Why Hardening Distortion Dominates Spiral Bevel Gear Accuracy
Spiral bevel and hypoid gearsets carry the final torque multiplication in automotive axles, industrial reducers, mining drives and rotorcraft transmissions. Their flanks are local-conjugate surfaces: contact is deliberately restricted to a small elliptic patch that migrates across the tooth under load, a design that tolerates small mounting errors and avoids edge loading. That patch is defined by a few tens of micrometres of flank geometry, generated by a rotating cutter head whose motion relative to the workpiece is prescribed by a small set of machine settings — roll ratio, radial and angular cutter position, vertical and axial workpiece offsets, sliding base and workpiece mounting angle.
Carburizing and quenching are mandatory for the load capacity of these parts, but the same cycle redistorts the flank. Thermal gradients and phase-transformation stresses, together with fixturing, quench-medium cooling rate and hardenability, displace the hardened surface from the cutting-induced geometry. Deviations of only 20–60 µm are enough to shift the contact patch toward the toe or the root, raise transmission error and concentrate stress at the flank ends. Because hard finishing of spiral bevel gear flanks is expensive, the usual industrial answer is lapping or grinding correction after hardening — a per-part remedy that consumes capacity rather than removing the cause.
A preventive route is more attractive: if the systematic part of the distortion is known before hardening, it can be pre-compensated in the cutting settings, so that the flank distorts into the design geometry. The obstacle is statistical. Destructive and slow inspection means that a production lot is characterized by a handful of measured parts — typically fewer than 30 — with no defensible distributional assumption. Classical estimation is weak exactly where it is needed. Bootstrap resampling, introduced by Efron in 1979, offers an alternative: repeated resampling with replacement from the observed batch generates an empirical distribution of a statistic without additional measurement, which is why it has been adopted for small-sample uncertainty quantification in medicine, bearing diagnostics and structural analysis.
The workflow described here combines four elements: an analytical model of the generated flank, a grid-based deviation measurement, a bootstrap model of the batch deviation statistics, and a NURBS-based reconstruction that converts the batch mean deviation surface into corrected machine settings. It was exercised on three gearsets, the largest being a 10 × 37 equal-depth hypoid pair from a Shandong gear and axle plant producing roughly 700,000 automotive spiral bevel gear sets per year on Gleason Phoenix and Klingelnberg machines.
Generating the Theoretical Flank
The cutting process is modelled as the envelope of a straight-sided blade mounted on a cutter head that rotates about its own axis while the cradle carries it around the generating-gear axis. In the cutter coordinate system $S_d(X_d,Y_d,Z_d)$, the active cutting cone and its unit normal are
$$\mathbf{r}_d(u_d,\theta_d)=\begin{pmatrix}\left(r_d \pm u_d\sin\alpha_d\right)\cos\theta_d\\ \left(r_d \pm u_d\sin\alpha_d\right)\sin\theta_d\\ -u_d\cos\alpha_d\\ 1\end{pmatrix},\qquad
\mathbf{n}_d=\begin{pmatrix}-\cos\alpha_d\cos\theta_d\\ -\cos\alpha_d\sin\theta_d\\ -\sin\alpha_d\end{pmatrix}$$
where $r_d$ is the blade tip radius, $\alpha_d$ the blade profile angle (positive for the outside blade cutting the convex flank, negative for the inside blade), and $u_d$, $\theta_d$ the surface parameters. Five coordinate systems are chained: cutter $S_d$, cradle $S_y$ rotating by $\varphi_y$, machine $S_j$, auxiliary $S_f$ and workpiece $S_g$, giving
$$\mathbf{r}_g=M_{gf}M_{fj}M_{jy}M_{yd}\,\mathbf{r}_d,\qquad
\mathbf{n}_g=L_{gf}L_{fj}L_{jy}L_{yd}\,\mathbf{n}_d$$
with $L$ the rotational 3 × 3 part of each 4 × 4 matrix. For example, the cradle-to-machine transform is a pure rotation about the cradle axis:
$$M_{jy}=\begin{pmatrix}\cos\varphi_y & \sin\varphi_y & 0 & 0\\ -\sin\varphi_y & \cos\varphi_y & 0 & 0\\ 0 & 0 & 1 & 0\\ 0 & 0 & 0 & 1\end{pmatrix}$$
The meshing condition that eliminates the cradle rotation from the solution is the standard tangency of relative velocity to the flank:
$$\mathbf{n}_d\cdot\mathbf{v}_j^{(yg)}=0,\qquad
\mathbf{v}_j^{(yg)}=\boldsymbol{\omega}_j^{(y)}\times\mathbf{r}_j-\boldsymbol{\omega}_j^{(g)}\times\left(\mathbf{r}_j-\mathbf{R}_g\right)$$
Because the flank is a free surface in $(u_d,\theta_d)$, a bounded region must be imposed before discretization. The tooth surface is rotated onto the axial plane through $X=x_d$ and $Y=\sqrt{y_d^{2}+z_d^{2}}$, the four corner points of the working flank are located from the gear blank geometry (face width, addendum, dedendum, pitch cone angle, outer cone distance), and the boundary polygon is shrunk inward: 10 % of the face width at toe and heel, and an inset of 5 % of the working depth — at least 0.6 mm — at tip and root, so that the measurement grid avoids the fillet and edge chamfers. Parametrizing the shrunk quadrilateral by bilinear interpolation of its four edges yields a 15 × 15 grid of 225 nodes; each node is then converted back to $(u_d,\theta_d)$ by Newton iteration, starting from the mid-face reference point whose axial coordinates follow from the pitch angle, mid-face cone distance, addendum and clearance.
| Item | Case A | Case B | Case C (production trial) |
|---|---|---|---|
| Module (mm) | 4.741 | 3.8929 | 12.78 |
| Teeth, pinion / gear | 10 / 41 | 8 / 39 | 10 / 37 |
| Hypoid offset (mm) | 31.8 | 31.75 | 30 |
| Mean pressure angle (°) | 20 | — | 22.5 |
| Mid-face spiral angle, pinion / gear (°) | 49.98 / 29 | 50 / 29.57 | 42.5 / 34.47 |
| Face width, pinion / gear (mm) | 33.53 / 28 | 38.8 / 32.18 | 79.7 / 75.0 |
| Cutter radius or diameter (mm) | 95.25 / 92.456 / 97.917 | — | 400 (diameter) |
Measuring the Deviation Field
Flank deviation is defined along the surface normal of the design flank, referred to the flank centre, and sampled on the same 15 × 15 grid used for the analytical model. Two acquisition strategies are available on a CNC gear measuring centre. Scanning moves the stylus continuously along a flank line and is fast, but stylus friction and the risk of interference in the fillet bias the recorded values. Point-by-point probing drives the stylus to each programmed node and records the geometric error directly; it is slower, yet it delivers the discrete, well-registered set of normal deviations that a machine-setting inversion requires, and it does not smear local features. For setting-correction purposes the point-by-point mode is therefore the appropriate choice.
All measurements reported below were taken on a Klingelnberg P65 measuring centre, with 20 gear pairs sampled before and after hardening for the production lot. Before heat treatment, the Case C gear flank showed a maximum convex deviation of 8.9 µm at the heel near the addendum and a maximum concave deviation of 19.4 µm at the toe addendum; the pinion reached −12.7 µm on the convex flank at the heel addendum and 21.5 µm on the concave flank at the heel dedendum. These values lie inside the design and shop-floor tolerance bands — the cutting process, in other words, was under control, and the distortion observed afterwards is attributable to the heat-treatment cycle rather than to the milling step.
Bootstrap Statistics for Small Batches
Let the measured deviations at one grid node over a lot form the sample $F=(F_1,F_2,\dots,F_m)$. A bootstrap sample of size $N$ is drawn with replacement from $F$, with every element equally likely and $F$ itself unchanged; the draw is repeated $B$ times, producing $B$ resamples
$$\mathbf{P}_{\text{Bootstrap}}=\left(P_1,P_2,\dots,P_b,\dots,P_B\right),\qquad
\bar{P}_b=\frac{1}{N}\sum_{i=1}^{N}P_b(i)$$
from which the mean, standard deviation and empirical distribution of the statistic follow directly:
$$\bar{P}=\frac{1}{B}\sum_{b=1}^{B}\bar{P}_b,\qquad
\sigma^2=\frac{1}{B-1}\sum_{b=1}^{B}\left(\bar{P}_b-\bar{P}\right)^2$$
Percentile confidence intervals are obtained by sorting the $B$ estimates and truncating at the order statistics $K_1=[B\alpha/2]$ and $K_2=[B(1-\alpha/2)]$, so that $P\{\hat{\theta}_{K_1}<\theta<\hat{\theta}_{K_2}\}=1-\alpha$. With $\alpha=0.05$ the interval carries 95 % confidence, which is the usual engineering convention.
The number of resamples and the size of the original sample were fixed empirically. With five-element samples and only 100 resamples, the relative error of the estimated variance reached 12.97 %; at 2000 resamples the error was minimal, so $B=2000$ was adopted. Increasing the original sample from five to ten elements reduced the variance error by 6.7 percentage points, whereas going from ten to fifteen reduced it by only 0.15 points — ten measured elements already capture the population characteristics, and larger samples buy little for the measurement effort they cost.
| Sample size | Gear convex flank | Pinion concave flank | ||
|---|---|---|---|---|
| Mean error | Variance error | Mean error | Variance error | |
| 5 | 0.45 % | 10.54 % | 0.23 % | 10.46 % |
| 10 | 0.26 % | 5.04 % | 0.17 % | 5.11 % |
| 15 | 0.18 % | 3.30 % | 0.13 % | 3.38 % |
| 20 | 0.12 % | 2.52 % | 0.02 % | 2.55 % |
The mean is recovered far more tightly than the variance — below 0.5 % even with five elements — and both errors decay monotonically, which is the behaviour expected when the resampling distribution converges on the empirical one. On the production lot, the 95 % interval for the mean of a five-element gear concave sample was (−1.5479, −1.5192) around a point estimate of −1.5335, and the corresponding standard-deviation interval was (3.5178, 3.5374) around 3.6542; widening the sample to twenty elements narrowed the mean interval to (−0.9133, −0.9031). The interval width is thus a usable measure of how much confidence a given inspection budget supports.
Reconstructing Measured Flanks with NURBS
The 225 probed deviations are discrete; compensation needs a continuous surface. A cubic NURBS representation is used,
$$\mathbf{P}(u)=\frac{\sum_{i=0}^{n}W_i\,\mathbf{d}_i\,N_{i,k}(u)}{\sum_{i=0}^{n}W_i\,N_{i,k}(u)}$$
with $\mathbf{d}_i$ the control points, $W_i$ the weights and $N_{i,k}$ the B-spline basis generated by the de Boor–Cox recursion
$$N_{i,0}(u)=\begin{cases}1, & s_i\le u\le s_{i+1}\\ 0, & \text{otherwise}\end{cases},\qquad
N_{i,k}(u)=\frac{u-s_i}{s_{i+k}-s_i}N_{i,k-1}(u)+\frac{s_{i+k+1}-u}{s_{i+k+1}-s_{i+1}}N_{i+1,k-1}(u)$$
The measured points are interpolated rather than approximated, so the control points are recovered by inverting the linear system relating data points to control points. Because the flank curvature varies strongly from toe to heel, centripetal parameterization is used; it distributes node parameters according to $\sqrt{|F_i-F_{i-1}|}$ and avoids the clustering that uniform or chord-length parameterization produces on tightly curved regions. For the surface, the face-width direction is taken as $u$ and the tooth-height direction as $v$; the node vectors of the section curves are averaged to give the surface node vector, control points are inverted first along $u$ and then, using the resulting control points as data points, along $v$.
For the Case A pinion concave flank the fitted surface gave a squared-deviation sum of 1.0383 mm² against 1.0760 mm² for the measured points, a 0.45 % relative difference; the gear convex flank gave 0.5877 mm² against 0.5852 mm², or 0.42 %. Typical $u$-direction node vectors were $U_1=[0,0,0,0,0.1427,\dots,0.8569,1,1,1,1]$ for the gear flank and $U_2=[0,0,0,0,0.1347,\dots,0.8492,1,1,1,1]$ for the pinion flank; the reconstruction preserves both the global curvature and the local deviation pattern.
Pre-Control Compensation via the Mirrored Mean Deviation Surface
Within a lot sharing material, cutting method and heat-treatment specification, the distortion is repeatable. Let $H_L$ be the design flank, $H_R$ the mean difference surface obtained by averaging the bootstrap-statistical deviations of the batch, and $H_D$ the surface obtained by mirroring $H_R$ about $H_L$ along the design normal $\mathbf{n}_L$. With $\mathbf{r}_L$ and $\mathbf{r}_R$ the position vectors of corresponding points and $\delta$ the normal deviation,
$$\mathbf{r}_R=\mathbf{r}_L+\delta\,\mathbf{n}_L,\qquad
\mathbf{r}_D=\mathbf{r}_L-\delta\,\mathbf{n}_L,\qquad
\mathbf{r}_D=\mathbf{r}_R-2\delta\,\mathbf{n}_L$$
If the flank is machined to $H_D$ before hardening, the distortion $\delta$ carries it onto $H_L$ afterwards: the pre-deformed surface is the pre-image of the design surface under the distortion map. Operationally, the correction is found by searching for machine-setting increments $\Delta\Phi_j$ that minimize the summed squared normal deviation over the 225 grid nodes,
$$f\left(\Delta\Phi_j\right)=\sum_{i=1}^{225}\left[\left(\mathbf{H}_L\left(u_i,\theta_i;\Phi_j\right)-\mathbf{H}_R\left(u_i,v_i\right)\right)\cdot\mathbf{n}_L\left(u_i,\theta_i;\Phi_j\right)\right]^{2}\rightarrow\min$$
with the settings updated iteratively as $\Phi_j^{(k+1)}=\Phi_j^{(k)}+\Delta\Phi_j^{(k)}$ and $\delta_i=\left(\mathbf{H}_R-\mathbf{H}_L\right)\cdot\mathbf{n}_L$ evaluated at each node. Because a finite set of seven settings cannot reproduce an arbitrary deviation field exactly, the corrected flank approaches but never coincides with the target; the residual is small enough to remain inside the flank tolerance band.
An intermediate verification on the Case B pinion (module 3.8929 mm, 8 × 39) showed the mechanism working end to end: after applying the computed corrections and repeating the same hardening cycle, the maximum flank deviation fell by 20.8 µm and the squared-deviation sum dropped from 0.2537 to 0.1104 mm².
Production Trial on a 10 × 37 Hypoid Set
The 469 rear-axle gearset (module 12.78 mm, 10 × 37, 400 mm cutter diameter) was milled on an Oerlikon C50 CNC machine using the settings derived from the design calculation, then carburized and quenched on the plant line. Twenty pairs were measured before hardening and twenty after, and the bootstrap model was applied to the gear concave and pinion convex data with five- and ten-element resamples at $B=2000$. Machine-setting corrections obtained from the mirrored mean difference surface are listed in Table 3; the gear received bed and radial-position corrections, while the pinion — the member that carries cutter tilt and swivel in this hypoid design — received corrections on eight settings.
| Setting | Gear | Pinion |
|---|---|---|
| Cutter tilt (°) | — | −0.0698 |
| Cutter swivel (°) | — | 0.0867 |
| Horizontal workpiece position (mm) | — | −0.2033 |
| Sliding base / bed (mm) | −0.2874 | 0.0349 |
| Workpiece mounting angle (°) | — | −0.0749 |
| Radial cutter position (mm) | −0.1592 | −0.2405 |
| Mid cradle angle (°) | — | −0.0309 |
| Vertical offset (mm) | — | 0.0593 |
Re-cutting with the corrected settings and repeating the identical heat-treatment cycle produced the results in Table 4. The peak deviation change dropped by 39.6 µm on the gear and 55.5 µm on the pinion, and the squared-deviation sums fell by 69.78 % and 76.66 % respectively; the deviation trend across the flank also flattened, indicating that the correction removed a systematic pattern rather than shifting a local extreme.
| Quantity | Gear, before | Gear, after | Pinion, before | Pinion, after |
|---|---|---|---|---|
| Change in peak deviation (µm) | — | −39.6 | — | −55.5 |
| Sum of squared normal deviations (mm²) | 0.4271 | 0.1291 | 0.4611 | 0.1076 |
| Reduction in squared-deviation sum | — | 69.78 % | — | 76.66 % |
The difference-surface comparison also illustrates why ten-element samples were preferred over five. Reconstructed with five elements, the bootstrap mean difference surface departed visibly from the measured mean difference surface at points sampled on the eighth row and fourth column and on the tenth row and eleventh column of the grid; with ten elements the gap narrowed at both locations, while the separation between the measured mean surface and the design surface stayed fixed, as it must. The bootstrap surface therefore converges on the empirical one as sample size grows, which is the property that justifies predicting distortion from a short inspection record.
Limits and Engineering Implications
The method converts a post-hoc, part-by-part correction into a setting-level correction applied once per lot, which suits high-volume production of spiral bevel gears where lapping capacity is the bottleneck. Its validity rests on three conditions that should be stated explicitly. First, the distortion must be systematic: identical material, identical cutting method and an unchanged heat-treatment recipe are prerequisites, and any change in quench agitation, fixturing or lot-to-lot hardenability will invalidate the mean difference surface. Second, the bootstrap estimates the statistics of the sampled population, not of the process; a biased inspection sample — for example, parts taken only from the top of a rack — will produce a confidently wrong correction. Third, the seven machine settings span a limited subspace of possible flank modifications, so the residual deviation after compensation is bounded below by that subspace’s approximation capability, not by the statistics.
Within those limits the approach is economical. A ten-element sample with 2000 resamples delivered mean estimates within 0.26 % and standard-deviation estimates within 5 % of the full-sample values, so the inspection burden is a handful of parts per lot rather than full inspection. The NURBS stage preserved measured flank geometry to within 0.45 %, well below the deviation magnitudes being corrected, and the compensation step reduced the squared-deviation sum by 70–77 % on both members of the production gearset. The remaining open questions are comparative rather than structural: other resampling or small-sample estimators might tighten the variance estimate, and the correction has so far been demonstrated at a single plant and a single heat-treatment line, so wider validation across quench equipment and lot sizes remains the necessary next step before the method can be treated as a standard production practice.