Manufacturing Tolerances and

Axle System NVH Performance

Zhaohui Sun, Mark Ranek, Michael Voight and Glen Steyer
American Axle & Manufacturing, Inc., Detroit, Michigan

A study of axle system noise, vibration, and harshness

(NVH) performance using design for six sigma (DFSS) meth-
ods is presented with a focus on system robustness to typical
product variations (manufacturing-based tolerances). Instead
of using finite-element analysis (FEA) as the simulation tool,
a lumped-parameter system dynamics model developed in
M ATLAB/Simulink is used in the study. This provides an effi-
cient way to conduct large analytical design of experiment
(DOE) and stochastic studies. The model’s capability to pre-
dict both nominal and variance performance is validated with
vehicle test data using statistical hypothesis test methods.
Major driveline system variables that contribute to axle gear
noise are identified and their variation distributions in pro-
duction are obtained through sampling techniques. Through
analytical DOE and analysis of variance (ANOVA) analyses,
the critical design parameters that control system NVH varia-
tions with respect to product variations and variations due to
operating conditions are covered. Design criteria of axle sys-
tems with respect to robust NVH performance are also dis- Figure 1. FEA model of rear-beam axle driveline.
cussed.

Axle gear whine is a major NVH concern in vehicle driveline Model Prediction
Vehicle test, 1st gear mesh
systems. The noise source is typically the hypoid-gear-mesh, Vehicle test, 2nd gear mesh
first-harmonic transmission error, which further transforms
into dynamic gear mesh force under operation. The dynamic
characteristics of the mesh force are controlled by the gear train
dB re 1 m/s2

torsional dynamics. The driveline system translational dynam-

ics and the impedance characteristics at the axle interface with 20 dB
the vehicle are also important in defining the final vehicle in-
terior gear-mesh noise level.
Considerable progress has been achieved in recent years in
understanding the physics of axle system gear noise. Using
computer-aided engineering (CAE) tools such as FEA to ana-
lyze system dynamics in regard to gear noise have been found 1st mesh harmonics 2nd mesh harmonics
to be extremely successful. 1-5 In most applications, linear sys-
tem assumption is satisfied, and a modeling approach called 0 Frequency (Hz) 1000
the building-block method has proven to be efficient and ac- Figure 2. Example of FEA model correlation to test data.
curate. In dealing with gear mesh geometry, a simplified lin-
ear gear-mesh model developed by Steyer et al. 4,5 is widely be effectively used to achieve a viable nominal design. A good
used. Figure 1 shows a typical rear-beam axle driveline FEA example of parametric design are the results presented by Sun
model and Figure 2 gives its accuracy with respect to vehicle et al., 1 where systems are desensitized to variations through
test data. enforcing NVH design specifications that are rolled down from
Despite the success of applying simulation tools in design/ analytical studies. Monte Carlo-type optimization methods for
tuning of system dynamics to achieve a quiet system for a par- parametric design are being developed.
ticular vehicle platform, it is even more important that the The tolerance design, on the other hand, deals with manu-
product has robustness in NVH performance in the following facturing process capability. This is to identify critical toler-
aspects: ances that contribute to system NVH performance variability
• The system meets the NVH requirements in the presence of and tighten specifications in an effective and economic fash-
‘variability’ due to both manufacturing process capability ion to minimize system variation. A tolerance optimization is
and environmental/operating conditions. most effectively done after a parameter optimization.
• The axle subsystem exhibits minimized system sensitivity Figure 3 illustrates the concept of both parametric and tol-
to different vehicle platforms. erance designs. Literature on parametric design studies and
In general, the robustness of a system can be established successful implementation can be found in recent years. How-
through two aspects of the design: parametric design and tol- ever, there are few studies on tolerance designs regarding axle
erance design. 7 The parametric design deals with nominal sys- system NVH. There are several reasons behind this:
tem design through design parameter settings to achieve re- • It requires an integrated analytical approach that combines
duced system sensitivity to variations. Current CAE tools can statistical and analytical DOE (design of experiment) tech-
niques with simulation tools. However, traditional simula-
Based on Paper #2005-01-2309, “Variation Reduction of Axle Sys-
tion tools such as FEA are still time consuming for this kind
tem NVH,” © 2005 SAE International, presented at the Noise & Vibra- of task. For instance, if one would conduct a three-level full
tion Conference, May 16-19, Traverse City, MI 2005. factorial DOE with six parameters, 729 runs are needed.

12 SOUND AND VIBRATION/APRIL 2006

 Before

Vehicle vibration
Axle
threshold for noise
vibration

Parametric


design
Contributes to
After customer
dissatisfaction

Tolerance
design

Figure 3. Axle subsystem robust design.

• It requires a tremendous effort to obtain the statistical data

Figure 4. Typical rear axle carrier assembly.
for the component-level parameters as inputs to the analyti-
cal models. In many cases, special tests have to be designed • Gear mesh line of action (LOA)
and conducted on large sample sizes to obtain meaningful • Gear pitch point location
results. • Effective prop torsional stiffness
The work here focuses on the tolerance design. To facilitate • Effective prop bending stiffness
the analytical study, a stand-alone, lumped-parameter axle • Gear mesh transmission error (MTE)
system dynamic model built in M ATLAB/Simulink is developed Notice from the above that the propeller shaft torsional and
and used instead of traditional FEA models. This model, as the bending dynamic stiffnesses are included in the study. This is
system solver, enables large analytical DOEs and statistical because the axle gear mesh dynamic force and system vibra-
analyses. HyperStudy from Altair Engineering, Inc. is used as tion responses are strongly controlled by the propeller shaft,
the analytical tool. Validation of the approach is presented and as previously discussed.1,2 Including the propeller shaft dy-
shows satisfactory accuracies in predicting both system nomi- namics in the analysis is essential in studying the interaction
nal and variance performance. between axle carrier assembly variations and the system gear-
Examples of probability distribution functions of modal pa- mesh dynamics. Modal models of the propeller shaft for bend-
rameters are also presented. In most of the cases, they are ob- ing and torsion are used in the lumped-parameter model, which
tained through testing on large populations of production is described further in the following section.
samples. Under certain circumstances, special analytical tools In a linear system as in this study, MTE is actually a linear
have to be used to obtain reasonable data on particular param- scaling factor. However, it is important to include it in the study
eters; for instance, the hypoid-gear-mesh, line-of-action varia- to evaluate its contributions to system NVH variation and com-
tion. parison with other parameters. The evaluated system response
The validated models are then used to conduct DOE studies in the study is the axle pinion nose vertical vibration, since it
and ANOVA analyses. Factors critical to quality are identified is the most appropriate indicator for axle gear mesh-induced
and discussed. pitch and bounce motion.

Scope of Study Analytical Approach

This study focuses on the axle carrier assembly (or third HyperStudy/M ATLAB Approach. To study the system varia-
member) regarding axle system gear noise variation reduction. tions, an analytical tool had to be used that is capable of doing
The axle carrier assembly includes the following major com- variation simulation (stochastic study), analytical DOE, and
ponents: axle carrier, differential assembly (with hypoid ring ANOVA analysis. Optimization capability is also a necessity.
gear bolted on the flange), pinion shaft and gear, pinion flange, Instead of developing our own analytical toolbox in perform-
differential bearings and pinion bearings (see Figure 4). ing this task, this paper uses HyperStudy provided by Altair,
For axle first-gear-mesh harmonic noise, the dynamics of dif- Inc., which is part of the HyperWorks package. Using this com-
ferential case assembly and axle carrier typically have little mercially available software enabled us to focus our effort and
effect and can be considered as lumped masses since their reso- time on problem solving instead of tool development.
nant frequencies are well above the gear-mesh frequency range Two options were evaluated regarding axle system gear-mesh
(typically 200-700Hz). Their mass and mass moment of iner- dynamics modeling: FEA and stand-alone, lumped-parameter
tia will affect the system gear-mesh force or vibration responses models. The later was chosen because, although FEA models
but are secondary compared with other variables. The axle are convenient and readily available, conducting analytical
carrier stiffness at bearing bores is important and can be in- DOE with these models would be time consuming. Stochastic
cluded into final effective bearing stiffness in supporting the studies with a large number of runs (say, >2000) would be im-
gearing system. The pinion shaft is relatively more compliant practical with FEA models.
in both torsional and lateral directions; typically it is modeled Instead, the authors decided to develop a new lumped pa-
with beam elements in FEA models. rameter model using M ATLAB /Simulink from The MathWorks,
With the knowledge base of the axle system gear noise Inc. The axle carrier assembly model (together with propeller
mechanism, the following lumped parameters will be studied shaft) is built in Simulink, while the model parameters and
in DOE and ANOVA analysis: intermediate variable calculations are given in the M ATLAB pro-
• Drive pinion tail bearing radial stiffness gram developed by the authors.
• Drive pinion head bearing radial stiffness The concept of the analytical approach to this study is shown
• Differential bearing radial and axial stiffness by the flow chart shown in Figure 5. HyperStudy is custom-

SOUND AND VIBRATION/APRIL 2006 13

ized to interface with the M ATLAB program. The next step is to acquire the test data for the parameters
M ATLAB/Simulink Program. The model is illustrated in Fig- of interest. For example, the bearing stiffnesses may be derived
ures 6 and 7, where the system is divided into subsystems and from appropriate frequency response function (FRF) measure-
each subsystem is represented by lumped parameters. That is, ments on the axle assembly. From the measurement of multiple
lumped masses are connected by springs and dampers with production assemblies the bearing stiffness PDFs can be deter-
space transformations of the motion. Modal models are used mined. Prop bending and torsional stiffnesses can also be de-
for the propeller shaft bending and torsional dynamics, and the rived from FRF measurements. PDFs can be estimated for all
modal parameters are determined in a separate M ATLAB pro- key NVH factors as (see Figure 8).
gram. In the case of characterizing the PDF for gearset MTE, it be-
The M ATLAB program was developed to perform the follow- comes a little more complicated. A sample of gearsets can be
ing tasks: read in the system configuration data; calculate and measured on a single flank tester at various buildable gearset
set the parameter values used in the Simulink model; and ob- pattern positions. These data combined with the percentages
tain the outputs from Simulink to conduct further mathemati- of axles built at each pattern position can yield an overall PDF
cal processing to obtain the system gear mesh force and vibra- of the “axle assembly” MTE delivered to the customer (see
tion responses. As illustrated in Figure 5, HyperStudy is Figure 9). A Weibull PDF typically fits this data best with the
interfaced with the M ATLAB program so that the model param- MTE values highly skewed to the low end.
eters can be easily replaced to conduct a large number of runs. PDFs from Analysis. Determining PDFs that are related to
gear-mesh geometry parameters is a challenging task. The pa-
Determination of Model Parameters rameters include gear mesh point location, which can be ex-
Determination of PDFs. To perform the variation simulations pressed as (xm , ym , zm) in a Cartesian coordinate system; and
of the parameters affecting NVH variation, the probability dis- the gear mesh line of action (LOA), which can be expressed as
tribution functions (PDFs) must be estimated for each of the (n x, ny , n z). Notice that there are only two independent param-
main parameters to be studied. The PDFs can be estimated by eters in the LOA vector. In defining the variation of the LOA
testing actual parts or purely analytically. The preference for vector, a variation cone surface is used, as illustrated in Fig-
this study has been to use the test method as much as possible. ure 10, where the apex angle α and cone angle γ are defined.
PDFs from Testing. The first step in establishing the test- If the vector ra is defined as being the rotation of nominal
derived PDFs is to obtain a representative sample of produc- LOA vector r n = (n x , ny , nz) around the global x-axis by apex
tion parts. Care must be taken to randomly select the samples angle a, then any vector rb on the cone surface defined by angle
over a long enough time and large enough group of production a can be expressed as: 8
machines to adequately represent typical manufacturing varia- rb = La ra = [I + (1 - cos g ) C2 + sin g C]ra (1)
tions. The size of the sample helps determine the confidence
intervals (error estimates) of the estimated ‘true’ population’s where I is the 3 ¥ 3 identity matrix, and
PDF parameters (µ,σ, etc.), with a larger sample size yielding
a more precise estimate. There are obvious trade-offs between Ê 0 - nz ny ˆ
sample size and its associated costs (parts, testing, etc.) and the Á ˜ (2)
C = Á nz 0 - nz ˜
precision of the estimate. ÁË - ny nx 0 ˜¯

Similarly, the relationship between rb and rn can be expres-

sed as:
Further postprocess
(Excel, MATLAB, Etc.) rb = (La Ln )rn (3)

where L n is the matrix defined similarly as L a for the

rotation from rn to r a .
HyperStudy
(DOE, stochastic study, optimization
Physically measuring the variations of the gear-mesh point
location and LOA vector is currently not available in practice.
Change values A test rig to conduct this kind of measurement is underway in
the author’s organization.
Modify Therefore, theoretical studies of the gear-mesh variation
Model MATLAB parameter Simulink through special gear analysis tools were conducted, and PDFs
Results
input file Read in program Run model
for mesh point (xm , ym , zm ) and LOA variables (α,γ) were ob-
simulation
tained through this study. The details of gear mesh geometry
Figure 5. Flow chart of analytical model. analytical studies are not be explicitly described here.

Mesh TE
In Output
Output to file

Pinion model

Carrier model
Differential / gear model
Mesh LOA
input

Pinion
noise
Output to file
Propshaft torsion

Propshaft bending

Figure 6. Axle carrier assembly system gear mesh model in MATLAB /Simulink.

14 SOUND AND VIBRATION/APRIL 2006

 or estimated so that the PDFs could be generated. The estimated
PDFs were inputs to the stochastic analysis module of
HyperStudy so that axle NVH performance variations could be
generated according to the PDFs. The process is illustrated in
Figure 12. Figure 13 shows a comparison of the actual experi-
mental variation compared to the analytically predicted varia-
tion. The data were reasonably Gaussian, so hypothesis tests
for ‘sigma’ were conducted at the same two key frequencies ref-
erenced above prop second bending resonance and the gear
mesh resonance at an α = 0.05 level of significance. As shown
in Figure 13 there is no statistically significant difference be-
tween the predicted model and experimental sigmas. Similar
results were obtained for the coast side, and consequently the
model was considered validated for predicting the variation in
vibration performance.

Figure 7. The carrier model in MATLAB /Simulink.

Analytical DOE
Time and other constraints necessitated a strategy of several
Model Validation exploratory screening DOEs to precede the main DOE. The
To obtain maximum confidence in model results, they should purpose of screening is to allow subsequent detailed analysis
be validated in regard to their ability to predict both system to focus on the most important factors and interactions.
mean and the variation of NVH performance. The model pre- Screening DOE. Figure 14 shows the main effects of a screen-
dictions were compared to experimental vehicle test results ing 12-factor, two-level DOE for the drive side. The two levels
and hypothetical tests were conducted to evaluate differences chosen represent the anticipated maximum and minimum or
in the mean and deviation. “tolerance range” of each of the factors. It is clear from Figure
Validation of Model’s Nominal Prediction. The model’s pre- 14 that the drive-side main effects are dominated by one fac-
diction of nominal axle vibration is compared to the experi- tor (Factor I). Coast-side results were similar.
mental average of 20 axles randomly selected from production. Figure 15 shows the typical low level of interactions that
Hypothesis tests for ‘means’ are conducted at two key frequen- were evident in the analysis. This is a desirable situation in a
cies; prop second bending resonance and the gear-mesh reso- driveline assembly, since it means that the system is predict-
nance at an α = 0.05 level of significance. As shown in Figure able and robust against normal manufacturing and product
11, there is no statistically significant difference between the variations.
predicted model and experimental means for the drive-side Three-Level, Full-Factorial DOE. Based on the screening
results. Similar results were obtained for the coast side, and DOE results, several significant factors were selected for fur-
consequently the model was validated for predicting nominal ther detailed study. A three-level DOE was conducted on the
vibration performance. six most significant factors and the main effects are shown in
Validation of Model’s Variation Prediction. The model’s Figure 16. Only minor nonlinear effects are noted. An evalua-
prediction of variation in axle vibration was compared to the tion of interactions in the three-level DOE again shows that they
experimental variation of 20 axles randomly selected from are of negligible magnitude. Typical three-level interactions are
production. The 20 axles had key NVH parameters measured shown in Figure 17.

0
Phase

-180

-360
100.00

Example of input PDF

10.00
derived from 20 randomly
selected axles
A/F (m/s2 / N)

1.000

0.1000

0.0100
300 1000 3000
Frequency (Hz)

Bearing stiffness PDFs derived

from drive pinion / flange FRFs

Input PDFs

Factors Distribution Mean Standard Dev Variance

Pinion tail bearing radial stiffness Normal 5.0e8 (N/m) 3.65E+07 1.33E+15
Pinion head bearing radial stiffness Normal 8.0e8 (N/m) 5.50E+07 3.03E+15
radial Normal 4.0e8 (N/m) 2.67E+07 7.13E+14
Diff bearing stiffness
axi al Normal 3.2e7 (N/m) 2.13E+06 4.54E+12
Mesh LOA angle change Normal 0 1.6667 2.78E+00
Mesh LOA angle cone Uniform 0-360 (degree)
Gearset TE Weibull Shape = 1.1245 Sc ale = 28.635
X Normal -85.289 ( mm) 1.73 2.99E+00
Pitch Point Y Normal -34.12 (mm) 0.60 3.60E-01
Z Normal 7.53 (mm) 1.57 2.46E+00
Test Meas urement Variation Normal 1 0.05 2.50E-01

Figure 8. Example of input PDF derived from FRFs of 20 randomly selected axles.

SOUND AND VIBRATION/APRIL 2006 15

 Position 1 Position 2
p = 0.29
A
p = 0.79

Acceleration, Log
Position 5
B A
Pinion nose vertical Experimental (µ = 3σ)
B
Drive Model (µ = 3σ)
Position 3 Position 4
Experimental (µ)
p = 0.24 p = 0.66
Model (µ)
TE 5 position
audit data 1000 1500 2000 2500 3000 3500
Propshaft RPM
Delivered TE PDFs
g
h tin Coast Figure 11. Hypothesis test results on model accuracy with respect to
eig
W system nominal and variations.
Assembled pattern position PDFs

Input PDFs

Model and hyperstudy

Pinion tall brg Mesh LOA Gearset MTE
Figure 9. Schematic diagram for obtaining PDFs of MTE. radial sitffness angle cone

Ra
Pinion head brg Mesh LOA Pinch point
Rb radial sitffness angle delta Y direction
(nx, ny, nz)

γ Diff brg Pinch point Measurement 100.00

radial sitffness Y direction variation Predicted NVH variation

z

Acceleration (m/s2)
10.00

Diff brg Pinch point 1.00

axial sitffness Z direction

y 0.10
1000 Propshaft RPM 4000
α
o
Figure 12. Flow chart of model prediction on system variations.
x

(a)

(xm, ym, zm)

Figure 10. Vector rotation on a cone surface.

Pinion Nose Vertical
Acceleration, LOG

It is shown from Figure 16 that Factors B and K have the most

significant contributions to the gear-mesh vibration variation
among the six factors. To better rank the sources of system NVH
variation, an ANOVA analysis was conducted.
ANOVA. An analysis of variance (ANOVA) was performed
to establish a rank of the sources of axle NVH variation. These
results are shown in Figure 18. It is shown that Factor I is re-
sponsible for the vast majority of normal product NVH varia-
tion (93% and 74% for drive and coast, respectively). The re-
sults also demonstrated that the rank of the factors and their 1000 Propshaft RPM 4000
contribution percentages are different under different vehicle/
(b)
axle load conditions. This is due to the change of axle gear
mesh kinematics and its interaction with system dynamics.
Therefore both vehicle drive and coast conditions should be
studied to capture all the significant factors.
These results clearly establish what factors should be the
primary focus of variation reduction efforts. Potential changes Pinion Nose Vertical
Acceleration, LOG

in factor tolerances can be confirmed or evaluated in a varia-

tion simulation by modifying a factor’s PDF to reflect the pro-
posed tolerance. Running the model multiple times with the
new tolerances and comparing the results with the original
tolerances will quantify the improvement in variation reduc-
tion. 500
Frequency (Hz)

Conclusions
It is possible to analytically estimate the variation in driv-
eline NVH performance due to typical product tolerances and 1000 Propshaft RPM 4000
manufacturing variations. The analytically derived variation
Figure 13. Model correlation to vehicle data related to system varia-
correlated well to the experimentally measured variations. tions. a) Experimental variation of 20 randomly selected production
Various stochastic studies, DOEs, etc., can be performed to try axles. b) Model predicted variation of 20 randomly selected simulated
to analyze and optimize product and manufacturing process axles (model runs).

16 SOUND AND VIBRATION/APRIL 2006

 Figure 16. Main effects from three-level, 10-factor DOE; Altair
hyperstudy analyses.
Figure 14. Main effects from two-level, 12-factor screening DOE;
driveside Altair hyperstudy analyses.

Figure 17. Typical variable interactions in three-level DOE.

Figure 15. Example of typical variable interactions.

tolerances to achieve the least NVH variation for given manu-

facturing and cost constraints.
A possible scenario for obtaining maximum benefit of these
new tools would be to conduct a parameter (nominal design)
optimization first to obtain several of the lowest nominal NVH
parameter combinations in the “design window.” Then each of
these ‘good’ candidate nominal designs can be evaluated for
tolerance sensitivity. The final design will be the best combi-
nation of low nominal NVH, low sensitivity to tolerances/
manufacturing variation, within existing cost/manufacturing/
etc. constraints.

Acknowledgments
The authors wish to thank American Axle & Manufacturing
in supporting this work. The authors would like to acknowl-
edge the support and contributions from our colleagues at
American Axle. Assistance from Altair, Inc. is also acknowl-
edged.

References
1. Sun, Z., Steyer, G., Meinhardt G., and Ranek, M.,”NVH Robustness
Design of Axle Systems,” SAE Noise and Vibration Conference,
03NVC-139, Traverse City, MI, 2003.
2. Sun, Z., Voight, M., and Steyer, G., “Driveshaft Design Guidelines
for Optimized Axle Gear Mesh NVH Performance,” FISITA 2004,
F2004V287, Barcelona, Spain, 2004.
3. Sun, Z., Steyer, G., and Ranek, M., “FEA Studies of Axle System
Dynamics,” SAE World Congress, 2002-01-1190, Detroit, MI, 2002.
4. Steyer, G., “Influence of Gear Train Dynamics on Gear Noise,” Pro-
ceedings of the National Conference on Noise Control Engineering,
The Penn State University, pp. 53-58, College Park, PA, 1987. Figure 18. Driveside (a) and coastside (b) results of ANOVA analysis;
5. Donley, M., Lim, T., and Steyer, G., “Dynamic Analysis of Automo- two-level, 10-factor Altair hyperstudy analysis.
tive Gearing Systems,” SAE Paper 920762, 1992.
6. Abe, T., Bonhard, B., and Cheng, M., et al., “High Frequency Gear 8. Litvin, F. L., Gear Geometry and Applied Theory, Chapter 1.3,
Whine Control by Driveshaft Design Optimisation,” SAE paper 2003- Prentice-Hall, 1994.
01-1478. 9. Thomson, W. T., Theory of Vibration with applications, Prentice-Hall,
7. Mahajan, S. K., Surella, M. M., and Single, T. C., “Design Six Sigma 1981.
Quality into a RWD IRS Driveline System for Improved Vehicle-Level
NVH performance,” SAE paper 2003-01-1494.
The author may be contacted at: glen.steyer@aam.com.

SOUND AND VIBRATION/APRIL 2006 17