Spiral bevel and hypoid gearsets carry torque through a small number of simultaneously meshing tooth pairs whose stiffness changes periodically as contact rolls across the flank. That periodic stiffness, superimposed on the deliberately non-conjugate flank form, produces a kinematic error at the output that is never a pure constant: the transmission error (TE). Because the error repeats once per mesh cycle, its harmonic content appears directly in the gear whine radiated by helicopter main reducers, electric-axle drives and other high-speed bevel applications. Tooth flank geometry — and therefore the shape of the TE curve — is fixed almost entirely by machine-tool settings, which makes TE an inexpensive place to attack the noise problem.
Two Transmission Errors, One Design Lever
It is useful to separate two quantities that are often conflated. The geometric transmission error is the deviation of the output rotation from the ideal ratio in the unloaded state; it is a pure function of flank topography and is obtained from tooth contact analysis (TCA). The loaded transmission error adds the compliance of the mesh: normal tooth deflection under torque, converted into an equivalent angular error, so that the resulting curve depends on load, on the actual contact ratio at that load, and on the edges of the contact pattern.
Both quantities matter for noise. The geometric curve sets what the gear train does when lightly loaded, where the tooth pairs hand over from one to the next; the loaded curve determines the dynamic excitation at operating torque. A design procedure therefore has to control two things at once: the symmetry of the geometric curve about the reference point, so that entry and exit behave alike and edge contact does not begin prematurely, and the peak-to-peak fluctuation of the loaded curve at working load, which is the direct source of stiffness excitation.
Where the Second-Order Parabolic Curve Stops Working
The standard answer to misalignment sensitivity is a parabolic TE: a second-order curve centred on the reference point whose slope is zero there. A parabola of this kind absorbs the linear component of manufacturing and assembly error, which is why it has been the default for decades and why high-contact-ratio variants built on the same idea improve load sharing. Its weakness is kinematic rather than geometric. At the transfer point, where contact passes from one tooth pair to the next, the incoming and outgoing branches of the error curve intersect at a finite angle. The slope of the TE curve is proportional to the instantaneous ratio deviation, so a corner in the curve is a step in angular velocity, and a step in velocity is an impact — the acceleration discontinuity that shows up as a high-order line in the spectrum.
Higher-order motion curves were introduced to blunt that corner. Gleason’s high-order TE concept replaces the parabola with a curve whose branches meet at a much larger included angle at the transfer point, and fourth- and seventh-order functions have been realised on modified-roll machines by raising the order of the roll-variation coefficients. Curvature-correction schemes achieve a similar effect by reshaping the flank rather than the motion. What these approaches share is a single analytic family stretched over the whole mesh cycle, so the shape at the transfer point and the shape inside the single-tooth contact zone are coupled.
The hybrid approach decouples them. The mesh cycle is split into three regions: a sinusoidal segment covering the single-tooth contact zone, and two higher-order parabolic segments covering the entry and exit regions where the neighbouring pairs share load. The sinusoid supplies the property that matters at handover — all of its derivatives are continuous — while the parabolas supply the amplitude and steepness control that keeps the contact pattern inside the flank.
Optimizing the Local Synthesis
Before any error curve can be imposed, the underlying conjugate design must be chosen. Three local synthesis parameters govern the shape and position of the contact pattern and the first-order behaviour of the ratio:
- $\Delta y$ — the tooth-height position of the reference point $M_0$ on the flank;
- $\eta_2$ — the angle between the contact path $AE$ and the root cone line $\Gamma_r$;
- $m’_{21}$ — the first derivative of the gear ratio at the reference point.
These are chosen by a weighted two-objective optimization. The first objective measures how evenly the mesh span is divided between entry and exit; the second measures the fluctuation of the loaded transmission error $T_e$ over one mesh cycle:
$$F(\Delta y,\ \eta_2,\ m’_{21})=\min\left\{\omega f_1+(1-\omega)f_2\right\}$$
$$f_1=\frac{\left|\phi_{1A}-\phi_{1C}\right|}{\left|\phi_{1E}-\phi_{1C}\right|},\qquad f_2=\max(T_e)-\min(T_e)$$
Here $\phi_{1A}$, $\phi_{1C}$ and $\phi_{1E}$ are the pinion rotation angles at the entry point, the reference point and the exit point, and $\omega$ is the weighting coefficient. The reference point itself is forced to be conjugate: at $M$ the local ratio equals the theoretical ratio, so $\delta_{2C}=0$ by construction. The entry and exit angles and their error amplitudes follow from the TCA equations combined with the geometry of simultaneous tip contact on pinion and gear; $T_e$ comes from a loaded tooth contact analysis (LTCA) model that computes normal tooth deflection over a mesh cycle and converts it into an equivalent rotation.
The constraints bound each variable by physical reasoning rather than by the optimizer:
$$\text{s.t.}\quad \Delta y_{\min}\le\Delta y\le\Delta y_{\max},\qquad \eta_{2\min}\le\eta_2\le\eta_{2\max},\qquad m’_{21\min}\le m’_{21}\le m’_{21\max}$$
$\Delta y$ is limited by the addendum and dedendum at the gear mid-point, $\eta_2$ by tooth strength and mounting accuracy, and $m’_{21}$ by the working load. Because the chain from design variables to objectives passes through local synthesis, TCA and LTCA in sequence, no closed-form mapping exists and the objective landscape has several local minima, so a genetic algorithm is used: encoding, initial population, fitness evaluation, selection, crossover and mutation, iterated until convergence.
Building the Three-Segment Curve
The hybrid curve is defined on five key points of the mesh cycle: the entry point $A$, the entry transfer point $B$, the reference point $C$, the exit transfer point $D$, and the exit point $E$. The transfer points are located half a mesh period $T_m$ either side of the reference point, $\phi_{1B}=\phi_{1C}-T_m/2$ and $\phi_{1D}=\phi_{1C}+T_m/2$, and their error amplitudes $\delta_{2B}$ and $\delta_{2D}$ follow from substituting those angles into the TCA equations. In the single-tooth contact zone the curve is sinusoidal:
$$\delta_2=A_m\sin\!\left[\omega_m(\phi_1-\phi_{10})-\omega_{m0}\right]+A_m$$
where $\omega_m$ is the angular frequency of the oscillation, $\omega_{m0}$ its phase, and $A_m=\delta_{2B}/2$ the amplitude, so that the segment swings between zero and the transfer-point value. The entry and exit regions carry higher-order parabolas:
$$\delta_2=a_{2B}(\phi_1-\phi_{1B})^{k}+\delta_{2B},\qquad k=2,4,\dots$$
$$a_{2B}=\frac{\delta_{2A}-\delta_{2B}}{(\phi_{1A}-\phi_{1B})^{k}}$$
with a mirror-image expression on the exit side, based at $D$ and anchored to the exit amplitude $\delta_{2E}$. The even order $k$ guarantees that $k-1$ derivatives vanish at the transfer point, so the parabola arrives with zero slope and the sinusoid leaves with zero slope. The two branches are therefore collinear at $B$ and at $D$: they meet at an included angle of 180°, the ratio deviation varies continuously through the handover, and the acceleration jump that a second-order parabola leaves behind disappears. Assembling the pieces gives the complete error function for one tooth pair, from entry to exit:
$$\delta_2=\begin{cases} a_{2B}(\phi_1-\phi_{1B})^{k}+\delta_{2B}, & \phi_1\in(\phi_{1A},\phi_{1B}]\\ A_m\sin\!\left[\omega_m(\phi_1-\phi_{10})-\omega_{m0}\right]+A_m, & \phi_1\in(\phi_{1B},\phi_{1D}]\\ a_{2C}(\phi_1-\phi_{1D})^{k}+\delta_{2D}, & \phi_1\in(\phi_{1D},\phi_{1E}] \end{cases}$$
The order $k$ controls how steeply the curve rises from a transfer point towards the entry or exit extremity. Increasing $k$ sharpens the shoulder of the curve but has little further influence on mesh behaviour, and it becomes progressively harder to realise on a machine, which is why $k=2$ is the practical choice.
Converting the Curve into Machine Settings
A curve is only useful if a grinding machine can produce it. The link is the definition of transmission error itself, which fixes the gear rotation for every pinion rotation:
$$\phi_2=\delta_2+\frac{Z_1}{Z_2}\left(\phi_1-\phi_{1C}\right)+\phi_{2C}$$
With $Z_1$ and $Z_2$ the pinion and gear tooth numbers and $\phi_{1C}$, $\phi_{2C}$ the initial mesh angles. Each of the three segments is discretized into $n-1$ equal intervals in $\phi_1$, and the corresponding $\phi_2$ values are obtained from the segment expression. Using the pinion roll ratio as parameter, an inverse TCA then returns the instantaneous roll ratio $r_a$ and the cradle rotation $\phi_{c1}$ at each discretized point. Fitting the resulting series with polynomials of order $N$ gives the modified-roll law that the machine must follow:
$$r_a=\begin{cases} b_{10}+b_{11}\phi_{c1}+b_{12}\phi_{c1}^{2}+\dots+b_{1N_1}\phi_{c1}^{N_1}, & \phi_1\in(\phi_{1A},\phi_{1B}]\\ b_{20}+b_{21}\phi_{c1}+b_{22}\phi_{c1}^{2}+\dots+b_{2N_2}\phi_{c1}^{N_2}, & \phi_1\in(\phi_{1B},\phi_{1D}]\\ b_{30}+b_{31}\phi_{c1}+b_{32}\phi_{c1}^{2}+\dots+b_{3N_3}\phi_{c1}^{N_3}, & \phi_1\in(\phi_{1D},\phi_{1E}] \end{cases}$$
The constant terms are the basic roll ratios of the three regions and the remaining coefficients are the higher-order modified-roll coefficients. Commercial CNC bevel grinding machines — Gleason Phoenix and Klingelnberg C-series machines among them — accept roll-variation polynomials up to the sixth power, so a three-segment law of this type is within reach of standard hardware without special options.
Worked Example
The procedure was applied to the working flanks of a 23/65 spiral bevel pair with a 3.9 mm module, a 25° mean spiral angle and a 37 mm face width, intended for a working torque of 400 N·m. Geometry and the optimized local synthesis parameters are listed in Tables 1 and 2.
| Parameter | Pinion | Gear |
|---|---|---|
| Number of teeth | 23 | 65 |
| Module (mm) | 3.9 | 3.9 |
| Mean spiral angle (°) | 25 | 25 |
| Hand of spiral | right | left |
| Shaft angle (°) | 90 | 90 |
| Pitch cone angle (°) | 70.5138 | 19.4861 |
| Face cone angle (°) | 71.6718 | 21.7770 |
| Root cone angle (°) | 68.2229 | 18.3281 |
| Mid-point cone distance (mm) | 115.951 | 115.951 |
| Face width (mm) | 37 | 37 |
| Addendum (mm) | 1.9844 | 4.6455 |
| Dedendum (mm) | 5.3787 | 2.7176 |
| Local synthesis parameter | Value |
|---|---|
| Contact path direction $\eta_2$ (°) | 15.713 |
| First ratio derivative $m’_{21}$ | −0.008512 |
| Contact ellipse major axis $2a$ (mm) | 11.1 |
| Elastic deflection $\delta$ (mm) | 0.00635 |
| Pitch cone offset $\Delta x$ (mm) | 0.05 |
| Tooth height offset $\Delta y$ (mm) | 0.5422 |
| Design contact ratio $\varepsilon_r$ | 2.2 |
Optimization used a population of 50 individuals over 100 generations with a crossover probability of 0.5 and a mutation probability of 0.1, settings chosen after trials on the problem size rather than by rule of thumb. The resulting machine settings for the working flank give the pinion a basic roll ratio of 2.80575 with second- and third-order modified coefficients of 0.0648 and 0.1066, against a basic roll ratio of 1.05990 and zero modified coefficients for the gear’s convex flank — that is, the modification is carried entirely on the pinion side.
| Key point | Pinion angle $\phi_1$ (°) | TE amplitude $\delta_2$ (arcsec) |
|---|---|---|
| $A$ — entry | −15.5498 | −80.0192 |
| $B$ — entry transfer | −6.0743 | −15.0000 |
| $C$ — reference | −1.7517 | 0 |
| $D$ — exit transfer | 9.5778 | −15.0000 |
| $E$ — exit | 18.9627 | −79.2132 |
Entry and exit amplitudes come out within 0.8 arcsec of each other (−80.02″ against −79.21″), so the two ends of the mesh cycle are loaded alike. For the hybrid curve itself the sinusoidal segment was set with $\omega_m=23$, $\omega_{m0}=\pi$ and $A_m=-7.5”$, giving an oscillation spanning 0 to −15.0″ between the transfer points, and the parabolic segments used $k=2$ with $a_{2B}=a_{2C}=-0.0115$. The inverse TCA discretized each region into $n_1=n_2=n_3=100$ divisions with polynomial orders $N_1=N_2=N_3=5$; the maximum fitting error of the instantaneous roll ratio stayed below $1.0\times10^{-5}$, which is the accuracy that makes the three-segment law usable as a machine input.
| Coefficient | Entry branch | Sinusoidal zone | Exit branch |
|---|---|---|---|
| Basic roll ratio | 2.81868 | 2.86216 | 2.90589 |
| 2nd order | −0.75340 | 0.07416 | −0.73830 |
| 3rd order | 15.38470 | 0.01231 | −15.60000 |
| 4th order | −154.50197 | −25.60220 | −153.45930 |
| 5th order | 814.93680 | 37.06650 | −818.28210 |
| 6th order | −1765.30910 | 274.51730 | −1792.60890 |
Contact Pattern, Flank Deviation and Loaded Behaviour
The hybrid curve is only a fair substitute if it leaves the contact pattern alone, and it does: the TCA contact ellipse, the design contact ratio of 2.2 and the entry and exit amplitudes of the hybrid design all sit close to those of the parabolic baseline, which allows the two designs to be compared on mesh behaviour rather than on contact geometry.
The price is a small amount of extra flank material. Taking the parabolic design as the reference surface and projecting the hybrid pinion flank along its normal gives a deviation map with a maximum of about +81 μm at the heel root, positive deviation also at the toe tip, and the opposite sign near the mid-flank reference region. The pattern follows directly from the curve: less modification is applied between the entry point and the entry transfer point, and between the exit transfer point and the exit point, than the parabola would apply there, while near the reference point the situation reverses.
Under torque the two designs behave in much the same way qualitatively. Loaded transmission error was computed for gear torques of 5, 100, 200, 300, 400, 500, 600, 700, 800, 900 and 1000 N·m. From light load up to the 400 N·m working point the error curves keep the shape of the geometric curve and their amplitude falls as load rises; beyond the working load the shape changes and the amplitude grows again. The reason is a change of contact ratio: near 400 N·m the mesh moves from two tooth pairs sharing load to three. The distinguishing feature of the hybrid design is quantitative — its peak-to-peak loaded fluctuation remains below that of the parabolic design over the load sweep, which is the direct measure of the stiffness excitation that drives vibration.
Harmonic Content of the Loaded Error
Fluctuation alone does not determine noise; the distribution of that fluctuation across harmonics does. A Fourier decomposition of the loaded transmission error at 400 N·m into its first six orders shows every order reduced for the hybrid design, not only the low orders that follow from the smoother transfer points. Averaged over the six orders, the amplitude falls from 0.6582″ for the parabolic design to 0.527″, a reduction of 19.86%.
| Metric at 400 N·m | Parabolic TE | Hybrid TE | Change |
|---|---|---|---|
| Mean amplitude, orders 1–6 (arcsec) | 0.6582 | 0.5270 | −19.86% |
| Order-by-order comparison | every order lower for the hybrid design | — | |
This is the result that connects the geometric construction to the acoustic problem. High-order harmonics of the loaded error are what radiate as high-frequency whine in a bevel stage, and a design that lowers the whole harmonic ladder at working load attacks that mechanism more broadly than one that only smooths the handover. The same ordering of amplitudes persists at other torques, with the curves shifting in the same way as at the working point.
Practical Constraints and the Boundaries of the Evidence
Three limitations deserve to be stated plainly. First, the acoustic benefit is inferred from the harmonic content of the computed loaded transmission error; no rig or vehicle measurement of sound pressure is part of this assessment, and the step from a 19.86% amplitude reduction to a measurable noise reduction depends on the structural response of the housing and the transfer path, neither of which is modelled. Second, the choice $k=2$ is a manufacturing compromise: higher orders sharpen the curve further but add little to mesh performance while making the roll law harder to realize on a grinder with a sixth-order polynomial limit. Third, the design is defined about a single reference point and a single working load; sensitivity to shaft misalignment and to load excursions well above 400 N·m is not quantified here, and the parabolic design retains its well-known advantage in absorbing linear mounting error that the hybrid curve inherits only through its parabolic flanks.
What the design does establish is a working division of labour within one mesh cycle: a sinusoidal core that guarantees derivative continuity where tooth pairs hand over, parabolic shoulders that control amplitude and flank position, and a three-segment modified-roll law that puts both onto an ordinary CNC bevel grinder. The construction is geometric, the machine input is a polynomial, and the benefit is concentrated where the noise problem lives — the higher harmonics of the loaded error at operating torque.