Why Grinding Residual Stress Is a System-Level Problem
Spiral bevel gears are used where a transmission must turn a corner under high torque: automotive axles and differentials, aircraft and helicopter drives, wind-turbine gearboxes, and heavy mobile machinery. Their curved teeth engage gradually, so load is shared over a contact zone that moves across the flank rather than applied at a single instant. This geometry gives high load capacity, smooth motion, and low noise, but it also makes the final manufacturing operation unusually important. Grinding is normally the last hard-finishing step; it determines tooth geometry, surface topography, and the near-surface stress state that the gear carries for its entire service life.
Residual stress is the self-equilibrated stress that remains after the grinding force and thermal field are removed. It is not a surface curiosity. It shifts the mean stress experienced by the tooth flank, influences crack initiation and propagation, and can distort the tooth after the part is released from the fixture. Compressive stress is generally beneficial for fatigue resistance; tensile stress is generally detrimental. The practical problem is therefore not simply whether the surface is compressive or tensile, but where the peak lies, how deep the field extends, and how steep the gradient is through the subsurface.
Grinding Kinematics and Contact Mechanisms
Spiral bevel gear grinding follows either a generating or a form-copying principle. In generating grinding, the wheel represents a tooth of a virtual generating gear—often described as an imaginary crown gear—and the wheel and workpiece maintain a timed conjugate motion. The instantaneous contact line sweeps across the tooth space, and successive positions envelope the flank. In form grinding, the wheel profile corresponds to the required tooth space and is fed along the tooth length; there is no generating roll, and the wheel shape is reproduced directly in the workpiece.
The generating wheel may be planar or conical. A planar generating wheel has a face-cone angle of 90°, with its wheel axis parallel to the machine cradle axis. A conical generating wheel has a face-cone angle other than 90° and nonparallel axes. The distinction matters because it changes the contact geometry, the effective wheel speed distribution, and the local cooling conditions along the flank.
Three grinding configurations are used in practice, and each produces a different thermo-mechanical contact state:
- Straight cup wheel grinding. The wheel section resembles the cutter blade of a face-milling head, and the grinding face behaves like the generating tooth surface. It is geometrically suited to pitch-cone angles up to about 50°. In form grinding of large wheels, however, the wheel can contact the tooth space across the full face width. This raises the normal force, restricts coolant access, and creates a burn risk near the root transition.
- Flaring cup wheel grinding. A wheel with a 30° cone angle is used so that the curvature of the outer and inner surfaces differs from that of the concave and convex flanks. The result is approximately single-point contact rather than full-width contact. Coolant can reach the contact zone, the instantaneous thermal load is lower, and the multi-axis machine path can sweep the wheel through the tooth space.
- Wguri eccentric-spindle grinding. The wheel spindle is mounted with an eccentricity of about 0.1 mm. Rotation about the eccentric axis creates an intermittent or clearance-type contact with the concave and convex flanks as the wheel is fed along the tooth length. A standard cylindrical wheel can replace a special flaring cup wheel, but the machine requires a precision eccentric spindle mechanism.
At the grit scale, material removal passes through three overlapping stages. In rubbing, the abrasive grain contacts the surface elastically and generates friction and heat with little removal. In plowing, the grain penetrates and pushes material to the sides, producing plastic flow, ridges, and work hardening. In cutting, the depth exceeds the critical threshold and a chip forms by shear. The balance among these stages controls the force ratio, the heat partition, the surface finish, and the residual stress profile.
Thermo-Mechanical Source Model and Finite-Element Representation
The grinding force per unit width is commonly separated into rubbing, plowing, and chip-formation components. The normal component is usually larger than the tangential component because the abrasive grain has an effective negative rake:
$$F’_t=F’_{ts}+F’_{tp}+F’_{tch},\qquad F’_n=F’_{ns}+F’_{np}+F’_{nch}$$
The heat generated at the contact is treated as a moving band source on a semi-infinite body. The heat flux entering the workpiece is estimated from the tangential force and the relative sliding speed:
$$q=R_w\,\frac{F_t\,(v_s\pm v_w)}{b\,l_k}$$
where \(R_w\) is the fraction of heat entering the workpiece, \(v_s\) is the wheel speed, \(v_w\) is the feed speed, \(b\) is the contact width, and \(l_k\) is the contact-arc length. The sign is set by the grinding direction: down grinding and up grinding distribute the relative velocity differently. The partition ratio is not a constant; it depends on the thermal properties of the abrasive, the workpiece, the coolant, and the effective number of active grains.
For the workpiece, AISI 4340 is a representative high-strength gear steel: yield strength about 792 MPa, tensile strength about 980 MPa, elastic modulus about 200 GPa, Poisson ratio 0.29, and thermal conductivity about 38 W/m·°C. The large strain, high strain rate, and steep temperature rise in grinding are usually described with the Johnson–Cook constitutive model:
$$\sigma=\left(A+B\varepsilon^n\right)\left(1+C\ln\frac{\dot\varepsilon}{\dot\varepsilon_0}\right)\left[1-\left(\frac{T-T_0}{T_m-T_0}\right)^m\right]$$
For AISI 4340, the parameters used in the analyzed case were \(A=792\) MPa, \(B=510\) MPa, \(n=0.26\), \(C=0.014\), and \(m=1.03\). Failure was represented with the Johnson–Cook damage criterion, in which the damage parameter accumulates as the ratio of incremental equivalent plastic strain to the failure strain.
The finite-element case examined a 44-tooth large wheel with module 3 mm, 20° pressure angle, 35° mean spiral angle, left-hand helix, 132 mm pitch diameter, 20.2 mm face width, and a 90° shaft angle. The single-grit model simplified the abrasive grain as a rigid 90° cone, 0.07 mm high, with a 0.01 mm edge radius. The workpiece was meshed with hexahedral elements and treated as an elastic–plastic body. The grinding step used an explicit solver, followed by a cooling step in a standard implicit solver. The workpiece base was fixed, the friction coefficient was 0.3, and the stress component in the grinding direction was extracted and averaged over 98 nodes at each depth layer to a depth of about 0.32 mm.
A multi-grit model was built with a virtual-lattice method. The wheel surface was divided into cells, one grit was assigned per cell, and the grit center was allowed a random offset within its cell. This produced a distribution that is macroscopically uniform but microscopically random—closer to a dressed wheel than a regular array, while remaining computationally tractable.
Residual Stress Profile and Parameter Sensitivity
The computed profiles have a consistent architecture. The as-ground surface is in compression. The maximum compressive stress lies below the surface, in the subsurface layer, and the field crosses into tension at greater depth before approaching zero. In the analyzed cases, the transition was essentially complete by about 0.3 mm. This shape reflects two competing effects: the mechanical compression and constrained plastic deformation near the surface, and the thermal gradient and elastic recovery that develop as the heat pulse decays.
Across the 16-run orthogonal grinding array, the maximum compressive stress ranged from −235.9 to −363.6 MPa, and the maximum tensile stress ranged from 9.1 to 26.7 MPa. Compressive stresses are reported with a negative sign; tensile stresses are positive.
| Run | Wheel speed \(v_s\) (m/s) | Depth of cut \(a_p\) (mm) | Feed \(v_w\) (m/min) | Maximum compressive \(S_1\) (MPa) | Maximum tensile \(S_2\) (MPa) |
|---|---|---|---|---|---|
| 1 | 10 | 0.01 | 3.0 | −235.882 | 12.381 |
| 2 | 10 | 0.02 | 3.5 | −266.544 | 12.668 |
| 3 | 10 | 0.03 | 4.0 | −295.398 | 14.212 |
| 4 | 10 | 0.04 | 4.5 | −363.551 | 26.726 |
| 5 | 15 | 0.01 | 3.5 | −249.112 | 12.162 |
| 6 | 15 | 0.02 | 3.0 | −285.248 | 11.913 |
| 7 | 15 | 0.03 | 4.5 | −332.917 | 22.652 |
| 8 | 15 | 0.04 | 4.0 | −356.948 | 20.861 |
| 9 | 20 | 0.01 | 4.0 | −287.359 | 11.619 |
| 10 | 20 | 0.02 | 4.5 | −321.676 | 18.303 |
| 11 | 20 | 0.03 | 3.0 | −335.188 | 17.054 |
| 12 | 20 | 0.04 | 3.5 | −357.614 | 20.602 |
| 13 | 25 | 0.01 | 4.5 | −279.252 | 9.098 |
| 14 | 25 | 0.02 | 4.0 | −347.571 | 22.155 |
| 15 | 25 | 0.03 | 3.5 | −347.897 | 23.767 |
| 16 | 25 | 0.04 | 3.0 | −342.336 | 20.512 |
The parameter ranking is consistent across the single-factor simulations and the statistical analysis. Depth of cut has the strongest influence, wheel speed is next, and feed is comparatively weak. Increasing depth from 0.01 to 0.04 mm enlarges the contact area and the plastic zone, raises the normal force, and drives the peak compression deeper. Increasing wheel speed raises the number of active grains per unit time; the chip thickness per grain falls, but the heat input rises, so the stress field intensifies. Increasing feed raises the removal rate and the thermal gradient, but it also shortens the contact time, so the net effect is smaller.
Multi-grit grinding produces the same qualitative profile as single-grit grinding, but a lower peak compressive stress. Neighboring grains do not act independently: the compressive field of one grain overlaps with the tensile field of an adjacent grain, and part of the compression is canceled. This is an important scaling result. It means that a single-grit model can predict the shape and trend of the depth profile, but it can overestimate the peak compression unless the overlap statistics of the real wheel are represented.
Multi-Objective Parameter Selection
Parameter selection was formulated as a constrained two-objective problem: maximize the compressive magnitude and minimize the tensile stress. A three-factor, four-level orthogonal array supplied 16 simulation points. Grey relational analysis converted the two responses into a single grade. Each response was normalized in the direction of its target—compressive magnitude larger-is-better, tensile stress smaller-is-better—and grey relational coefficients were calculated with a distinguishing coefficient of 0.5. The grey relational grade is the average of the two coefficients; a higher grade indicates a better combined response.
The mean grade by parameter level was:
- Wheel speed: 0.5776, 0.5890, 0.6117, 0.6248 at 10, 15, 20, and 25 m/s; range 0.0472.
- Depth of cut: 0.6033, 0.5764, 0.5728, 0.6504 at 0.01, 0.02, 0.03, and 0.04 mm; range 0.0776.
- Feed: 0.5842, 0.5920, 0.6109, 0.6159 at 3.0, 3.5, 4.0, and 4.5 m/min; range 0.0317.
The range ranking is depth > speed > feed, which agrees with the single-factor simulations. The depth effect was not monotonic: the mean grade fell slightly from 0.01 to 0.03 mm and then rose sharply at 0.04 mm, indicating interaction between depth and the other factors. The highest grey relational grade occurred at 25 m/s, 0.04 mm, and 4.5 m/min.
Response-surface regressions were then fitted to the two stress responses and used as objective functions inside a particle-swarm optimization (PSO) routine. The PSO used a population of 100, 500 iterations, a linearly decreasing inertia weight from 0.9 to 0.4, and learning factors of 2. The regression models had coefficients of determination of 97.11% for compressive stress and 97.18% for tensile stress, with adjusted values of 94.58% and 93.95%. The optimization produced five non-dominated candidates. An entropy-weighting model was used to score the grey relational solution and the five Pareto candidates.
| Candidate | \(v_s\) (m/s) | \(a_p\) (mm) | \(v_w\) (m/min) | Predicted \(S_1\) (MPa) | Predicted \(S_2\) (MPa) | Composite score | Rank |
|---|---|---|---|---|---|---|---|
| Grey relational | 25.0000 | 0.0400 | 4.5000 | −411.047 | 38.596 | 0.899 | 1 |
| PSO 1 | 22.5517 | 0.0395 | 3.4265 | −357.301 | 22.335 | 0.143 | 2 |
| PSO 2 | 21.4123 | 0.0400 | 3.4488 | −355.888 | 21.087 | 0.128 | 4 |
| PSO 3 | 21.7891 | 0.0400 | 3.4307 | −356.042 | 21.359 | 0.129 | 3 |
| PSO 4 | 20.7137 | 0.0400 | 3.4436 | −353.826 | 20.091 | 0.101 | 6 |
| PSO 5 | 22.0855 | 0.0400 | 3.3742 | −354.199 | 21.033 | 0.102 | 5 |
The entropy-weighted ranking favored the grey relational point because its predicted compressive magnitude was substantially larger—about 411 MPa versus 354–357 MPa for the Pareto candidates. The trade-off is equally important: that point also carried the highest predicted tensile stress, 38.6 MPa versus 20.1–22.3 MPa. The selection is therefore a weighted decision, not a strict Pareto dominance result. If a design rule places more emphasis on avoiding tensile stress, a candidate near 20.7 m/s, 0.04 mm, and 3.44 m/min would be a reasonable alternative. The experimentally validated parameter set was the grey relational point: 25 m/s, 0.04 mm, and 4.5 m/min.
Experimental Validation and Error Structure
Validation was performed on an H650GA eight-axis, five-link CNC spiral bevel gear grinding machine with an eccentric grinding spindle. The ground wheel was wire-cut into single-tooth specimens at room temperature with coolant, and the cutting line was placed far from the tooth surface to limit cutting-induced stress. Residual stress was measured by X-ray diffraction with an Xstress 3000G3 analyzer. Three locations on the tooth flank were measured, and electrolytic polishing was used to remove thin layers for the depth profile.
| Quantity | Finite-element simulation | Experiment 1 | Experiment 2 | Experiment 3 |
|---|---|---|---|---|
| Maximum compressive \(S_1\) (MPa) | −396.5 | −450.2 | −463.8 | −454.5 |
| Maximum tensile \(S_2\) (MPa) | 36.9 | 42.9 | 43.4 | 42.4 |
The measured and simulated depth profiles followed the same trend: surface compression, a subsurface compression peak, a transition to tension, and a return toward zero at greater depth. The absolute magnitudes differed systematically. Relative to the regression prediction, the finite-element results differed by 3.67% for compressive stress and 4.52% for tensile stress. The experimental results differed from the prediction by 8.7–11.37% for compression and 8.94–11.16% for tension. Relative to the finite-element results, the experiments were 11.9–14.5% more compressive and 12.88–15.01% more tensile. All differences were below 16%.
The direction of the discrepancy is informative. The experiments were consistently more compressive and more tensile than the simulation, which suggests that the model captured the mechanism but under-represented the intensity of multi-grit interaction. The actual abrasive density on the dressed wheel may have been higher than the stochastic model, increasing the overlap of stress fields. The simulation also used idealized cooling, a fixed base, and a homogeneous workpiece, whereas the actual process includes temperature fluctuation, machine vibration, nonuniform coolant distribution, and prior manufacturing stresses. Phase transformation and initial heat-treatment residual stress were not included in the model.
Process Design Implications and Model Limits
For spiral bevel gear grinding, depth of cut is the primary residual-stress lever, wheel speed is secondary, and feed is comparatively weak. That ranking is useful for process window design, but it does not justify simply maximizing all three. A deeper cut raises the compressive peak and deepens the affected layer, yet it also raises the tensile stress at greater depth and increases the risk of burn, wheel loading, and thermal damage. The preferred target is a controlled profile—subsurface compression of sufficient magnitude, limited surface tension, and a gradual gradient—rather than a single peak value.
Wheel topography is another controllable variable that the single-grit model cannot fully represent. Grit density, protrusion, and dressing quality determine how many active cutting edges overlap in the contact zone. Since multi-grit overlap lowers the peak compression relative to a single grit, changes in wheel condition can shift the residual-stress profile even when the nominal grinding parameters are unchanged. Coolant delivery and the choice of straight cup, flaring cup, or eccentric-spindle grinding also alter the heat partition and the burn risk.
The quantitative evidence supports the use of finite-element simulation for trend analysis, parameter ranking, and relative comparison. It does not support treating the simulated absolute stress as a substitute for measurement. The validation covered one gear geometry, one workpiece material, one wheel condition, and three measurement locations; the errors are a measure of consistency within that window, not a general accuracy guarantee. Coupling phase transformation, measured wheel topography, and in-situ diffraction measurements would reduce the systematic gap between simulation and experiment. Until then, the most reliable process design combines simulation-based ranking with experimental calibration of the specific wheel–material–machine system.