In a spiral bevel gear set, the pinion is the member that sets the meshing quality of the whole drive: its flank topography determines contact pattern, transmission error and noise. The completing method cuts both flanks of the pinion in a single setup with one cutter head and one set of machine settings, which is why it has displaced the older five-cut sequence in high-volume production. The same property creates its characteristic difficulty. Any parameter changed to reduce error on the working flank also moves the non-working flank, so correction is a coupled, two-objective problem rather than a single-flank fitting exercise.
Free-form CNC gear cutting machines provide a way out of this coupling. Their axes are numerically controlled and need not follow the rigid geometric relations of a mechanical cradle, so each axis can be driven by a polynomial in the cradle angle, and every coefficient of that polynomial becomes an independent correction parameter. The engineering question is then how many orders are worth using, which coefficients control which feature of the flank, and how to distribute the correction between the two flanks without sacrificing one to improve the other.
From cradle settings to six-axis motion
The conventional cradle-type model is defined by nine settings. Four locate the cutter head in the machine plane: cutter tilt it, cutter swivel js, radial setting Sr and angular setting q. Five locate the workpiece: vertical offset Em, sliding base XB (coupled to cradle rotation through the helical motion coefficient H), blank mounting angle γm, horizontal offset Xg, and the roll ratio Rap, which fixes the workpiece rotation as φ = Rapφc. In the cradle frame the generating motion is a rotation of the cradle about its axis; the flank is the envelope of the cutter positions swept during that rotation.
The equivalent Cartesian machine has three linear axes CX, CY, CZ and rotary axes ψa (cutter spindle), ψb (workpiece tilt) and ψc (workpiece rotation). Because the cutter spindle does not interpolate with the other axes during milling, ψa is held constant and only five axes participate in generation. The workpiece surface follows from the chain of transformations
$$\mathbf{r}_w=M_{wd}(\psi_c)\,M_{da}(C_X,C_Y,C_Z,C_d,\psi_b,\Delta\psi_c)\,\mathbf{r}_t$$
where Cd is the distance from the pitch cone apex to the machine origin (taken as zero here) and Δψc is the additional rotation between cutter and workpiece that is superimposed on the nominal roll ψc = Rapφc.
Since the relative position and orientation of cutter and workpiece must be identical in both models, the two transformation matrices are equated, Mda = Met. Matching corresponding entries gives the axis positions in closed form:
$$C_X=(a_{14}+C_d)\cos\psi_b-(a_{34}\cos\Delta\psi_c+a_{24}\sin\Delta\psi_c)\sin\psi_b$$
$$C_Y=a_{24}\cos\Delta\psi_c-a_{34}\sin\Delta\psi_c$$
$$C_Z=-(a_{14}+C_d)\sin\psi_b-(a_{34}\cos\Delta\psi_c+a_{24}\sin\Delta\psi_c)\cos\psi_b$$
$$\psi_b=\arcsin(a_{13}),\qquad \Delta\psi_c=\arctan\!\left(\frac{a_{12}}{a_{11}}\right)$$
Expanding these expressions as a Taylor series about φc = 0 up to sixth order converts the nine cradle settings into five polynomials with up to seven coefficients each. The expansion is exact at the reference cradle position; away from it, the higher-order terms absorb the difference between the rigid mechanical coupling and a freely programmable path. This is precisely what makes the formulation useful for correction: the low-order coefficients carry the global motion of the cradle, while the high-order coefficients describe departures from it that a mechanical machine cannot produce.
For a representative hypoid pinion (outer blade radius 115.45 mm, inner blade radius 112.41 mm, blade pressure angles 18.00° and 22.00°), the calculated machine settings are listed in Table 1 and the resulting axis polynomials in Table 2.
| Setting | Symbol | Value |
|---|---|---|
| Cutter tilt | it | 2.24° |
| Cutter swivel | js | 136.97° |
| Radial setting | Sr | 106.7915 mm |
| Angular setting | 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° |
| Axis | Expression in φc |
|---|---|
| CX | 52.996218 − 91.270203φc − 25.947066φc2 + 14.341102φc3 + 2.104369φc4 − 0.565574φc5 − 0.057888φc6 |
| CY | −92.973799 − 52.132105φc + 44.750647φc2 + 8.587348φc3 − 3.313978φc4 − 0.4006628φc5 + 0.067463φc6 |
| CZ | 6.760839 + 2.674010φc − 0.001595φc2 − 0.001658φc3 + 0.000132φc4 + 0.000082φc5 − 0.0000004φc6 |
| ψb | 0.439330 + 0.016726φc + 0.016324φc2 − 0.002658φc3 − 0.001310φc4 + 0.000110φc5 + 0.000037φc6 |
| ψc | 0.018785 + 0.036673φc − 0.008946φc2 − 0.005887φc3 + 0.000611φc4 + 0.000251φc5 − 0.000006φc6 |
Surface definition and the deviation metric
Substituting the polynomial motion into the transformation chain yields the pinion flank as a function of the cutter surface parameters u, θ and the generating angle φc. The unit normal and the meshing condition are
$$\mathbf{n}_w=\frac{\dfrac{\partial \mathbf{r}_w}{\partial u}\times\dfrac{\partial \mathbf{r}_w}{\partial \theta}}{\left|\dfrac{\partial \mathbf{r}_w}{\partial u}\times\dfrac{\partial \mathbf{r}_w}{\partial \theta}\right|},\qquad f(u,\theta,\varphi_c)=\mathbf{n}_w\cdot\frac{\partial \mathbf{r}_w}{\partial \varphi_c}=0$$
A grid on the flank is generated in the axial plane through the rotational projection XL = rw,1, RL = (rw,22 + rw,32)1/2, so that measured and calculated points can be compared on a common reference. To compare two flanks, one is rotated until the grid midpoints coincide, after which the normal separation δ at any corresponding point pair is obtained from
$$x_w+\delta\,n_{xw}=x’_w,\qquad y_w+\delta\,n_{yw}=y’_w,\qquad z_w+\delta\,n_{zw}=z’_w$$
The transformation itself was checked by computing the same flank directly from the cradle model and from the polynomial CNC model. The largest difference over the grid was 0.00056 μm, i.e. below one nanometre, attributable to the decimal truncation of the polynomial coefficients. For gear metrology this is several orders of magnitude below any realistic measurement uncertainty, so the equivalent conversion can be treated as exact.
Which coefficients move which part of the flank
Each polynomial coefficient was perturbed by an equal amount — 0.1 mm for the linear axes, 0.1° for the rotary axes — and the resulting flank deviation recomputed. Taking the CZ axis as the illustrative case, the peak deviation produced by each order is given in Table 3.
| Order | 0 | 1 | 2 | 3 | 4 | 5 | 6 |
|---|---|---|---|---|---|---|---|
| Max. deviation (μm) | 7.0295 | 13.1374 | 5.8477 | 2.0718 | 0.69369 | 0.22767 | 0.07518 |
Apart from the constant term, influence falls monotonically with order, and above the third order a single coefficient moves the flank by less than a micrometre. The mechanisms differ by axis and by order rather than being a uniform scaling. The constant term of the CZ polynomial shifts the spiral angle over the whole flank; the first-order term changes the spiral angle as well, but applies a different pressure-angle correction at the toe and at the heel, producing diagonal contact — on the concave flank a band running from toe root to heel top, on the convex flank the mirrored direction from toe top to heel root. Across all five axes, the CX, CY and CZ coefficients act mainly on spiral and pressure angle and therefore on the orientation of the contact diagonal, whereas the ψb and ψc coefficients produce local warping of the flank.
Two consequences follow for correction practice. First, the parameter set is naturally hierarchical: low-order coefficients are the right tool for global spiral-angle and pressure-angle errors, high-order coefficients for local shape deviations, and the two groups barely interfere. Second, high-order coefficients provide little correction authority per unit change, so using them aggressively buys small error reductions at the cost of large, poorly conditioned axis motion.
Weighted two-flank correction
Let aij denote the coefficient of order j on axis i, with n coefficients in total. Linearising the difference between the target flank r1 and the flank produced by the perturbed coefficients r2 and projecting onto the target normal gives the sensitivity relation
$$\{\Delta\delta_p\}=[\eta_{pq}]\{\Delta a_q\},\qquad p=1,\dots,k,\quad q=1,\dots,n$$
where Δδp is the normal error at grid point p, Δaq the coefficient correction, and ηpq ≈ Δδ/Δaq the sensitivity of that grid point to that coefficient, evaluated by finite difference. Because both flanks of the pinion are generated by the same set of axis polynomials, a single correction vector acts on working and non-working flank simultaneously. A weight λ is therefore applied to the working-flank error,
$$\{\Delta\delta_p\}=\lambda\{\Delta\delta_1\}+\{\Delta\delta_2\}$$
with {Δδ1} the working-flank and {Δδ2} the non-working-flank error. 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:
$$\{\Delta a_q\}=\left([\eta_{pq}]^{\mathsf T}[\eta_{pq}]\right)^{-1}[\eta_{pq}]^{\mathsf T}\{\Delta\delta_p\}$$
The weight is thus an explicit design choice: λ > 1 buys accuracy on the working flank at the expense of the coast flank, λ < 1 does the reverse, and λ = 1 treats them equally.
Cutting and measuring a pinion
The polynomial motion of Table 2 was post-processed and used to cut a pinion on a Gleason 600HC CNC hypoid gear cutting machine; the part was then measured on a Gleason 650 gear measuring centre. Before correction, the largest error over both flanks reached 75.7 μm, located at the toe root of the non-working flank, and both flanks carried substantial spiral-angle and pressure-angle errors — a state that would measurably degrade the mesh. Applying the correction model with working-flank weights of 0.8, 1.0 and 1.2 gave the predicted residual errors in Table 4.
| Condition | Working flank (μm) | Non-working flank (μm) |
|---|---|---|
| Before correction, measured | — | 75.7 |
| Predicted, λ = 0.8 | 2.1 | 1.7 |
| Predicted, λ = 1.0 | 2.4 | 1.8 |
| Predicted, λ = 1.2 | 2.7 | 2.0 |
| Measured after correction, λ = 1.0 | 6.7 | 5.5 |
Increasing λ from 0.8 to 1.2 raised the residual working-flank error from 2.1 μm to 2.7 μm and the non-working-flank error from 1.7 μm to 2.0 μm, all within the flank tolerance requirement; the trade-off between the two sides is smooth and can be tuned to the application. The correction corresponding to λ = 1 was adopted, giving the revised axis polynomials in Table 5.
| Axis | Expression in φc |
|---|---|
| CX | 53.0373196 − 92.0004226φc − 26.1775563φc2 + 14.4538129φc3 + 2.1288772φc4 − 0.5693242φc5 − 0.0589989φc6 |
| CY | −92.6708881 − 52.5824837φc + 45.1009008φc2 + 8.6741732φc3 − 3.3386377φc4 − 0.4065351φc5 + 0.0676434φc6 |
| CZ | 6.6959074 + 2.6474309φc + 0.0137491φc2 + 0.027715φc3 − 0.0114576φc4 − 0.0001385φc5 + 0.0000382φc6 |
| ψb | 0.4394788 + 0.0172291φc + 0.0161894φc2 − 0.0027395φc3 − 0.0013015φc4 + 0.000114φc5 + 0.0000373φc6 |
| ψc | 0.0193487 + 0.0363665φc − 0.009218φc2 − 0.0058455φc3 + 0.000631φc4 − 0.0002505φc5 − 0.0000069φc6 |
A second pinion was cut with the corrected motion and re-measured. The peak error dropped to 6.7 μm on the working flank and 5.5 μm on the non-working flank. The measured values are roughly 2.5 to 3 times the predicted 2.4 μm and 1.8 μm, a gap consistent with machine-tool geometric error and workpiece mounting error that the model does not contain: the sensitivity analysis describes how the flank responds to a coefficient change, not how accurately the machine realises the commanded motion. Even so, the residual is well inside the engineering requirement, and the flank quality is acceptable for service.
Practical reading of the result
The value of the polynomial formulation is not that it adds degrees of freedom for their own sake, but that it gives a structured hierarchy of them, mapped onto identifiable flank features. Global spiral-angle and pressure-angle errors are removed with the first two orders; the remaining orders handle shape, with rapidly diminishing leverage. In production this suggests a staged correction: fit the low-order coefficients first and stop when the residual is dominated by measurement and machine noise, rather than spending axis motion on coefficients whose effect is smaller than the machine’s own repeatability.
The weighted formulation addresses the structural constraint of the completing method — one parameter set, two flanks. Because the weight enters the least-squares problem explicitly, the same model covers the whole family of preferences from working-flank priority to symmetric treatment, without changing the underlying computation. The measured gap between predicted and achieved error also marks the natural boundary of the approach: geometric correction of the flank is only one term in the error budget, and further improvement requires separating machine and fixture errors from the generated surface rather than refining the polynomial further.
Nothing in the procedure depends on the cutter being a milling cutter or on the machine being a six-axis hypoid generator. Any process in which the machine settings of a cradle-type model can be mapped onto numerically controlled axes, and in which the resulting surface can be measured and expressed as a normal deviation field, admits the same treatment: machined and ground spiral bevel and hypoid gears, and other generating methods whose kinematics reduce to a small number of coupled axes.