Spiral bevel gears transmit torque between intersecting shafts through curved teeth that engage gradually and carry load over several contacting tooth pairs. This geometry is efficient, but it also makes the contact pattern sensitive to small changes in shaft position. In an aeroengine drive, a few hundredths of a millimetre of axial displacement or a few thousandths of a radian of shaft-angle deviation can move the contact pattern toward the tooth edges, raise local load, and distort the loaded transmission error. Tightening manufacturing and assembly tolerances reduces these effects, but at a cost that rises quickly with precision. An alternative is to design the tooth surface so that its contact pattern and transmission error remain acceptable over the installation-error distribution, not only at one nominal alignment.
Installation error as a four-parameter random vector
For spiral bevel gears mounted in a practical assembly, the relevant misalignment is commonly represented by four components: pinion axial displacement \(H_p\), gear axial displacement \(H_g\), shaft offset \(V\), and shaft-angle deviation \(\Sigma\). The sign conventions follow the mounting practice specified in GB 11365-1989. Axial displacement is positive toward the large end of the tooth and negative toward the small end. Shaft offset is positive when the gear moves upward as viewed from the gear front. Shaft-angle deviation is positive when the shaft angle increases.
The pinion and gear coordinate systems \(S_1(O_1,X_1,Y_1,Z_1)\) and \(S_2(O_2,X_2,Y_2,Z_2)\) are connected to the fixed frame \(S_h(O_h,X_h,Y_h,Z_h)\) through an auxiliary frame \(S_d\). The transformations that carry the pinion to the fixed frame, the auxiliary frame to the fixed frame, and the gear to the auxiliary frame are
$$
M_{h1}=
\begin{bmatrix}
1&0&0&H_p\\
0&\cos\phi_1&\sin\phi_1&0\\
0&-\sin\phi_1&\cos\phi_1&0\\
0&0&0&1
\end{bmatrix},
$$
$$
M_{hd}=
\begin{bmatrix}
\cos(\Gamma+\Sigma)&0&\sin(\Gamma+\Sigma)&H_g\cos(\Gamma+\Sigma)\\
0&0&0&V\\
-\sin(\Gamma+\Sigma)&0&\cos(\Gamma+\Sigma)&-H_g\sin(\Gamma+\Sigma)\\
0&0&0&1
\end{bmatrix},
$$
$$
M_{d2}=
\begin{bmatrix}
1&0&0&0\\
0&-\cos\phi_2&\sin\phi_2&0\\
0&-\sin\phi_2&-\cos\phi_2&0\\
0&0&0&1
\end{bmatrix}.
$$
Here \(\Gamma\) is the nominal shaft angle, and \(\phi_1,\phi_2\) are the rotation angles of the two members. The full gear-to-fixed-frame transformation is \(M_{h2}=M_{hd}M_{d2}\). The error vector \(\mathbf E=[H_p,H_g,V,\Sigma]^{\mathsf T}\) enters the contact equations through these matrices. In a deterministic tooth contact analysis, \(\mathbf E\) is a fixed offset. In a robust design, it is a random vector, and the response statistics are computed over its distribution.
Misalignment sensitivity has been treated quantitatively since Baxter’s 4×2 matrix formulation at Gleason in the early 1960s, which used finite differences to estimate how alignment changes affect second-order meshing characteristics. Litvin, Simon, and Argyris developed tooth contact analysis for misaligned spiral bevel and hypoid gears and showed how bearing contact shifts and transmission error changes. Later sensitivity-matrix and V/H-check analyses connected axial and offset errors to contact-point motion, and active-design methods used those sensitivities to select machine-tool settings. The robust formulation extends that logic from a single sensitivity coefficient to the mean and variance of the response under a distribution of installation errors.
Contact-pattern metrics
The contact pattern is the most direct indicator of meshing quality, and it is also the basis for the constraints used in the optimization. A loaded contact analysis produces a sequence of contact ellipses along the tooth surface. The pattern is reduced to three quantities: area, centroid position, and direction.
If the major-axis endpoints of the contact ellipses are connected in order, the pattern becomes an approximate polygon with vertices \((x_i,y_i)\). Its area is
$$
S=\sum_{i=1}^{n-2}\frac12
\left|
\det
\begin{bmatrix}
x_i&y_i&1\\
x_{i+1}&y_{i+1}&1\\
x_{i+2}&y_{i+2}&1
\end{bmatrix}
\right|.
$$
The centroid is
$$
\bar x=\frac1n\sum_{i=1}^{n}x_i,\qquad
\bar y=\frac1n\sum_{i=1}^{n}y_i,
$$
and the pattern direction is taken from the first to the last contact point:
$$
\theta=\operatorname{atan2}(y_n-y_1,\;x_n-x_1).
$$
For optimization, these quantities are converted into normalized equivalents \(S’\), \(d’\), and \(e’\), where \(e’\) represents the transmission-error metric. Normalization allows area, direction, and transmission error to be compared on a common scale and used as constraints with consistent limits.
Loaded contact analysis with installation error
The tooth surface is first designed by local synthesis. Three second-order contact parameters are prescribed at a reference point: the angle between the contact trace and the root cone, \(\eta\); the first derivative of the transmission-ratio function, \(m’_{21}\); and the ratio of the contact-ellipse semi-major axis to the face width, \(\delta\). These parameters control the orientation of the contact path, the slope of the predesigned transmission-error curve, and the size of the contact ellipse. They are therefore the natural design variables for changing misalignment sensitivity.
Unloaded tooth contact analysis solves for the contact point and the rigid-body rotations that satisfy position and normal-vector equality:
$$
\begin{aligned}
\mathbf r_h^{(1)}(s_p,\theta_p,\phi_{r1},\varphi_1,\mathbf E)
-\mathbf r_h^{(2)}(s_g,\theta_g,\phi_{r2},\varphi_2,\mathbf E)&=0,\\
\mathbf n_h^{(1)}(s_p,\theta_p,\phi_{r1},\varphi_1,\mathbf E)
-\mathbf n_h^{(2)}(s_g,\theta_g,\phi_{r2},\varphi_2,\mathbf E)&=0,\\
f_{1p}(s_p,\theta_p,\phi_{r1},\mathbf E)=0,\qquad
f_{2p}(s_g,\theta_g,\phi_{r2},\mathbf E)&=0.
\end{aligned}
$$
The variables \(s_p,s_g,\theta_p,\theta_g\) are cutter generating-surface parameters, and \(\phi_{r1},\phi_{r2}\) are the generating roll angles. The equations impose tangency of the two surfaces and the generating constraints of the two members.
Unloaded analysis does not include tooth compliance, load sharing, or contact deformation. Loaded tooth contact analysis adds these effects through a compliance matrix and a complementarity condition:
$$
\mathbf F\mathbf p+\mathbf w=\Theta\mathbf r_1+\mathbf d,\qquad
\mathbf p^{\mathsf T}\mathbf r_2=T_1.
$$
Here \(\mathbf F\) is the flexibility matrix, \(\mathbf p\) is the contact-load vector, \(\mathbf w\) is the initial gap vector, \(\Theta\) is the gear rotation caused by elastic deformation, \(\mathbf r_1\) and \(\mathbf r_2\) are the contact-point radius vectors on pinion and gear, \(\mathbf d\) is the deformed gap vector, and \(T_1\) is the pinion torque. The contact condition is \(p_i\ge0\), \(d_i=0\) for a closed contact and \(p_i=0\), \(d_i>0\) for a separated point. The solution minimizes strain energy subject to these constraints.
When installation error is included, the initial gap and radius vectors become functions of \(\mathbf E\). The error-loaded formulation is
$$
\mathbf F\mathbf p+\mathbf w(\mathbf E)=\Theta(\mathbf E)\mathbf r_1(\mathbf E)+\mathbf d(\mathbf E),\qquad
\mathbf p^{\mathsf T}\mathbf r_2(\mathbf E)=T_1.
$$
This error-loaded contact analysis is the response model used inside the robust optimization. It preserves the unloaded contact solution but recomputes the loaded contact under the misaligned geometry.
6σ robust optimization formulation
A deterministic optimization seeks the best nominal response, usually by minimizing a single contact-pattern error or transmission-error measure. That approach can place the optimum close to a constraint boundary, where a small installation error causes a large performance change. The 6σ robust formulation instead treats installation error as a random input and requires the response distribution to remain inside the functional limits.
The design vector is
$$
\mathbf X=[\eta,\;m’_{21},\;\delta,\;H_p,\;H_g,\;V,\;\Sigma].
$$
The objective combines mean performance and response variance:
$$
\min f=
\frac{W_1}{S_1}\left(\mu_{I_E}-I_{E0}\right)^2+
\frac{W_2}{S_2}\sigma_{I_E}^2.
$$
Here \(I_E=x’+y’\) is the normalized equivalent of the contact-pattern center position, \(I_{E0}\) is its target value, \(\mu_{I_E}\) and \(\sigma_{I_E}\) are the mean and standard deviation of that equivalent, \(W_1\) and \(W_2\) are weighting coefficients, and \(S_1\) and \(S_2\) are normalization coefficients. The two terms represent the two requirements of robust design: the pattern should be close to the target on average, and it should not scatter widely when the installation error varies.
The constraints are
$$
[S’,d’,e’]=f_{\mathrm{ELTCA}}(\eta,m’_{21},\delta,H_p,H_g,V,\Sigma),
$$
$$
S’\le0.6,\qquad d’\le0.6,\qquad e’\le1,
$$
$$
x_{\min,i}+4\sigma_{x_i}\le\mu_{x_i}\le x_{\max,i}-4\sigma_{x_i},
\qquad i=1,\ldots,7.
$$
The first constraint evaluates the normalized contact area \(S’\), direction \(d’\), and transmission-error equivalent \(e’\) through the error-loaded contact analysis. The area and direction limits are set at 0.6, and the transmission-error limit at 1. The last set of constraints keeps the mean of each design variable at least four standard deviations from its bound. This is an 8σ-wide tolerance band in total, corresponding to a one-sided reliability of 99.9937% at a bound. In the general \(n\sigma\) form, \(n=6\) gives a two-sided conformance of 99.9999998%; the gear case uses the 4σ margin on each side because the response is evaluated at both variable bounds.
Monte Carlo sampling is used to estimate the mean and standard deviation of the response for each candidate design. The installation errors are sampled as normal random variables, with their dispersion set from the tabulated limit deviations. During the robust search, the second-order contact parameters are allowed to vary uniformly within their design ranges; in the final assessment they are fixed at the optimized values and only the installation errors are sampled. A multi-island genetic algorithm (MIGA) drives the search. MIGA divides the population into islands, allows periodic migration, and preserves diversity better than a single-population genetic algorithm, which is useful when the objective is a noisy Monte Carlo estimate and the constraints are evaluated by repeated ELTCA calls.
Case study: a 34/43-tooth aviation pair
The procedure was applied to an aviation spiral bevel gear pair with the basic data in Table 1. The initial second-order contact parameters were \(\eta=60^\circ\), \(m’_{21}=0.008\), and \(\delta=0.18\). The design-variable bounds and the statistical treatment are listed in Table 2.
| Parameter | Pinion | Gear |
|---|---|---|
| Number of teeth | 34 | 43 |
| Module at large end (mm) | 2.0 | 2.0 |
| Face width (mm) | 15 | 15 |
| Pressure angle (°) | 20 | 20 |
| Shaft angle (°) | 90 | 90 |
| Hand | Right | Left |
| Variable | Symbol | Lower | Upper | Treatment |
|---|---|---|---|---|
| Contact-trace angle to root cone (°) | \(\eta\) | 30 | 90 | Uniform during robust search |
| First derivative of transmission-ratio function | \(m’_{21}\) | -0.008 | 0.008 | Uniform during robust search |
| Contact-ellipse semi-major axis / face width | \(\delta\) | 0.15 | 0.20 | Uniform during robust search |
| Pinion axial displacement (mm) | \(H_p\) | -0.1 | 0.1 | Normal; dispersion from limit deviation |
| Gear axial displacement (mm) | \(H_g\) | -0.1 | 0.1 | Normal; dispersion from limit deviation |
| Shaft offset (mm) | \(V\) | -0.08 | 0.08 | Normal; dispersion from limit deviation |
| Shaft-angle deviation (rad) | \(\Sigma\) | -0.0045 | 0.0045 | Normal; dispersion from limit deviation |
Three designs were compared: the initial tooth surface, a deterministic optimum obtained from a sensitivity-coefficient method, and a 6σ robust optimum. The installation-error vector was held at \(H_p=-0.0596\) mm, \(H_g=-0.0508\) mm, \(V=0.0762\) mm, and \(\Sigma=0.00485\) rad in all three evaluations, so the comparison isolates the effect of the second-order contact parameters. The optimized parameters are listed in Table 3.
| Quantity | Initial | Deterministic optimum | 6σ robust optimum |
|---|---|---|---|
| \(\eta\) (°) | 60 | 60.5133 | 74.6895 |
| \(m’_{21}\) | 0.008 | 0.0047 | 0.0037 |
| \(\delta\) | 0.18 | 0.1601 | 0.1523 |
The deterministic optimum reduced \(m’_{21}\) and \(\delta\) only modestly and kept the contact-trace angle close to its initial value. The robust optimum moved the trace angle to 74.6895°, reduced the transmission-ratio derivative to 0.0037, and shortened the contact ellipse to 0.1523 of the face width, close to the lower bound. This is a design trade-off: the nominal contact ellipse becomes shorter and the contact path changes orientation, but the pattern is distributed so that it remains inside the functional limits over the installation-error distribution. The area and direction constraints prevent the shorter ellipse from being obtained at the cost of an unacceptable contact area.
For each design, 100 Monte Carlo samples of the installation-error vector were passed through the error-loaded contact analysis. The resulting statistics for the contact-pattern center and the transmission-error equivalent are given in Table 4.
| Response | Statistic | Initial | Deterministic optimum | 6σ robust optimum |
|---|---|---|---|---|
| Contact-pattern center equivalent | Mean | 0.8174 | 0.4702 | 0.4334 |
| Standard deviation | 0.0104 | 0.0206 | 0.0057 | |
| Reported σ level | — | 3.52 | 8 | |
| Transmission-error equivalent | Mean | 0.7598 | 0.3981 | 0.3838 |
| Standard deviation | 0.0096 | 0.0057 | 0.0054 | |
| Reported σ level | 2.454 | 8 | 8 |
Results and interpretation
The initial design is clearly not robust. Its contact-pattern center has a large mean and its transmission-error equivalent reaches only a 2.454σ level, so the upper transmission-error bound can be violated under the installation-error distribution. The deterministic optimization improves the nominal response substantially: the mean contact-pattern center falls from 0.8174 to 0.4702 and the mean transmission-error equivalent falls from 0.7598 to 0.3981. However, the standard deviation of the contact-pattern center increases from 0.0104 to 0.0206, and the contact-pattern response reaches only a 3.52σ level. The deterministic design therefore improves the nominal point while increasing scatter, which is the classic weakness of nominal optimization.
The 6σ robust design avoids that trade-off. Its contact-pattern center mean is 0.4334, lower than the deterministic value, and its standard deviation is 0.0057, about 72% below the deterministic 0.0206. Both the contact-pattern center and the transmission-error equivalent reach the 8σ band, and the one-sided reliability at the active bounds is 99.9937%. The robust design also reduces the minimum transmission error in magnitude: from \(-0.0204^\circ\) for the initial design to \(-0.0067^\circ\) for the deterministic optimum and \(-0.0053^\circ\) for the robust optimum, reductions of 67.12% and 73.33%, respectively.
The standard deviation of the equivalent transmission error changes only slightly between the deterministic and robust designs, from 0.0057 to 0.0054. The more important difference is in the shape and extremes of the loaded transmission-error curve. The robust design reduces the fluctuation range of that curve by more than 30% relative to the deterministic design and lowers the sensitivity of both the contact pattern and the transmission error to installation error by more than 30%. The improvement is therefore not only a smaller sample variance but also a less peaky, less edge-loaded response under misalignment.
These results have a practical interpretation for tooth-surface design. The robust optimum does not attempt to maximize nominal contact area or minimize nominal transmission error alone. It selects a slightly shorter contact ellipse, a different trace angle, and a lower transmission-ratio derivative so that the contact pattern stays centered and the transmission-error curve stays flat when the shafts move. The 6σ formulation converts that requirement into a probabilistic constraint: the design variable means must remain far enough from their bounds that the response distribution stays inside the functional limits. The numerical values are conditioned on the stated load case, the 34/43-tooth geometry, and the assumed installation-error distributions. A different assembly distribution or load level would change the optimum, but the same ELTCA-based 6σ procedure can be rerun for those conditions.
Design implications
The method provides a quantitative way to allocate tooth-surface preset values and tolerances at the same time. Instead of tightening every manufacturing and assembly tolerance, the designer can choose second-order contact parameters that make the gear pair less sensitive to the errors that are difficult to control. The optimization also identifies which response is the limiting one: in the deterministic design, the transmission error reached the 8σ band while the contact-pattern center remained at 3.52σ, so the contact pattern was the weak constraint. The robust design brought both responses into the same reliability band.
The procedure is computationally demanding because each candidate design requires repeated error-loaded contact solutions inside a Monte Carlo loop and a MIGA search. Its value lies in the resulting design logic: the contact pattern and transmission error are treated as distributions rather than single values, and the optimizer is forced to find a tooth surface whose performance remains acceptable when the installation is not ideal. For aviation spiral bevel gears, where alignment cannot be made perfect at acceptable cost, that is a more realistic design target than a nominal optimum alone.