The coupling that limits flank correction
Spiral bevel gears carry the final torque stage in aircraft engine and helicopter drive trains, so their tooth-surface accuracy propagates directly into mesh quality, contact-pattern stability and transmission noise. The completing method has largely displaced the classical five-cut route for volume production: instead of roughing and finishing the pinion with separate machine settings, a single pass with one cutter head generates both flanks, cutting cycle time substantially. The trade-off appears at the correction stage. Because the working (drive) flank and the non-working (coast) flank are formed from the same set of machine settings, any adjustment introduced to reduce error on one flank simultaneously displaces the other. A conventional cradle-type machine offers nine setting parameters plus the roll ratio — too few independent variables to shape two flanks that respond in a coupled way.
Free-form CNC gear cutting machines remove that structural limitation. Their axes need not follow the linear or sinusoidal laws imposed by a mechanical cradle; each axis position can be prescribed as a polynomial in the cradle angle, so an equivalent kinematic transformation turns the traditional setting list into a much larger coefficient set. With five interpolating axes and seven coefficients per axis, the correction problem acquires 35 independent variables — enough, in principle, to attack both flanks at once, provided the mapping from coefficient to flank deviation is known and the two-flank objective is posed correctly.
Equivalent conversion from cradle settings to Cartesian axes
The starting point is the traditional cradle-type model. Four parameters fix the cutter head in the machine frame: cutter tilt it, cutter swivel js, radial setting Sr and angular setting q. Five more fix the workpiece: vertical offset Em, sliding base XB combined with the helical motion coefficient H, blank mounting angle γm, and horizontal offset Xg. During generation the cradle rotates through angle φc while the blank rotates through φ = Rapφc, with Rap the roll ratio. The resulting tooth surface in the workpiece frame follows from the chain of homogeneous transformations:
$$\mathbf{r}_w(u,\theta,\varphi_c)=\mathbf{M}_{we}(\phi)\,\mathbf{M}_{et}\left(i_t,j_s,S_r,q,E_m,X_B,X_g,\gamma_m,R_{ap}\right)\mathbf{r}_t(u,\theta)$$
where u and θ are the surface parameters of the cutter blade and rt is the blade surface in the cutter frame.
The Cartesian machine model in turn contains three translational axes CX, CY, CZ and three rotary axes ψa, ψb, ψc. ψb is the blank mounting angle and ψc the blank rotation; CZ controls tooth depth, while CX and CY locate the cutter centre. The cutter spindle rotation ψa is excluded because it does not interpolate with the other axes during milling. With the centre distance Cd taken as zero, the surface is
$$\mathbf{r}_w=\mathbf{M}_{wd}(\psi_c)\,\mathbf{M}_{da}\left(C_X,C_Y,C_Z,C_d,\psi_b,\Delta\psi_c\right)\mathbf{r}_t(u,\theta)$$
where Δψc is the additional rotation superposed on the nominal roll, i.e. the deviation of the blank from the linear law ψc = Rapφc. The two models must describe the same relative position and motion of cutter and blank, which is enforced by requiring the transformation matrices to be identical:
$$\mathbf{M}_{da}=\mathbf{M}_{et}$$
Equating corresponding entries gives closed-form expressions for the five axis coordinates in terms of the cradle angle and the original settings — for example ψb = asin(a13) and Δψc = atan(a12/a11), with CX, CY and CZ obtained from the translation column and the a24, a34 terms. Each axis is then expanded in a Taylor series about φc = 0 and truncated at sixth order:
$$q_k(\varphi_c)=\sum_{j=0}^{6}a_{k,j}\,\varphi_c^{\,j},\qquad k\in\{X,Y,Z,B,C\}$$
A representative pinion was defined by the settings in Table 1; the resulting axis polynomials are given in Table 2 in compact form.
| Parameter | Symbol | Value |
|---|---|---|
| Cutter tilt angle | it | 2.24° |
| Cutter swivel angle | js | 136.97° |
| Radial cutter position | Sr | 106.7915 mm |
| Angular cutter position | q | 60.60° |
| Vertical offset | Em | −0.3235 mm |
| Sliding base | XB | 2.3564 mm |
| Helical motion coefficient | H | 2.5913 mm·rad−1 |
| Blank mounting angle | γm | 27.07° |
| Horizontal offset | Xg | −0.0123 mm |
| Roll ratio | Rap | 2.181756 |
| Outer blade radius | r11 | 115.45 mm |
| Outer blade pressure angle | α11 | 18.00° |
| Inner blade radius | r12 | 112.41 mm |
| Inner blade pressure angle | α12 | 22.00° |
For the CZ axis, for instance, the expansion about the reference cradle position reads 6.760839 + 2.674010φc − 0.001595φc2 − 0.001658φc3 + 0.000132φc4 + 0.000082φc5 − 0.0000004φc6, in the length units of the axis command. The higher-order terms are numerically small, but they are precisely the terms that later carry the correction.
Surface generation and verification of the conversion
Substituting the axis polynomials into the machine transformation yields a tooth surface expressed entirely through polynomial coefficients. The unit normal follows from the surface derivatives,
$$\mathbf{n}_w=\frac{\dfrac{\partial \mathbf{r}_w}{\partial u}\times\dfrac{\partial \mathbf{r}_w}{\partial \theta}}{\left\|\dfrac{\partial \mathbf{r}_w}{\partial u}\times\dfrac{\partial \mathbf{r}_w}{\partial \theta}\right\|}$$
and the equation of meshing closes the system:
$$f(u,\theta,\varphi_c)=\mathbf{n}_w\cdot\frac{\partial \mathbf{r}_w}{\partial \varphi_c}=0$$
Grid points on the flank are obtained by rotating the surface into a projection plane,
$$X_L=r_{w,1},\qquad R_L=\sqrt{r_{w,2}^{2}+r_{w,3}^{2}}$$
so that a prescribed grid in the axial–radial plane back-solves to the surface parameters (u, θ, φc) and hence to a three-dimensional point. Deviations between two surfaces are evaluated by rotating one surface until the grid midpoints coincide, then solving along the surface normal for the signed distance δ at every grid point.
Comparing surfaces generated directly from the cradle model with those generated from the truncated polynomial axes gave a maximum deviation of 5.6 × 10−4 μm over both flanks — an artefact of the rounding applied to the polynomial coefficients rather than a kinematic discrepancy, and several orders of magnitude below any manufacturing tolerance. The equivalence of the conversion is therefore established to numerical precision.
How polynomial coefficients shape the flank topology
Each coefficient was perturbed by a fixed amount — 0.1 mm on the translational axes, 0.1° on the rotary axes — and the resulting two-flank deviation field was recomputed. For the CZ axis, the peak deviation produced by perturbing one coefficient at a time falls steeply with order (Table 2).
| Perturbed coefficient order (CZ) | Peak two-flank deviation (μm) |
|---|---|
| 0 (constant) | 7.0295 |
| 1 | 13.1374 |
| 2 | 5.8477 |
| 3 | 2.0718 |
| 4 | 0.69369 |
| 5 | 0.22767 |
| 6 | 0.07518 |
The zeroth-order term acts essentially as a shift of the blank along the axis and produces a spiral-angle error across the whole flank. The first-order term is the most influential: besides a spiral-angle component, it changes the pressure angle differently from the mid-face towards the toe and the heel, so the contact pattern rotates into a diagonal — running from toe root to heel tip on the concave flank and from toe tip to heel root on the convex flank. Second- and third-order terms bend the flank without a strong rigid-body component, while terms of fourth order and above produce only local, low-amplitude warping. The same ordering behaviour holds for the other axes, though the character differs: the X, Y and Z translations act mainly on spiral and pressure angle (and hence on the diagonal), whereas the B and C rotations act mainly on local twist of the flank.
This hierarchy is the practical basis for correction. Low-order coefficients supply the large, smooth adjustments; high-order coefficients provide fine shaping with limited collateral effect on the opposite flank, and their diminishing sensitivity means a correction algorithm can use them without destabilizing the solution.
Sensitivity matrix and weighted two-flank correction
Writing the coefficient vector as aij, with i over the five axes and j over the seven orders (n = 35 coefficients in total), the residual between the target surface r1 and the perturbed surface r2 is
$$\Delta \mathbf{r}=\mathbf{r}_1(a_{ij})-\mathbf{r}_2(a_{ij}+\Delta a_{ij})$$
Linearizing and projecting onto the target unit normal gives, for each grid point, a linear relation between the normal error and the coefficient adjustments, with the sensitivity of coefficient n at grid point k approximated by a finite difference ηnk ≈ Δδ/Δan. Assembling k grid points over the working and non-working flanks yields the system
$$\{\Delta\delta_p\}=[\eta_{pq}]\{\Delta a_q\},\qquad p=1,2,\dots,k,\quad q=1,2,\dots,n$$
Because the two flanks share one set of axis polynomials, a correction that reduces working-flank error may either reduce or increase the error on the non-working flank. To keep the two objectives controllable, a weight λ is applied to the working-flank errors:
$$\{\Delta\delta_p\}=\lambda\{\Delta\delta_1\}+\{\Delta\delta_2\}$$
with {Δδ1} and {Δδ2} the working- and non-working-flank error vectors. Since the number of coefficients is smaller than the number of grid points, the system is overdetermined and is solved in the least-squares sense:
$$\{\Delta a_q\}=\left([\eta_{pq}]^{\mathrm{T}}[\eta_{pq}]\right)^{-1}[\eta_{pq}]^{\mathrm{T}}\{\Delta\delta_p\}$$
Cutting and measurement trials
The corrected motion data were post-processed and used to cut the pinion on a Gleason 600HC CNC hypoid gear cutting machine; the tooth surfaces were then measured on a Gleason 650 gear measuring centre. In the uncorrected state, the largest deviation, 75.7 μm, appeared at the toe root of the non-working flank, and both flanks carried substantial spiral-angle and pressure-angle errors, i.e. the error pattern was dominated by exactly the low-order terms identified in the sensitivity study.
Before committing to hardware, the correction model was evaluated for three working-flank weights (Table 3). All three settings bring the predicted peak deviation below 3 μm on both flanks, with the peaks shifting modestly as the weight changes — the maxima reported are grid-point extremes, so the peak metric need not vary monotonically with λ even though the error distribution does.
| Working-flank weight λ | Predicted peak deviation, working flank (μm) | Predicted peak deviation, non-working flank (μm) |
|---|---|---|
| 0.8 | 2.1 | 1.7 |
| 1.0 | 2.4 | 1.8 |
| 1.2 | 2.7 | 2.0 |
λ = 1 was adopted as the balanced choice, and the resulting coefficient adjustments were applied to all five axis polynomials. The changes are not confined to the constant terms; for CZ, the constant moved from 6.760839 to 6.695907 and the quadratic coefficient from −0.001595 to 0.013749, showing that the least-squares step redistributes magnitude into mid-order terms rather than simply offsetting the surface.
A second cutting trial with the corrected polynomials was measured under the same conditions. The peak deviation fell to 6.7 μm on the working flank and 5.5 μm on the non-working flank, against the 75.7 μm initially present on the non-working flank (Table 4). The measured values are roughly 2.5–3 times the model prediction, a gap consistent with machine-tool geometric error, workpiece mounting error and cutter wear that the kinematic model does not represent — but the flank quality meets the accuracy requirement for the application.
| Stage | Peak deviation, working flank (μm) | Peak deviation, non-working flank (μm) |
|---|---|---|
| Before correction (measured) | large spiral- and pressure-angle errors | 75.7 (toe root) |
| After correction, predicted (λ = 1) | 2.4 | 1.8 |
| After correction (measured) | 6.7 | 5.5 |
What the approach delivers and where it stops
Two elements make the method work. First, the sixth-order polynomial representation of the CNC axes preserves the kinematic equivalence with the cradle model to within coefficient-rounding error, while expanding the correction space from nine settings to 35 coefficients. Second, the weighted sensitivity formulation converts the coupled two-flank problem into a single overdetermined least-squares problem in which the relative priority of the working and non-working flanks is an explicit, tunable parameter rather than an accident of the setting list. The weight does not create accuracy — it allocates it.
The limits follow from the same sensitivity hierarchy. Coefficients of order four to six move the surface by well under 1 μm per 0.1 mm or 0.1° of perturbation, so their usefulness is confined to local shaping; the bulk of any correction must come from the first three orders, which are also the terms with the strongest cross-flank coupling. Residual error after correction is set by machine and installation accuracy rather than by the polynomial order, which is why the measured peaks exceeded the predicted ones. Within those bounds the correction procedure is not specific to the completing method: any spiral bevel gear process whose machine axes can be described by coefficient-bearing motion laws — five-cut cutting, grinding or other generating routes — can use the same coefficient-to-error sensitivity formulation to reduce two-flank deviations.