt (°) | \(\sigma\) | 3.44 | 3.440614 | 0.5985 | 0.6173 |
| Cutter swivel (°) | \(\zeta\) | 2.5969 | 2.596959 | 14.88 | 14.88 |
| Vertical offset (mm) | \(E_M\) | 0.12558 | 0.05488 | 0.0000 | 0.0512 |
| Installation angle (°) | \(\gamma_m\) | 20.13 | 20.1016 | 23.73 | 23.754 |
| Horizontal offset (mm) | \(X_B\) | 0.43873 | 0.438954 | 0.00000 | −1.3304 |
| Bed position (mm) | \(X_D\) | −3.1264 | −3.1439 | −3.86914 | −7.07774 |
| Roll ratio | \(m\) | 2.5258999 | 2.512648 | 2.075437 | 2.080307 |
| Helical correction | \(H_{eli}\) | 6.68482 | 6.684794 | 0.00000 | 0.00000 |
After four correction cycles, the accumulated error and the maximum error both fell on the two working flanks. The accumulated error dropped to 46.765% and 14.8% of the initial values, and the maximum error dropped to 3.77% and 10.62%, for the pinion concave and wheel convex flanks respectively. The final maximum residual was below 15 µm on both flanks.
| Flank | Correction cycles | Accumulated error after correction (% of initial) | Maximum error after correction (% of initial) | Final maximum residual |
|---|---|---|---|---|
| Pinion concave | 4 | 46.765 | 3.77 | <15 µm |
| Wheel convex | 4 | 14.8 | 10.62 | <15 µm |
The residual surfaces are smooth rather than randomly scattered, and points along the same grid line carry similar residuals. This banded pattern is consistent with a systematic machine-setting effect. The pinion concave flank shows larger residuals at the four corners of the grid, while the wheel convex flank is biased toward negative residuals. Those features indicate where the ten-setting model has less authority; they are geometric observations, not predictions of contact performance.
Interpretation, scope and limits
The method converts a trial-and-error setting search into a measurement-based correction loop: generate the design flank from the kinematic model, measure the actual flank, form the normal-deviation sensitivity matrix, solve a damped least-squares update, and apply the screened setting changes. In the 31/38 case, this loop brought both working flanks below 15 µm maximum residual in four cycles. The different reduction rates for the accumulated and peak errors show why a single scalar objective is insufficient; the acceptance rule needs both indicators.
The word adaptive here refers to successive measurement-based updates of machine settings, not to real-time control of the grinding wheel during a pass. Each cycle still requires a ground and measured part. The practical benefit is a reduction in the number of trial parts and a more predictable direction for each adjustment, not the elimination of measurement.
The scope is geometric. The model minimizes normal deviation between the actual and design flanks; it does not model loaded tooth contact, transmission error, noise or grinding burn. A flank that is geometrically close to the design surface can still have an unacceptable contact pattern under load. The case is also a single 31/38 design with two corrected flanks, not a statistical validation across batches, machine thermal states or cutter wear conditions. Production use would require verification on multiple parts and a check that the recommended settings remain within the machine’s resolution and adjustment range.
Within those limits, the example supports a clear technical conclusion: for non-orthogonal aerospace spiral bevel gears, a ten-setting kinematic model combined with a damped, trust-region-controlled least-squares update can reduce both the aggregate and peak flank deviations in a small number of correction cycles. The remaining corner residuals and the negative bias on the wheel flank define the next modeling problem: extending the setting space or adding local compensation so that the edge regions converge as well as the central flank.