Study on Roughness Parameters

Yang Duo
State Key Laboratory of High Performance
Complex Manufacturing,
Screening and Characterizing
Central South University,
Changsha, Hunan 410083, China Surface Contact Performance
e-mail: [email protected]

Tang Jinyuan1
Based on Sensitivity Analysis
State Key Laboratory of High Performance
As microtopography can influence the contact behavior of materials, it is of great signifi-

Downloaded from http://asmedigitalcollection.asme.org/tribology/article-pdf/144/4/041502/6732952/trib_144_4_041502.pdf by University Of Leeds user on 05 September 2021

Complex Manufacturing,
Central South University, cance to study the correlation between morphology characterization parameters and
Changsha, Hunan 410083, China contact performance. In the light of complex relevance of parameters, a method for screen-
e-mail: [email protected] ing roughness parameters (RP) to characterize contact performance is constructed to get
the maximum influence parameters on the contact stress (CS) and avoid the error of expe-
riential selection. First, Pearson’s coefficient and back propagation (BP) neural network
Zhou Wei are utilized to elaborate on correlation level between RP and CS and to build the regression
Hunan Provincial Key Laboratory of High Efficiency
model. Then global sensitivity analysis (Sobol) and local sensitivity analysis (MIV and
and Precision Machining of Difficult-to-Cut
Garson) are introduced to demonstrate RP quantitative influences on CS and select main
Material,
RP for characterizing contact performance. The research shows (1) in the correlation anal-
Hunan University of Science and Technology,
ysis, RP with high correlation and noncollinearity on σmax are Sa, Sdq, S5p, Spk, and Svk;
Xiangtan, Hunan 411201, China
With regard to Mpmax and τmax, Sa, S5p, Sdq, and Vmp are on display, (2) RP importance
e-mail: [email protected]
sequence based on the results of correlation analysis is Sa, Spk, Sdq, Svk, S5p for σmax, and
Sa, Vmp, Sdq, S5p for Mpmax and τmax, and (3) For the comprehensive main parameters
Wen Yuqin model, RP for characterizing contact performance under the three contact stresses are
State Key Laboratory of High Performance
Sa, Spk, and Vmp, belonging to height parameter, function parameter, and volume param-
Complex Manufacturing,
eter, respectively. According to definition, all of them can significantly affect the stress con-
Central South University,
centration and distribution on contact surface of materials, which validates the rationality
Changsha, Hunan 410083, China
of the method. [DOI: 10.1115/1.4051733]
e-mail: [email protected]
Keywords: surface roughness, backward propagation, sensitivity analysis, correlation
analysis

1 Introduction forward the three-dimensional roughness parameters in the

roughness standard draft ISO/TC213N756 and divided them into
Since there is no absolutely smooth surface in engineering prac-
five categories: height parameters, volume parameters, functional
tice, factors such as machine tool vibration and tool friction will ser-
parameters, mixed parameters, and spatial parameters, with a total
iously affect the surface quality of components and make the
of 23 parameters. Finally, ISO25178 product geometry technical
surface morphology present different characteristics. The surface
specification standard was planted in 2010 [7,8].
morphology is directly related to wear resistance, tightness, and
With the increase of the number and types of three-dimensional
fatigue contact resistance of the components [1,2].
parameters, it is easier to obtain more complete and accurate infor-
In order to describe the surface topography characteristics more
mation about surface topography. The engineering application
accurately, the German, G. Schmalz, made the first quantitative
and correlation of different three-dimensional surface roughness
measurement of surface topography in 1929 and proposed the rele-
parameters were widely discussed and analyzed. For example, He
vant surface evaluation parameters [3]. Subsequently, different
et al. studied the correlation between the critical load and three-
parameters standards for two-dimensional surface morphology
dimensional roughness parameters Spc, Sq, Vvc, and Sdq in micro-
were formulated in various countries around the world, and rough-
connected structures and endowed the main performance of different
ness parameter Ra or Rq was generally used to characterize and
roughness parameters [9]. Draganovská et al. used Pearson correla-
evaluate the two-dimensional topography [4–6]. However, the two-
tion coefficient to analyze the correlation between two-dimensional
dimensional surface characterization parameters based on contour
roughness parameters [10]. Qi et al. established different levels
features cannot fully reflect all the surface topography characteris-
of “parameter tree” models among three-dimensional roughness
tics. As a result, the surface has an approximate value Ra, but
parameters, so as to observe the correlation between parameters
actual morphology is significantly different, which makes the
[11]. Sedlač ek et al. focused on the correlation between roughness
surface unable to meet the high-precision processing requirements
parameters such as Ssk and Sku and contact surfaces friction proper-
in modern industry.
ties on the premise of the fact that the values of Sa (Ra) were basi-
Considering the loss of surface information in the two-
cally the same [12,13].
dimensional morphology characterization, surface morphology
In spite of the correlation between roughness parameters and
can be better characterized by the three-dimensional roughness
material properties getting comprehended, researches have mainly
parameters based on the topography midplane. The Technical Com-
focused on the influence of Sa (Ra), leaving the separate influence
mittee of the International Organization for Standards (ISO) put
of different roughness parameters on contact performance still
unclear. Meanwhile, the complexity of the correlation between
1
roughness parameters has added difficulties to related studies.
Corresponding author.
Contributed by the Tribology Division of ASME for publication in the JOURNAL OF
Moreover, the selection method of roughness parameters from
TRIBOLOGY. Manuscript received May 1, 2021; final manuscript received June 30, 2021; experience to characterize the surface properties makes it impossi-
published online July 27, 2021. Assoc. Editor: Yong Hoon Jang. ble to clarify the rationality. Therefore, it is of importance to find

Journal of Tribology Copyright © 2021 by ASME APRIL 2022, Vol. 144 / 041502-1
out which roughness parameters are the main factors affecting the
surface contact and characterization of the contact performance
and to know the correlation between the roughness parameters.
In the first place, correlation between roughness parameters and
contact stress is investigated with the help of Pearson correlation
analysis. Having established the regression model between CS
and RP based on BP neural network, global sensitivity analysis
Sobol, mean impact value (MIV) and Garson neural network sensi-
tivity analysis are introduced for the quantitative calculation of RP
impact factor on CS to get the main roughness parameters under dif-
ferent contact stress parameters. Ultimately, the ranking results of
main roughness parameters based on influence under different
contact stresses are verified, and the characterization parameters

Downloaded from http://asmedigitalcollection.asme.org/tribology/article-pdf/144/4/041502/6732952/trib_144_4_041502.pdf by University Of Leeds user on 05 September 2021

are obtained. The research reveals the correlation between surface
Fig. 1 Topography measurement
microstructure characteristics and surface contact, which provides
a basis for the design and manufacture of anti-fatigue contact.
2.2 Contact Stress Calculation. Considering that roughness
surface contact stress field at the micromorphology scale cannot
2 Data Source be directly measured by experimental instruments, researchers gen-
erally use finite element and numerical algorithms to analyze the
2.1 Roughness Parameters Measurement and Calculation. contact state of rough surfaces and calculate the surface contact
First, multiple sets of rough surfaces were obtained from the ultra- stress. The three-dimensional rough surface modeling and contact
sonic vibration grinding test device. The material surface property analysis method of ellipsoidal asperities proposed by Wen et al.
and experiment conditions are shown in Tables 1 and 2. in the literature [15,16] solves the problem of low solution speed
The white light interferometer Wyko NT9100, shown in Fig. 1, of finite element and numerical algorithms and overcomes the diffi-
was used to measure the surface morphology of the workpieces. culty in convergence of iteration, making it a leading rough surface
To reduce the data error caused by topography measurement, ten contact algorithm. Therefore, Wen’s model is used to calculate the
sampling areas were set on the surface to gather the height matrix surface contact stress in the paper. The model is briefly described as
z(i,j) with the unit of µm. The height matrix of sampling area was follows.
obtained by ConfoMap Surface Analysis software. The percentage The watershed algorithm is used to divide the peaks and valleys
of unmeasured points was 15. The sampling interval was 0.99 μm, of the topography to obtain different asperities in Wen’s model, and
and the total area was An = Lx × Ly = 0.4 × 0.56 = 0.224 mm2. A total then through numerical fitting and the principle of the minimum
of 228 sets of samples were obtained. root mean square error, the most suitable ellipsoid equation for
With surface original information processed in Wen’s rough the surface asperities is determined. The fitting asperity schematic
surface contact model [15,16] and very little white noise in the mea- diagram is shown in Fig. 2.
surement, it did not get filtered during the data sampling phase for
the sake of ensuring true surface initial information. Roughness
parameters were calculated from the initial information to represent Table 3 Parameters
surface characteristics as realistically as possible. Then based on
Classification Symbol Definition
ISO 25178-2:2012 international evaluation standard [14] for three-
dimensional roughness parameters, height parameters, volume Height parameters Sa Arithmetical mean height
parameters, functional parameters, spatial parameters, etc. were Sz Maximum height
obtained. The definition of parameters is displayed in Table 3 Sq Root-mean-square height
with RP formula and description given in literature [14]. Ssk Skewness
Sku Kurtosis
Sp Maximum peak height
Table 1 Surface mechanical property Sv Maximum pit depth
Hybrid parameters Sdq Root-mean-square gradient
Name Value Sdr Developed interfacial area ratio
Elastic modulus 209 GPa Feature parameters Spd Density of peaks
Poisson’s ratio 0.29 Spc Arithmetic mean peak curvature
Yield strength 850 MPa S5p Five-point peak height
Hardness 2.38 GPa S5v Five-point pit height
S10z Ten-point height of surface
Functional parameters Sk Core height
Table 2 Grinding experimental conditions Spk Reduced peak height
Svk Reduced valley height
Smr1 Material ratio in peak
Name Value
Smr2 Material ratio in valley
Sxp Peak extreme height
Grinding wheel CBN grinding wheel (120 mesh)
Wheel radius 100 mm Volume parameters Vmp Peak material volume
Material 12CrNi4A (9 × 9 × 16 mm) Vmc Core material volume
Wheel speed 1500 rpm Vvc Core void volume
Cutting speed 200 mm/min Vvv Dale void volume
Cutting depth 5–25 µm
Spatial parameters Sal Autocorrelation length
Grinding way Plane grinding
Str Texture aspect ratio
Cooling fluid 120 l/min (Castrol Syntilo 2000)
Vibration feed direction Transverse Contact stress parameters σmax Maximum mises stress
Vibration amplitude 0–10 µm τmax Maximum shear stress
Vibration frequency 20 kHz Mpmax Maximum surface pressure

041502-2 / Vol. 144, APRIL 2022 Transactions of the ASME

 Downloaded from http://asmedigitalcollection.asme.org/tribology/article-pdf/144/4/041502/6732952/trib_144_4_041502.pdf by University Of Leeds user on 05 September 2021
Fig. 4 Schematic diagram of surface stress calculation

 
  
P (1 − 2ν) z xy xyz 3xyz
τxy = 1 − − − (4)
2π r2 ρ r 2 ρ3 ρ5

3P xz2
τxz = − (5)
2π ρ5

3P yz2
τyz = − (6)
2π ρ5
where ρ = (x2 + y2 + z2 )1/2 and r 2 = x 2 + y 2.
In the case of a given load of 1 N, this paper uses the model of
Wen et al. to calculate the CS of the rough surface. Figure 5
shows the local topography and stress cloud diagram of the
Fig. 2 Single-asperity fitting sketch workpiece.

After the rough surface morphology is fitted by the asperity,

according to the normal deformation ω of the asperity and John-
son’s contact mechanics theory [17], the contact pressure p and
contact stress including σmax, τmax, and Mpmax of different asperities
can be calculated iteratively under a given load. The deformation of
a single asperity is shown in Fig. 3. The surface contact stress cal-
culation diagram is shown in Fig. 4 and the formulae are as follows:
 
  
P (1 − 2ν) z x2 − y2 zy2 3zx2
σx = 1 − + − (1)
2π r2 ρ r2 ρ3 ρ5

 
  
P (1 − 2ν) z y2 − x2 zx2 3zy2
σy = 1 − + − (2)
2π r2 ρ r2 ρ3 ρ5

3P z3
σz = − (3)
2π ρ5

Fig. 5 (a) Topography and (b) stress cloud

Fig. 3 Deformation of single asperity

Journal of Tribology APRIL 2022, Vol. 144 / 041502-3

3 Principles for Screening “Main Roughness and the partial variance Di1 ,i2 ,...is of each subfunction is
Parameters” that Affect and Characterize Contact Stress 1
3.1 Back Propagation Neural Network. There is no tradi- Di1 ,i2 ,...is = fi1 ,i2 ,...is 2 (xi1 , xi2 , . . . , xis )dxi1 dxi2 . . . dxis (12)
0
tional display formula between RP and CS, but the relationship
between them is often nonlinear. BP neural network, composed Then Sobol sensitivity coefficient is defined as
of input layer, hidden layer, and output layer, is usually expert in
digging out multivariable nonlinear relationship with error back Di1 ,i2 ,...is
Si1 ,i2 ,...is = (13)
propagation algorithm [18–20]. D
The way of hidden layer neuron dealing with the input signal is as Due to the properties that the sum of all sensitivity coefficients is
follows: 1 [25,26]

   
·
n n n n
Hj = f 1 1
w( j,i) xi (7) Si + Sij + · · · + S12···n = 1 (14)

Downloaded from http://asmedigitalcollection.asme.org/tribology/article-pdf/144/4/041502/6732952/trib_144_4_041502.pdf by University Of Leeds user on 05 September 2021

i=0 i=1 i=1 j=1
where xi represents the ith input layer node (x0 is the input layer i≠j
bias, generally assigned as 1); w1(j,i) is the weight value of the ith
where sensitivity coefficient Si is the first-order impact factor of xi
input layer node and the jth hidden layer node; f1(X ) is the activa- (excluding the interaction of variables). To better evaluate the com-
tion function of the hidden layer; Hj is the output value of the jth
prehensive impact of xi on the model, STi is proposed as the overall
hidden layer node.
impact factor
In the same way, there exists the following relationship between
the hidden layer and the output layer: STi = 1 − S∼i (15)


·
m
where S∼i is expressed as the sum of all Si1 ,i2 ,...is without xi.
yp = f 2 w2(p,j) Hj (8)
j=0
3.2.2 Disturbance Sensitivity Analysis Method: Mean Impact
where yp is the pth node in the output layer.
Value. Mean impact value method, is able to quantitatively
Usually, the MSE is applied to evaluate the fitting effect of the
evaluate the influence of variables in the neural network on the
model:
results by variables disturbance [31–33]. The calculation process
 2 is as follows:
k
yp − yt
MSE = (9) Set X as the l × k-order input matrix of the initial training sample
p=1
2k of BP neural network:
⎡ ⎤
where k is the number of nodes in the output layer and yt is the true x11 x12 . . . x1k
value of the sample. ⎢ x21 x22 . . . x2k ⎥
To further improve the simplification and robustness of BP neural X=⎢ ⎣··· ··· ··· ···⎦
⎥ (16)
network structure, the additional momentum method [21,22] and xl1 xl2 . . . xlk
correlation pruning algorithm [23–25] are provided in the paper
to strengthen the stability and robustness. where k is the number of input variables, and l is the number of
initial training samples. With a variable xm (1 ≤ m ≤ k) arbitrarily
3.2 Sensitivity Analysis. Although the relationship between selected, two new input matrices Xm+0.1 and Xm−0.1 are generated
roughness parameters is complex, the sensitivity analysis method by increasing or decreasing 10% of column xm :
can quantitatively calculate the influence degree of a single rough- ⎡ ⎤
x11 x12 . . . x1m (1 + 10%) . . . x1k
ness parameter on the contact stress parameters, which provides a ⎢ x21 x22 . . . x2m (1 + 10%) . . . x2k ⎥
basis for screening “main roughness parameters” that affect and Xm+0.1 = ⎢⎣ ... ... ...
⎥ (17)
characterize contact stress. ... ... ...⎦
xl1 xl2 . . . xlm (1 + 10%) . . . xlk
3.2.1 Global Sensitivity Analysis Method: Sobol. Sobol sensi- ⎡ ⎤
tivity analysis, based on mathematical statistics [26,27], is also x11 x12 ... x1m (1 − 10%) . . . x1k
known as variance decomposition method proposed in 1993 and ⎢ x21 x22 ... x2m (1 − 10%) . . . x2k ⎥
Xm−0.1 ⎢
=⎣ ⎥ (18)
applied in multiple fields [28–30]. The core of the method is to ... ... ... ... ... ...⎦
decompose the model into functions that combine different vari- xl1 xl2 ... xlm (1 − 10%) . . . xlk
ables. It can well explain the influence degree of a single indepen-
dent variable on the dependent variable in the case of multiple Input the two new samples Xm+0.1 and Xm−0.1 into the completed
independent variables. The calculation process is as follows: training BP neural network to obtain two target responses Ym+0.1
First, the model is decomposed into the sum of subfunctions of and Ym−0.1 . Then the MIV value of xm can be obtained:
different dimensions:
  1 l
f (x) = f0 + fi (xi ) + fij (xi , xj ) + · · · MIVxm = (Y +0.1 (i) − Ym−0.1 (i)), m = 1, 2, · · ·k (19)
i 1≤s<j≤n
l i=1 m
+ f12−n (x1 , x2 , · · · , xn ) (10) And then the average influence value of each input variable on the
output variable can be obtained from formula (19).
Second, the ratio of the partial variance of the subfunction to the
total variance is used to indicate the influence of the independent
variable on the dependent variable of the model. The total variance 3.2.3 Weight Judgment Method: Garson. Garson sensitivity
D of the dependent variable f (x) is analysis is a method based on connection weights of network,
which uses the product of weights to calculate the input variables
D= f 2 (x)dx − f02 (11) influence or relative contribution on output variables [34–36].
In The formula for calculating the influence factors of xi on yk is as

041502-4 / Vol. 144, APRIL 2022 Transactions of the ASME

follows:



L 
N
wij v jk wij
j=1 i=1
Iik = 

N 
 L 
N (20)
wij v jk wij
i=1 j=1 i=1
(i = 1, . . . , N; j = 1, . . . , L; k = 1, . . . , M)
Here, i, j, k represents the node of input layer, hidden layer, and
output layer, respectively; N, L, M is the number of nodes in the
model; wij is the weight between the input layer and the hidden
layer; v jk is the weight between the hidden layer and the output

Downloaded from http://asmedigitalcollection.asme.org/tribology/article-pdf/144/4/041502/6732952/trib_144_4_041502.pdf by University Of Leeds user on 05 September 2021

layer. The same goes for the following formula.
However, due to the existence of positive and negative values of
wij and v jk in formula (20), the influence of xi on yk will be wea-
kened. To avoid this problem, Goh [37] proposed an improved
and widely used Garson formula. This paper adopts the improved
Garson formula:


L   N 
 
wij v jk  wij 
j=1 i=1
Iik = 

N 
 L   N 
  (21)
wij v jk   wij 
i=1 j=1 i=1
(i = 1, . . . , N; j = 1, . . . , L; k = 1, . . . , M)

4 Method Design and Results Fig. 6 Method flow

4.1 Process of Parameter Relationship Analysis. For
purpose of digging out the “main parameters” affecting different
contact stress parameters from dozens of roughness parameters to 4.2 Pearson Correlation Analysis. The correlation coeffi-
characterize the contact performance, the following parameter cor- cient results of three-dimensional roughness parameters and
relation analysis method is developed in this paper: contact stress parameters are shown in Fig. 7. It can be seen from
the figure that the parameter that has the highest correlation
(1) Analyze the correlation degree between parameters based on with the three contact stress parameters is Sa, which is consistent
Pearson correlation coefficient. Calculate correlation coeffi- with the practical engineering experience.
cient rij , rik of roughness parameters xi , xj and contact Based on the screening criteria described above, select the RP
stress parameters yk , select roughness parameters xi that whose correlation coefficients with contact stress is greater than
satisfy rij ≤ 0.7 (nonstrong correlation) and rik ≥ 0.3 0.3, and sort the parameters to construct the ranking parameter set
(nonweak correlation) to ensure that RP initially screened X{x1, x2, …, xn}according to the absolute value of correlation
have no collinearity and are highly correlated on contact
stress parameters [38]. Meanwhile, the data dimension is
reduced to facilitate the calculation of BP model.
(2) Take σmax, τmax, and Mpmax as separate output responses
(without considering the correlation between stress parame-
ters). Use RP initially screened in (1) and contact stress
parameters to build three BP neural networks. Compare the
parameters ranking results of Sobol, MIV, and Garson sensi-
tivity analysis to verify the parameters order under different
sensitivity analysis methods.
(3) Select the top two roughness parameters under different
contact stress parameters in (2) as the “main roughness
parameters,” and compare the “main roughness parame-
ters–contact stress parameters” BP model and (2) “RP for
preliminary screening–contact stress parameters” BP model
relative error to verify the selection results of “main rough-
ness parameters” and provide a basis for removing “redun-
dant” roughness parameters.
(4) Synthesize the “main roughness parameters” selected in (3),
construct a “comprehensive main roughness parameters–
contact stress parameters” BP model, and use the sensitivity
analysis method to obtain the parameter ranking under σmax,
τmax, and Mpmax. Then comprehensive main roughness Fig. 7 Correlation coefficient between roughness parameters
parameters order under different contact stress parameters and contact stress parameters. (1) The difference of coefficients
is compared. The analysis process of roughness parameters in the figure represents the correlation coefficient and (2) correla-
and contact stress parameters is shown in Fig. 6. tion coefficient was treated with absolute value.

Journal of Tribology APRIL 2022, Vol. 144 / 041502-5

 Table 4 Screened parameters based on correlation Table 5 BP model structure in correlation analysis

CS RP CS Initial structure Pruning

σmax Sa, Sdq, S5p, Spk, Svk σmax 5-10-1 5-9-1

τmax Sa, S5p, Sdq, Vmp τmax 4-10-1 4-7-1
Mpmax Sa, S5p, Sdq, Vmp Mpmax 4-10-1 4-9-1

coefficient. So that the parameters introduced are highly correlated Thanks to the excellent results of BP model, sensitivity analysis
with the contact stress. Then take the ranked first parameter x1 as the can be used to calculate the sensitivity coefficients of preliminary
benchmark, and sequentially filter other parameters whose correla- screened RP and assign their order. It can be seen from Fig. 9:
tion coefficient is less than 0.7 in set X to ensure that the parameters
(1) For the order of the roughness parameters of σmax from large
are relatively independent. In the last place, repeat the above steps

Downloaded from http://asmedigitalcollection.asme.org/tribology/article-pdf/144/4/041502/6732952/trib_144_4_041502.pdf by University Of Leeds user on 05 September 2021

to small, the results of three sensitivity analysis methods are
to complete the screening of all the morphology parameters, and
consistent: Sa, Spk, Sdq, Svk, S5p; As for τmax and Mpmax, it
select the benchmark parameters to be the target.
is Sa, Vmp, Sdq, S5p. The order of roughness parameters is
RP for preliminary screening about σmax, τmax, and Mpmax are
not affected by the sensitivity analysis method, which inter-
shown in Table 4.
actively verify the reliability of ranking results. But Miv and
After correlation screening, RP with high correlation on σmax and
Garson methods are more distinguishable than Sobol
no collinearity between variables are Sa, Sdq, S5p, Spk, and Svk,
method.
which are derived from height parameters, hybrid parameters,
(2) From the observation of CS, in the parameter ranking of σmax,
feature parameters, and functional parameters. The above results
the relative values of Sdq, Svk and S5p in three sensitivity
indicate that the extreme characteristics of the surface amplitude
analysis methods are all less than 0.2, which is far less than
and the steepness of the surface inclination, as well as the height
the relative values of Sa and Spk. For τmax and Mpmax, Sdq
of the peaks and valleys of the surface, have the greatest correlation
and S5p in different sensitivity analysis methods are basically
with σmax, significantly affecting its value, while other surface prop-
less than 0.2, which is also less than the sum of roughness
erties have little effect on σmax. As for τmax and Mpmax, Sa, S5p,
parameter coefficients ranking in the first two places.
Sdq, and Vmp are the target parameters, belonging to height param-
(3) In the sensitivity analysis of “RP for preliminary screening–
eters, feature parameters, hybrid parameters, and volume parame-
contact stress parameters” BP model, the average relative
ters. Comparing the results of σmax, it can be found that the peak
value of Spk in three sensitivity coefficients of σmax is
volume has a higher correlation with τmax and Mpmax.
0.25, and Vmp in τmax and in Mpmax is 0.22 and 0.21, respec-
The screened parameters under different stress parameters are
tively. It demonstrates that the influence of Spk on σmax is
partly the same, which shows that some roughness parameters
slightly stronger than the influence of Vmp on τmax and
have a universal effect on contact performance, but they will be dif-
Mpmax.
ferent when evaluating different aspects of the contact stress.
(4) Based on the order of RP under different contact stress param-
eters, Sa always plays the maximum influence parameter on
4.3 Screening Parameters. Because of unit differences surface contact, Sdq is in the third place, and S5p is located
between roughness parameters and contact stress parameters, data at the bottom. Combining the analysis of parameter sensitivity
standardization is essential to improve BP network training effect. coefficient in (2), it is easy to find out that the sum of Sa and
The normalization formula is as follows: the next parameter sensitivity coefficient is significantly

 greater than other parameters, indicating that the overall influ-
Xi − Xmin ence of the roughness parameters ranking in the first two
Xinew = (22)
Xmax − Xmin places is greater than that of other parameters. Therefore,
the roughness parameters ranked in the top two positions
where Xi is unnormalized data, Xmin is the minimum value of Xi, can be regarded as the “main roughness parameters.”
Xmax represents the maximum value, and Xnew i is normalized data.
The initial hidden layer nodes of “RP for preliminary screening–
contact stress parameters” BP model, “main roughness parame- 4.3.2 Main Roughness Parameters–Contact Stress Parameters.
ters–contact stress parameters” BP model and “comprehensive To verify the rationality of the “main roughness parameters”
main roughness parameters–contact stress parameters” BP model
are set based on the node selection method in Refs. [39,40].
Target error is 0.015, and the maximum iterations is 5000. When
replacing different contact stress parameters as output variables in
the BP model, after random interference with the sample order,
about 70% of the samples are treated as training samples and the
rest is taken as test samples to prevent errors caused by sample
allocation.

4.3.1 Roughness Parameters for Preliminary Screening–

Contact Stress Parameters. With the help of Pearson correlation
analysis, the preliminary screened RP corresponding to different
contact stress parameters are obtained. The BP model structure of
σmax, τmax, and Mpmax is shown in Table 5. It gives the number
of nodes in the input layer, hidden layer and output layer,
respectively.
The training set of the model converges to the target error value
within iterations, which indicates the model fitting effect is desir-
able. Then in order to test the generalization performance of the
model, the relative error of test samples is calculated. The result,
shown in Box Fig. 8, is basically within 0.1. Fig. 8 Relative error of preliminary screened RP

041502-6 / Vol. 144, APRIL 2022 Transactions of the ASME

 Table 6 BP model structure in main roughness parameters

CS Initial structure Pruning

σmax 2-5-1 2-3-1

τmax 2-5-1 2-3-1
Mpmax 2-5-1 2-3-1

with RP for preliminary screening BP model. The results prove that

the removed roughness parameters do have a relatively low impact
on the contact stress parameters, which verifies the rationality of the
selection of the “main roughness parameters.”

Downloaded from http://asmedigitalcollection.asme.org/tribology/article-pdf/144/4/041502/6732952/trib_144_4_041502.pdf by University Of Leeds user on 05 September 2021

Meanwhile, for the sake of comparing the consistency of the
common parameters in the “RP for preliminary screening BP
model” and the “main roughness parameters BP model,” sensitivity
analysis is carried out in “main roughness parameters BP model”
again.
The order of main roughness parameters on σmax is Sa > Spk. For
τmax and Mpmax, the order is Sa > Vmp. The order of parameters in
the two models is completely consistent, which verifies the rational-
ity of the main roughness parameters selection.

4.3.3 Comprehensive Main Roughness Parameters–Contact

Stress Parameters. Although the main parameters of different
contact stress parameters have been obtained, there are differences
in the parameters. For example, the main roughness parameters of
σmax are Sa and Spk, and Sa and Vmp for τmax and Mpmax. So it
is impossible to know the influence of Vmp on σmax and the influ-
ence of Spk on τmax and Mpmax.
In order to study the influence of all the main roughness param-
eters on the contact stress parameters, BP network and sensitivity
analysis method are used to rank all the main roughness parameters
to clarify the influence on the contact stress parameters. There are 7
nodes in the initial hidden layer, and the model structure is shown in
Table 7.
Table 7 shows that the BP models of the three contact stress param-
eters before and after pruning are all consistent, and the fitting
and generalization performance of the model are shown in Fig. 11.
As is demonstrated in Fig. 11, the models show good fitting effect
and data generalization performance. Based on the trained model,
all roughness parameters are subjected to sensitivity analysis, and
the order of comprehensive main roughness parameters is shown
in Fig. 12.
The above research shows:
(1) Regarding the influence of roughness parameters on different
contact stress parameters, the order of the comprehensive
roughness parameters is: Sa, Spk, Vmp, verifying the ratio-
nality of the ordering. Comparing the sensitivity coefficients
of the roughness parameters of σmax, τmax, and Mpmax, the
relative values of Sa in different sensitivity analysis
methods are all around 0.5, while Vmp and Spk are both
around 0.25. The influence of Sa on surface contact can be
Fig. 9 Sensitivity analysis of roughness parameters. Sensitivity
coefficient is dealt with relative value. (a) σmax (b) τmax, and regarded as twice that of Vmp and Spk.
(c) Mpmax. (2) In the sensitivity analysis of the “comprehensive main
parameters model,” the sensitivity coefficients of Spk to
σmax, τmax, and Mpmax are 0.27, 0.27, and 0.23, respectively;
the sensitivity coefficients of Vmp to σmax, τmax, and Mpmax
proposed in (1), the analysis is carried out again with the help of BP are 0.24, 0.25, and 0.22, respectively. Therefore, the average
network to observe the fitting effect and generalization performance value of Spk about contact stress parameters is 0.256, and the
of “main roughness parameters–contact stress parameters” BP value is 0.236 for Vmp. In conclusion, Spk has a slightly
model. Set the number of nodes of the initial hidden layer as 5, stronger influence on surface contact than Vmp, which is
and the main roughness parameter BP network structure is shown consistent with the sensitivity analysis results of “RP for pre-
in Table 6. liminary screening–contact stress parameters” BP model.
The main roughness parameters BP model still quickly converges (3) From the perspective of the three contact stress parameters,
to the target error. Its generalization performance is shown in Box Sa, Spk, and Vmp are the parameters with the greatest influ-
Fig. 10. Figure 10 shows that when only two “main roughness ence among many roughness parameters. Sa, the average
parameters” are introduced, the fitting and generalization perfor- height of the rough surface, basically reflects the overall
mance of “main roughness parameters” BP are basically consistent characteristics of surface quality and micromorphology,

Journal of Tribology APRIL 2022, Vol. 144 / 041502-7

 Downloaded from http://asmedigitalcollection.asme.org/tribology/article-pdf/144/4/041502/6732952/trib_144_4_041502.pdf by University Of Leeds user on 05 September 2021
Fig. 10 Relative error of main roughness parameters

Table 7 BP model structure in comprehensive main parameters

CS Initial structure Pruning

σmax 3-7-1 3-3-1

τmax 3-7-1 3-3-1
Mpmax 3-7-1 3-3-1

Fig. 12 Sensitivity analysis of comprehensive main parameters.

Fig. 11 Relative error of comprehensive main roughness Sensitivity coefficient is dealt with relative value. (a) σmax (b) τmax,
parameters and (c) Mpmax.

5 Conclusion
and significantly affects surface contact stress. Spk is the
average height of the peak above the core surface, character- (1) With the help of nonstrong and nonweak correlation con-
izing peaks protrusion, which will significantly affect the cepts, preliminary screened parameters with high correlation
stress concentration in contact. Vmp, actually the ratio of with σmax but no collinearity among RP in correlation analy-
the peak top material to the entire peak, is the support sis are Sa, Sdq, S5p, Spk, and Svk. As for the result of τmax
volume of the peak region, which can significantly affect and Mpmax, Sa, S5p, Sdq, and Vmp are on display. The
the stress concentration and contact stress value in contact. research shows that preliminary screened parameters do not
Although the above parameters come from different param- only cover the information about surface amplitude, statisti-
eter classification, they actually include surface peak infor- cal distribution, and functional characteristics, but also
mation, consistent with the cognition of literature [41]. It greatly reduce the number of original parameters that charac-
shows that the peak feature has a stronger influence on the terize contact performance, which provides a theoretical
contact stress than other features of the morphology, which basis for selecting the main roughness parameters that
can well characterize the surface contact. affect surface contact state.

041502-8 / Vol. 144, APRIL 2022 Transactions of the ASME

 (2) On the premise of not considering the correlation between [13] Sedlač ek, M., Podgornik, B., and Viž intin, J., 2009, “Influence of Surface
contact stress parameters in comprehensive main roughness Preparation on Roughness Parameters, Friction and Wear,” Wear, 266(3–4),
pp. 482–487.
parameters BP model, the coefficient of Sa in different sen- [14] International Organization for Standardization, 2012, “ISO 25178-2: 2012:
sitivity analysis methods is around 0.5, and the value is Geometrical Product Specifications (GPS)-Surface Texture: Areal–Part 2:
around 0.25 for Vmp and Spk. The average influence of Terms, Definitions and Surface Texture Parameters,” https://www.iso.org/
Spk in different contact stress parameters is 0.256, and standard/42785.html
[15] Wen, Y., Tang, J., Zhou, W., and Zhu, C., 2018, “A New Elliptical Microcontact
the influence value of Vmp is 0.236. Sensitivity analysis Model Considering Elastoplastic Deformation,” Proc. Inst. Mech. Eng. Part J:
quantifies the influence of the screened roughness parame- J. Eng. Tribol., 232(11), pp. 1352–1364.
ters in correlation analysis on the contact stress, and the [16] Wen, Y., Tang, J., Zhou, W., and Li, L., 2020, “A Reconstruction and Contact
sensitivity coefficients are all positive, belonging to positive Analysis Method of Three-Dimensional Rough Surface Based on Ellipsoidal
Asperity,” ASME J. Tribol., 142(4), pp. 1271–1283.
influence. Therefore, Sa, Vmp, and Spk should be reduced [17] Johnson, K. L., 1985, Contact Mechanics, Cambridge University Press, London.
in the actual machining to provide the surface bearing [18] Muñoz-Escalona, P., and Maropoulos, P. G., 2010, “Artificial Neural Networks
capacity. for Surface Roughness Prediction When Face Milling Al 7075-T7351,”

Downloaded from http://asmedigitalcollection.asme.org/tribology/article-pdf/144/4/041502/6732952/trib_144_4_041502.pdf by University Of Leeds user on 05 September 2021

(3) Combined with correlation analysis and sensitivity analysis, J. Mater. Eng. Perform., 19(2), pp. 185–193.
[19] Rumelhart, D. E., Hinton, G. E., and Williams, R. J., 1986,
the screened roughness parameters with maximum influence “Learning Representations by Back-Propagating Errors,” Nature, 323(6088),
on different contact stress parameters to characterize surface pp. 533–536.
contact performance are Sa, Spk, and Vmp, respectively [20] Lipiń ski, D., and Bałasz, B., 2018, “Modelling of Surface Roughness and
attached to height parameters, functional parameters, and Grinding Forces Using Artificial Neural Networks with Assessment of the
Ability to Data Generalization,” Int. J. Adv. Manuf. Technol., 94(1–4),
volume parameters. Based on the definition of roughness pp. 1335–1347.
parameters, what the three parameters have in common is [21] Li, J., Cheng, J., Shi, J., and Huang, F., 2012, “Brief Introduction of Back
that they can significantly affect the stress concentration Propagation (BP) Neural Network Algorithm and its Improvement,” Advances
and distribution state of the contact surface and reflect the in Computer Science and Information Engineering. Advances in Intelligent and
Soft Computing, D. Jin, and S. Lin, eds., vol 169, Springer, Berlin/Heidelberg,
characteristics of the peak. The conclusion proves the ratio- pp. 553–558.
nality of the method used to screen the roughness parame- [22] Zhang, X., and Gao, Y., 2011, “Based on Momentum Method BP Neural
ters that characterize the surface contact. Therefore, the Network,” Proceedings of the SPIE 8193, International Symposium on
method can be used to guide the selection of main charac- Photoelectronic Detection and Imaging 2011: Advances in Infrared Imaging
and Applications, Beijing, China, Sept. 8, p. 81932N.
terization parameters in actual processing to measure [23] Castellano, G., Fanelli A, M., and Pelillo, M., 1997, “An Iterative Pruning
surface quality. Algorithm for Feedforward Neural Networks,” IEEE Trans. Neural Networks,
8(3), pp. 519–531.
[24] Wang, J., Liu, L., and Pan, X., 2020, “Pruning Algorithm of Convolutional Neural
Network Based on Optimal Threshold,” Proceedings of the 2020 5th International
Acknowledgment Conference on Mathematics and Artificial Intelligence, New York, pp. 50–54.
[25] Li, X., and Li, X., 2013, “An Improved Correlation Pruning Algorithm for
The author(s) disclosed receipt of the following financial Artificial Neural Network,” Electr. Des. Eng., 21(8), pp. 65–67.
support for the research, authorship, and/or publication of this [26] Sobol, I. M., 2001, “Global Sensitivity Indices for Nonlinear Mathematical
article: National Key R&D Program of China (Grant No. Models and Their Monte Carlo Estimates,” Math. Comput. Simul., 55(1–3),
pp. 271–280.
2018YFB2001300), National Natural Science Foundation of [27] Sobol, I., 1993, “Sensitivity Estimates for Nonlinear Mathematical Models,”
China (NSFC) through Grant No. 51705142. Math. Modell. Comput. Exp., 1(4), pp. 407–414.
[28] Zadeh, F. K., Nossent, J., Sarrazin, F., Pianosi, F., van Griensven, A., Wagener,
T., and Bauwens, W., 2017, “Comparison of Variance-Based and
Moment-Independent Global Sensitivity Analysis Approaches by Application
to the SWAT Model,” Environ. Modell. Softw., 91, pp. 210–222.
Conflict of Interest [29] Sobol’, I. M., Asotsky, D., Kreinin, A., and Kucherenko, S., 2011, “Construction
There are no conflicts of interest. and Comparison of High-Dimensional Sobol’generators,” Wilmott, 2011(56),
pp. 64–79.
[30] Ballester-Ripoll, R., Paredes, E. G., and Pajarola, R., 2019, “Sobol Tensor Trains
for Global Sensitivity Analysis,” Reliab. Eng. Syst. Saf., 183, pp. 311–322.
[31] Tian, P., Zhan, G. F., and Nai, L., 2015, “A Variable Selection for Asphalt
References Pavement Performance Based on RBF Neural Network,” Appl. Mech. Mater.,
[1] Whitehouse, D. J., 2010, Handbook of Surface and Nanometrology, CRC Press, 3843(1490), pp. 1222–1225.
London. [32] Zhang, X., Chen, L., Sun, Y., Bai, Y., Huang, Y., and Chen, K., 2018,
[2] Shi, Z., Jiang, H., and Zhang, M., 2016, “Multi-Scale Representation for Skin “Determination of Zinc Oxide Content of Mineral Medicine Calamine Using
Models in Geometrical Product Specifications,” Opt. Precis. Eng., 24(7), Near-Infrared Spectroscopy Based on MIV and BP-ANN Algorithm,”
pp. 1647–1654. Spectrochim. Acta, Part A, 193, pp. 133–140.
[3] Li, C., and Dong, S., 2000, “The Trends of Characterizing 3D Surface [33] Dombi, G. W., Nandi, P., Saxe, J. M., Ledgerwood, A. M., and Lucas, C. E.,
Microtopography,” China Mech. Eng., 11(5), pp. 16–20. 1995, “Prediction of Rib Fracture Injury Outcome by an Artificial Neural
[4] Han, X., Chen, X., Yang, X., Bai, H., Li, Z., and Su, X., 2006, “Application,” Network,” J. Trauma Acute Care Surg., 39(5), pp. 915–921.
Proceedings of the IEEE, Zhuhai, China, Jan. 18–21, pp. 131–135. [34] Ghanizadeh, A. R., Heidarabadizadeh, N., and Jalali, F., 2020, “Artificial Neural
[5] Temsamani, A. B., Vandenplas, S., and van Biesen, L., 2001, “Optimal Model for Network Back-Calculation of Flexible Pavements With Sensitivity Analysis
the Diffraction Effect in the Ultrasonic Field of Piston Transducers,” J. Comput. Using Garson’s and Connection Weights Algorithms,” Innov. Infrastruct.
Acoust., 9(2), pp. 461–476. Solut., 5(2), pp. 1–19.
[6] Li, B., and Liu, Y., 2004, “Study of Developing Tendency of Surface [35] Maozhun, S., and Ji, L., 2017, “Improved Garson Algorithm Based on Neural
Roughness,” Tool Eng., 38(1), pp. 63–67. Network Model,” 29th Chinese Control and Decision Conference (CCDC),
[7] Mainsah, E., Greenwood, J. A., and Chetwynd, D. G., 2013, Metrology and IEEE, Chongqing, China, May 28–30, pp. 4307–4312.
Properties of Engineering Surfaces, Springer, Berlin. [36] Garson, D. G., 1991, “Interpreting Neural Network Connection Weights,” AI
[8] Leach, R., 2013, Characterisation of Areal Surface Texture, Springer, Berlin. Expert, 6(4), pp. 47–51.
[9] He, B., Petzing, J., Webb, P., and Leach, R., 2015, “Improving Copper Plating [37] Goh, A. T., 1995, “Back-Propagation Neural Networks for Modeling Complex
Adhesion on Glass Using Laser Machining Techniques and Areal Surface Systems,” Artif. Intell. Eng., 9(3), pp. 143–151.
Texture Parameters,” Opt. Lasers Eng., 75, pp. 39–47. [38] Schober, P., Boer, C., and Schwarte, L. A., 2018, “Correlation Coefficients:
[10] Draganovská, D., Iž aríková, G., Brezinová, J, and Guzanová, A., 2015, “The Appropriate Use and Interpretation,” Anesth. Analg., 126(5), pp. 1763–1768.
Study of Parameters of Surface Roughness by the Correlation Analysis,” Mater. [39] Stathakis, D., 2009, “How Many Hidden Layers and Nodes?,” Int. J. Remote
Sci. Forum, 818(15), pp. 15–18. Sens., 30(8), pp. 2133–2147.
[11] Qi, Q., Li, T., Scott P, J., and Jiang, X., 2014, “A Correlational Study of Areal [40] Sheela, K. G., and Deepa, S. N., 2013, “Review on Methods to Fix
Surface Texture Parameters on Some Typical Machined Surfaces,” 13th CIRP Number of Hidden Neurons in Neural Networks,” Math. Probl. Eng., 2013, pp.
Conference on Computer Aided Tolerancing: CAT 2014, Hangzhou, China, 1–11. .
May 11–14. [41] Dong, W. P., Sullivan, P. J., and Stout, K. J., 1994, “Comprehensive Study of
[12] Sedlač ek, M., Podgornik, B., and Viž intin, J., 2012, “Correlation Between Parameters for Characterising Three-Dimensional Surface Topography: III:
Standard Roughness Parameters Skewness and Kurtosis and Tribological Parameters for Characterising Amplitude and Some Functional Properties,”
Behaviour of Contact Surfaces,” Tribol. Int., 48, pp. 102–112. Wear, 178(1–2), pp. 29–43.

Journal of Tribology APRIL 2022, Vol. 144 / 041502-9