Spiral bevel gears transmit torque between intersecting shafts with a curved tooth trace that lengthens contact, reduces impact, and allows the contact pattern to be placed away from the tooth edges. Their flanks are not cut to a simple profile. They are generated by a coordinated roll between a rotating cutter head and the workpiece, and the geometry of the flank is therefore controlled by the relative motion, not only by the cutter shape. The nominal roll ratio i0 fixes the average ratio between cradle rotation and workpiece rotation. Modern CNC gear machines can superimpose a nonlinear correction on that ratio, so that the instantaneous roll varies along the generating stroke. This is the role of high-order roll ratio correction coefficients: they redistribute the generated flank without changing the cutter profile or the basic machine setup.
Traditional proportional correction depends heavily on operator experience. A gear measuring center can supply a grid of normal deviations, but a grid alone does not indicate which machine parameter should be changed. The useful quantity is a sensitivity coefficient—the change in a specific shape mode per unit change in a correction coefficient. Once those coefficients are known, the measured topology can be decomposed into tilt, twist, and curvature, and the appropriate order can be selected for adjustment.
Generating model with a polynomial roll correction
Consider a left-hand pinion in the machine coordinate system. The origin O is at the machine center, the i–j plane is the machine plane, and the k axis points into the cradle. The cutter spindle is positioned by the radial cutter position S, angular cutter position q, vertical offset Em, bed position XB, horizontal position Xp, and workpiece mounting angle δM; i0 is the roll ratio. The straight portion of the cutter edge generates the working flank, while the circular corner generates the root fillet. Because the measured grid is inset from the tooth boundaries, the root fillet is normally outside the evaluated region.
The cutter edge is represented by
\[
\mathbf r_c=\mathbf r_0+s_1\mathbf t_1
\]
where \(\mathbf r_0\) is the tip vector, \(s_1\) is the distance from the tip along the edge, and \(\mathbf t_1\) is the unit vector of the straight edge. A point on the workpiece flank is obtained by rotating the cutter point about the workpiece axis \(\mathbf p\) through the generating angle \(\phi\):
\[
\mathbf r_w=(\mathbf p\cdot \mathbf r_c)\mathbf p+\cos\phi\left[(\mathbf p\times \mathbf r_c)\times\mathbf p\right]+\sin\phi\,(\mathbf p\times \mathbf r_c)
\]
\[
\mathbf n_w=(\mathbf p\cdot \mathbf n_c)\mathbf p+\cos\phi\left[(\mathbf p\times \mathbf n_c)\times\mathbf p\right]+\sin\phi\,(\mathbf p\times \mathbf n_c)
\]
Here \(\mathbf n_c\) is the cutter-edge normal at the same point and \(\mathbf n_w\) is the corresponding workpiece-flank normal. The meshing equation supplies the contact condition and determines the cradle angle, cutter-phase angle, and workpiece rotation at each contact point.
The high-order correction is introduced into the roll relation as a polynomial in the cradle-angle increment \(\Delta q\):
\[
\phi=i_0\left(\Delta q+K_2\Delta q^2+K_3\Delta q^3+K_4\Delta q^4+K_5\Delta q^5+K_6\Delta q^6\right)
\]
The coefficients are signed. In the notation used for the machine setting, \(K_2=2c\), \(K_3=6d\), \(K_4=24e\), \(K_5=120f\), and \(K_6=720g\). Setting all \(K_n\) to zero recovers the uniform roll. A nonzero \(K_n\) makes the instantaneous roll ratio vary with cradle position, so the generating motion is no longer a constant-ratio roll. The resulting flank change is not a uniform offset: it has tilt, twist, and curvature components distributed over the tooth surface.
Discretization and normal-deviation topology
For measurement and sensitivity analysis, the theoretical flank is discretized in its axial projection. A 9 × 5 grid is used: nine columns from toe to heel and five rows from root to tip, giving 45 points. The grid is inset from the boundaries to avoid edge burrs and root fillets. The general margin is 10% of face width at the toe and heel and 5% of full tooth height at the tip and root, with a minimum of 0.6 mm, following the AGMA 2009-B01 practice. For the test pinion, the actual insets were 1.5 mm at the toe, 2.5 mm at the heel, 1.0 mm at the tip, and 1.7 mm at the root.
In the axial section, let \(x\) be the distance from the crossing point along the gear axis and \(y\) the radial distance from the axis. These are obtained from the workpiece vector as
\[
y=\left|\mathbf r_w\times\mathbf p\right|,\qquad x=-\mathbf r_w\cdot\mathbf p
\]
Given a grid point \((x,y)\), a two-variable iteration solves for the corresponding cradle angle and cutter-phase angle. The tooth-surface vector and normal are then calculated from the model and transformed into the measurement coordinate system. This is the basis for both the theoretical point set and the measured deviation grid.
To compare an error flank with the theoretical flank, the error flank is rotated about the gear axis until its grid center coincides with the theoretical center. The signed distance along the normal between the two flanks at each grid point is then the normal flank deviation. In the topology plots, a heavy solid line denotes the theoretical flank; a fine solid line indicates that the theoretical flank lies below the actual flank, and a fine dashed line indicates that it lies above. Corner labels give representative deviations in micrometres.
| Parameter | Wheel | Pinion |
|---|---|---|
| Hand | Right | Left |
| Number of teeth | 17 | 17 |
| Shaft angle | 90° | |
| Mean spiral angle | 40° | |
| Module | 3 mm | |
| Pressure angle | 17.5° | |
| Face width | 12 mm | |
Topology patterns produced by the correction orders
The test pinion was assumed to be produced by the duplex helical method. Starting from the nominal machine setting, the second- through sixth-order coefficients were changed one at a time by \(+0.02\), \(+0.2\), \(+2\), \(+20\), and \(+200\), respectively. The increments increase by a decade with order so that the induced deviations remain visible in the topology plots; they are not equal relative changes. Three consistent features appear.
First, the largest deviations are concentrated near the tooth tip, particularly at the toe on the concave flank and at the heel on the convex flank. The mid-flank and root regions change much less. A roll correction therefore behaves more like a local end-and-tip trim than like a uniform tooth-thickness change.
Second, the heel shows approximately the same trend on both flanks, whereas the toe depends on the order parity. The second-, fourth-, and sixth-order changes produce a concave actual flank below the theoretical flank and a convex actual flank above it. The third- and fifth-order changes reverse this relation: the concave actual flank lies above the theoretical flank and the convex actual flank below it. This alternation is consistent with the alternating signs of the polynomial terms and with the corresponding sensitivity coefficients.
Third, the deviation is not a simple offset. The contour lines rotate and bend across the grid, which means that tilt and curvature change at the same time. The sign can reverse from toe to heel, so a correction chosen to reduce a toe deviation can increase a heel deviation if the wrong order is used. Because the concave and convex flanks are generated by different cutter edges and respond differently, both flanks must be checked after each adjustment.
Sensitivity coefficients and the shape modes they represent
To separate the topology into interpretable modes, the normal deviation \(Z\) over the grid is approximated by a quadratic surface:
\[
Z=a_1X+a_2Y+a_3XY+a_4X^2+a_5Y^2
\]
where \(X\) is the lengthwise coordinate, \(Y\) is the profile-height coordinate, and the origin is at the grid center. The coefficients have direct geometric meanings: \(a_1\) is the spiral-angle error sensitivity, \(a_2\) the pressure-angle error sensitivity, \(a_3\) the tooth-surface twist sensitivity, \(a_4\) the lengthwise-curvature sensitivity, and \(a_5\) the profile-curvature sensitivity. The first two terms describe tilt; the last three describe bending and twist.
For the 45 grid points, the error equations form an overdetermined system
\[
\{\mathbf Z\}=[\mathbf S]\{\mathbf a\}
\]
where each row of \([\mathbf S]\) contains \(X_i\), \(Y_i\), \(X_iY_i\), \(X_i^2\), and \(Y_i^2\). The least-squares solution is
\[
\{\mathbf a\}=\left([\mathbf S]^{\mathsf T}[\mathbf S]\right)^{-1}[\mathbf S]^{\mathsf T}\{\mathbf Z\}
\]
The fitted coefficients for the pinion are listed in the table. They are per-unit changes in the corresponding correction coefficient and are used here for relative magnitude and sign; a combined change can be approximated by superposition within the small-correction range.
| Flank | Correction | \(a_1\) | \(a_2\) | \(a_3\) | \(a_4\) | \(a_5\) |
|---|---|---|---|---|---|---|
| Concave | \(K_2\) | −0.001459 | −0.000085 | 0.030401 | −0.032993 | 0.001665 |
| Concave | \(K_3\) | 0.000669 | −0.000398 | −0.024025 | 0.026346 | −0.000377 |
| Concave | \(K_4\) | −0.000855 | 0.000232 | 0.023961 | −0.025666 | 0.000699 |
| Concave | \(K_5\) | 0.000385 | −0.000179 | −0.012491 | 0.013369 | −0.000246 |
| Concave | \(K_6\) | −0.000227 | 0.000088 | 0.006956 | −0.007385 | 0.000159 |
| Convex | \(K_2\) | 0.000101 | 0.001265 | 0.023800 | −0.024071 | −0.001349 |
| Convex | \(K_3\) | 0.000372 | 0.000912 | 0.009852 | −0.010666 | −0.001245 |
| Convex | \(K_4\) | 0.000245 | 0.000797 | 0.011255 | −0.011419 | −0.001034 |
| Convex | \(K_5\) | 0.000167 | 0.000395 | 0.000395 | −0.004645 | −0.000555 |
| Convex | \(K_6\) | 0.000075 | 0.000192 | 0.002383 | −0.002416 | −0.000265 |
The second-order coefficient has the largest overall effect. On the concave flank, the fitted curvature terms are \(|a_3|\approx 0.0304\) and \(|a_4|\approx 0.0330\); on the convex flank they are \(|a_3|\approx 0.0238\) and \(|a_4|\approx 0.0241\). Third- and fourth-order terms are somewhat smaller but still comparable. Fifth- and sixth-order terms fall further, which is why they are better suited to fine local trimming than to removing a large global deviation.
For the concave flank, \(a_3\) and \(a_4\) are the two largest terms for every order, and \(a_5\) is the smallest. For the convex flank, \(a_4\) is consistently large; \(a_3\) is close to it for \(K_2\) through \(K_4\) and for \(K_6\), while the fifth-order twist term is weaker. The tilt terms \(a_1\) and \(a_2\) are generally smaller on the concave flank. On the convex flank, the pressure-angle term is not negligible: for \(K_2\), \(a_2=0.001265\), about fifteen times the magnitude of the concave value. A convex-flank pressure-angle error is therefore more likely to respond to roll correction than the same error on the concave flank.
The sign pattern is also informative. On the concave flank, \(a_4\) alternates with order: negative for \(K_2\), positive for \(K_3\), negative for \(K_4\), positive for \(K_5\), and negative for \(K_6\). On the convex flank, \(a_4\) remains negative for all five orders. This difference is the mathematical counterpart of the alternating toe behaviour seen in the topology plots. Because the five modes are not orthogonal, the coefficients should be read together: a large \(a_3\) with a small \(a_4\) indicates twist, a large \(a_4\) indicates lengthwise bending, and a dominant \(a_5\) indicates profile curvature. If the measured error is mostly \(a_5\), roll correction is a weak tool; a cutter-profile or pressure-angle adjustment is more direct.
Experimental check and shop-floor use
Grinding trials were run on an H350GH CNC spiral bevel gear grinder using the same pinion data. The second-order coefficient was changed by \(+0.02\), and the resulting flank was measured on a gear measuring center over the 9 × 5 grid. The measured topology reproduced the predicted shape: the same sign of deviation at the toe and heel and comparable point errors. The residual differences were at the level expected from random and measurement errors. Because only the \(K_2=+0.02\) case was reported in detail, the experimental check is strongest for the second-order mode; it confirms the sign and shape trend for that case rather than every high-order coefficient.
The practical sequence is to measure the flank, fit the five coefficients, and identify the dominant mode before changing the machine. If the topology is dominated by lengthwise curvature and twist near the toe, the second-order correction is a reasonable first trial. Third- and fourth-order terms can refine the distribution and sign; fifth- and sixth-order terms are fine trims for local end effects. If the error is predominantly pressure angle on the concave flank or profile curvature, roll correction has limited leverage and other machine or cutter adjustments should be considered.
The numerical values in the tables are specific to the 17-tooth, 12 mm face-width, 3 mm module example. The transferable result is the mode hierarchy: high-order roll correction acts mainly on twist and lengthwise curvature, concentrates near the tooth tip and ends, and produces opposite responses on the concave and convex flanks. The model is kinematic and assumes rigid cutter and workpiece; machine geometric errors, thermal distortion, and cutting dynamics remain separate error sources. Within those limits, the sensitivity coefficients provide a systematic way to choose the correction order and to reduce trial cuts during flank adjustment.