Spiral bevel gears are used in vehicle drivelines, machine tools, and other transmissions where a smooth, high-load connection between intersecting shafts is required. The conventional five-cut process separates the gear and pinion into five operations: gear roughing, gear finishing, pinion roughing, pinion concave finishing, and pinion convex finishing. A two-cut alternative completes roughing and finishing for each member on one machine. For the pinion, the duplex helical method performs both flanks in a single setup, reducing non-productive time and producing a continuous root fillet. The constraint is equally important: the concave and convex flanks share the same machine settings, so a setting error or correction acts on both flanks at once.
The engineering problem is therefore twofold. First, the generated flank must be represented as a function of the machine settings, including the additional helical motion that distinguishes duplex helical machining from a fixed-bed process. Second, the influence of each setting on the flank topology must be ranked so that a correction vector can be chosen when measured or prescribed deviations appear. The model below addresses both points for a representative 18-tooth pinion and 35-tooth gear pair.
Kinematic model of duplex helical generation
Let \(S_m\) be the machine frame, with origin \(O_m\) and the \(X_mO_mY_m\) plane in the machine plane. The cutter frame is \(S_t\), the cradle frame is \(S_c\), and the workpiece frame is \(S_w\). Two intermediate frames, \(S_a\) and \(S_b\), are attached to the cutter-tilt and cutter-swivel mechanisms. The cutter is located by the tilt angle \(i\), swivel angle \(j\), radial cutter position \(S_r\), and angular cutter position \(q\). The workpiece is located by the vertical offset \(E_m\), bed position \(X_B\), mounting angle \(\gamma\), horizontal offset \(X_g\), and helical motion coefficient \(H\). The roll ratio \(R_{ap}\) fixes the ratio between cradle speed \(\omega_c\) and workpiece speed \(\omega_w\). During generation, \(\phi_c\) is the cradle angle and \(\phi_w\) is the corresponding workpiece rotation.
The distinctive term is the bed motion. In a conventional fixed-bed model, the bed coordinate remains at \(X_B\). In duplex helical machining, the effective bed position along \(Z_m\) changes with the cradle angle as \(X_B-H\phi_c\), where \(H\) has units of mm/rad. The generated workpiece surface is therefore
$$
\mathbf r_w(u,\theta,\phi_c)
= M_{we}(R_{ap})M_{ed}(\gamma)M_{dm}(E_m,X_B,H)M_{mc}(\phi_c)M_{cb}(S_r,q)M_{ba}(j)M_{at}(i)\mathbf r_t(u,\theta)
$$
where \(\mathbf r_t(u,\theta)\) is the cutter surface in \(S_t\), and \(u\) and \(\theta\) are the cutter surface parameters. The matrix containing the helical bed motion is
$$
M_{dm}(E_m,X_B,H)=
\begin{bmatrix}
1 & 0 & 0 & 0\\
0 & 1 & 0 & E_m\\
0 & 0 & 1 & -(X_B-H\phi_c)\\
0 & 0 & 0 & 1
\end{bmatrix}.
$$
The remaining matrices are rigid-body rotations and translations between the machine, cradle, cutter, and workpiece frames. When \(H=0\), the bed is fixed and the formulation reduces to the conventional five-cut fixed-bed case. The duplex helical model is therefore a superset of the fixed-bed model rather than a separate kinematic convention.
The workpiece normal is obtained from the surface derivatives:
$$
\mathbf n_w=
\frac{\mathbf r_{w,u}\times\mathbf r_{w,\theta}}
{\left\|\mathbf r_{w,u}\times\mathbf r_{w,\theta}\right\|},
$$
where subscripts \(u\) and \(\theta\) denote partial derivatives. The meshing condition is
$$
f(u,\theta,\phi_c)=\mathbf n_w\cdot
\frac{\partial\mathbf r_w}{\partial\phi_c}=0.
$$
For each point on a rotary projection grid, the meshing equation and the projection conditions determine \(u\), \(\theta\), and \(\phi_c\). Substituting these values back into the surface equation gives a discrete tooth surface. A 45-point grid is used on each flank in the example below; the axial coordinate and the radial distance from the workpiece axis define the grid.
Flank deviation and machine-setting sensitivity
Let \(\Sigma_1\) be the theoretical surface generated by the nominal settings and \(\Sigma_2\) the actual surface generated with setting errors. Before comparison, \(\Sigma_2\) is rotated about the \(X_m\) axis until its grid midpoint coincides with the midpoint of \(\Sigma_1\). This removes the rigid-body offset and leaves the shape deviation. At a corresponding grid point, the scalar deviation \(\delta\) along the theoretical normal satisfies
$$
\mathbf r_w(u,\theta,\phi_c)+\delta\,\mathbf n_w(u,\theta,\phi_c)
=\mathbf r_w'(u,\theta,\phi_c),
$$
where \(\mathbf r_w’\) is the rotated actual surface. Positive \(\delta\) means that the actual surface lies outside the theoretical surface along the outward normal.
To separate the deviation into interpretable modes, the grid deviation \(Z\) is fitted to a second-order surface:
$$
Z(X,Y)=a_1X+a_2Y+a_3XY+a_4X^2+a_5Y^2.
$$
Here \(X\) is the lengthwise direction, \(Y\) is the profile direction, and \(Z\) is the normal deviation. The coefficient \(a_1\) represents the spiral-angle slope, \(a_2\) the pressure-angle slope, \(a_3\) the flank twist, \(a_4\) the lengthwise crowning, and \(a_5\) the profile crowning. With 45 grid points per flank, the five coefficients are obtained by linear least squares. The fit is a local description of the deviation field, not a replacement for the full generated surface.
The example gear pair and its nominal machine settings are listed in Tables 1 and 2. The gear is right-hand with 35 teeth, and the pinion is left-hand with 18 teeth. The gear setting uses \(H=0\), while the pinion uses \(H=5.0336\) mm/rad, so the pinion is the member for which the helical bed motion is active.
| Parameter | Gear | Pinion |
|---|---|---|
| Shaft angle \(\Sigma\) (°) | 90 | |
| Module \(m\) (mm) | 7.914 | |
| Spiral angle \(\beta\) (°) | 35 | |
| Pressure angle \(\alpha_0\) (°) | 20 | |
| Face width \(b\) (mm) | 53 | |
| Number of teeth \(z\) | 35 | 18 |
| Hand | Right | Left |
| Addendum \(h_a\) (mm) | 4.04 | 9.96 |
| Dedendum \(h_f\) (mm) | 11.58 | 5.66 |
| Parameter | Gear | Pinion |
|---|---|---|
| Cutter tilt \(i\) (°) | 2.73 | 2.78 |
| Cutter swivel \(j\) (°) | 216.62 | 188.42 |
| Radial cutter position \(S_r\) (mm) | 131.4036 | 131.6021 |
| Angular cutter position \(q\) (°) | 71.62 | 72.42 |
| Vertical offset \(E_m\) (mm) | 0 | 0.018 |
| Bed position \(X_B\) (mm) | 9.1304 | 6.4497 |
| Helical motion coefficient \(H\) (mm/rad) | 0 | 5.0336 |
| Mounting angle \(\gamma\) (°) | 57.85 | 25.14 |
| Horizontal offset \(X_g\) (mm) | 0 | -0.1459 |
| Roll ratio \(R_{ap}\) | 1.120322 | 2.186118 |
A perturbation study with +0.1 mm for linear settings and +0.1° for angular settings shows how each setting changes the flank topology. Cutter tilt \(i\) mainly changes the pressure-angle slope. Cutter swivel \(j\) has a small effect. Radial cutter position \(S_r\) mainly changes the spiral-angle slope. Angular cutter position \(q\) also has a small effect. Vertical offset \(E_m\) affects both the spiral-angle and pressure-angle slopes. Bed position \(X_B\) has a relatively small influence. Helical motion \(H\) affects both slope terms, but with different diagonal trends on the two flanks: on the convex flank it mainly acts near the heel addendum and toe dedendum, while on the concave flank it acts near the heel dedendum and toe addendum. This makes \(H\) a useful setting for diagonal contact control. Mounting angle \(\gamma\) mainly affects the pressure angle. Horizontal offset \(X_g\) mainly affects the pressure angle and also changes the spiral angle. The roll ratio \(R_{ap}\) affects the pressure angle and, to a lesser extent, the spiral angle.
The second-order coefficients give a more quantitative ranking. For the pinion concave flank, the spiral-angle coefficient is most sensitive to \(S_r\), followed by \(i\), \(E_m\), \(H\), and \(\gamma\). The pressure-angle coefficient is most sensitive to \(E_m\), \(i\), \(\gamma\), \(X_g\), and \(X_B\). The twist and crowning coefficients are most sensitive to \(i\), \(\gamma\), \(E_m\), \(H\), and either \(S_r\) or \(X_g\), depending on the term. The convex flank differs in detail, so the coefficients must be recomputed for each gear pair rather than transferred directly.
When the full 45-point grid is used, the global squared-error ranking for the example is given in Table 3. The two flanks do not respond identically: the bed position \(X_B\) is more important on the non-working flank than on the working flank, while \(X_g\) is more important on the working flank. These rankings apply to the stated perturbation magnitudes and to this 18/35 pair.
| Rank | Working flank | Non-working flank |
|---|---|---|
| 1 | Roll ratio \(R_{ap}\) | Roll ratio \(R_{ap}\) |
| 2 | Radial cutter position \(S_r\) | Radial cutter position \(S_r\) |
| 3 | Vertical offset \(E_m\) | Vertical offset \(E_m\) |
| 4 | Cutter tilt \(i\) | Cutter tilt \(i\) |
| 5 | Horizontal offset \(X_g\) | Bed position \(X_B\) |
| 6 | Mounting angle \(\gamma\) | Horizontal offset \(X_g\) |
| 7 | Helical motion \(H\) | Mounting angle \(\gamma\) |
| 8 | Bed position \(X_B\) | Helical motion \(H\) |
| 9 | Cutter swivel \(j\) | Cutter swivel \(j\) |
| 10 | Angular cutter position \(q\) | Angular cutter position \(q\) |
Linearized correction model
Let \(\boldsymbol\psi\) be the vector of \(m\) machine settings and \(\Delta\boldsymbol\psi\) a correction applied to the nominal settings. The residual between the theoretical surface \(\mathbf r_1\) and the actual surface \(\mathbf r_2\) is
$$
\Delta\mathbf r=\mathbf r_1(\boldsymbol\psi)-\mathbf r_2(\boldsymbol\psi+\Delta\boldsymbol\psi).
$$
For small corrections, first-order expansion gives
$$
\delta(\Delta\mathbf r)
=-\sum_{k=1}^{m}
\frac{\partial\mathbf r_2}{\partial\psi_k}\Delta\psi_k.
$$
Projecting this residual onto the theoretical normal \(\mathbf n_w\) gives the scalar deviation at grid point \(s\):
$$
\Delta\delta_s
=-\sum_{k=1}^{m}
\left(
\frac{\partial\mathbf r_2}{\partial\psi_k}\cdot\mathbf n_w
\right)\Delta\psi_k.
$$
The sensitivity coefficient \(\eta_{sk}\) is the change in deviation at point \(s\) caused by a unit change in setting \(k\), with the sign included. In matrix form,
$$
\{\Delta\delta_s\}=[\eta_{sk}]\{\Delta\psi_k\},
\qquad s=1,\ldots,2p,\quad k=1,\ldots,m,
$$
where \(p\) is the number of grid points per flank and the factor \(2p\) stacks the concave and convex flanks. The sensitivity matrix is populated by finite differences: one setting is perturbed, the surface is regenerated, the change is projected onto the normal, and the result is divided by the perturbation. With \(m=10\) settings and \(2p=90\) grid points, the system is overdetermined. The least-squares correction is
$$
\{\Delta\psi_k\}
=\left([\eta_{sk}]^{\mathsf T}[\eta_{sk}]\right)^{-1}
[\eta_{sk}]^{\mathsf T}\{\Delta\delta_s\}.
$$
Because both flanks share the same settings, the solution minimizes the combined squared deviation rather than correcting each flank independently. This is the central compromise of duplex helical correction: a setting that improves the concave flank may worsen the convex flank unless the sensitivity matrix is solved as a coupled system.
Numerical correction example
The numerical test begins with a prescribed initial deviation field, not a measured physical flank. The field is used to demonstrate the correction procedure. Applying the least-squares model gives the setting corrections in Table 4. Some corrections are large because the initial field is prescribed rather than a small manufacturing error; they should be read as a test of the method, not as a production calibration.
| Setting | Correction |
|---|---|
| Cutter tilt \(i\) (°) | 0.4275 |
| Cutter swivel \(j\) (°) | -1.0816 |
| Radial cutter position \(S_r\) (mm) | -1.744 |
| Angular cutter position \(q\) (°) | -2.15 |
| Vertical offset \(E_m\) (mm) | -1.743 |
| Bed position \(X_B\) (mm) | 1.3465 |
| Helical motion coefficient \(H\) (mm/rad) | -0.0867 |
| Mounting angle \(\gamma\) (°) | 0.5190 |
| Horizontal offset \(X_g\) (mm) | -0.1397 |
| Roll ratio \(R_{ap}\) | -0.0231 |
After correction, the convex-flank maximum deviation falls from 32 μm to -7.8 μm near the heel addendum, and the concave-flank maximum falls from 17.7 μm to -9.5 μm. The sign change indicates that the corrected surface crosses the theoretical surface rather than remaining on one side. The largest remaining deviations are concentrated near the flank edges, where they have less influence on the loaded contact pattern.
| Flank | Before correction (μm²) | After correction (μm²) |
|---|---|---|
| Concave | 9,000 | 623 |
| Convex | 1,800 | 293 |
| Combined | 10,800 | 916 |
The combined squared error falls by about 91.5%, from 10,800 μm² to 916 μm². The concave flank falls by about 93.1%, and the convex flank by about 83.7%. The improvement is therefore not limited to one flank, although the residual is not zero. This is expected from a linearized correction with a finite set of machine settings and a prescribed initial field.
Implications for process design
The kinematic model provides a single formulation for duplex helical machining and for the fixed-bed five-cut case through \(H=0\). For process design, the important result is the coupling between the two flanks. The roll ratio and radial cutter position dominate the global squared error in this example, but the local error shape is controlled by a different combination: cutter tilt and mounting angle strongly affect pressure angle and twist, vertical offset affects both slope terms, and helical motion provides a diagonal adjustment that is not available in a fixed-bed process. A correction strategy that uses only one dominant setting will therefore reduce the total error at the cost of an unfavorable error distribution.
The sensitivity ranking is gear-pair specific. Changes in tooth count, spiral angle, face width, or nominal cutter tilt alter the lever arms and the contact geometry, so the matrix should be rebuilt for each case. The linearized solution is also local: for large deviations, the correction should be applied in steps, with the surface regenerated and the sensitivity matrix updated between steps. The numerical example shows that the method can reduce a prescribed deviation field substantially, but production use still requires measured flank data and iterative verification on the machine.