Professional Documents
Culture Documents
Measurement
journal homepage: www.elsevier.com/locate/measurement
A R T I C L E I N F O A B S T R A C T
Keywords: Uncertainty evaluation is playing an increasingly important role in assessing the performance, safety and re-
Measurement uncertainty evaluation liability of complex physical systems in the absence of adequate amount of experimental data. This paper
Whiplash test presents a quantification of the measurement uncertainty in whiplash test models. We researched the analysis
Back propagation neural network techniques of uncertainty for the complex nonlinear systems, as well as the advantages and disadvantages of the
Least squares support vector machine
proposed methodology. By introducing the finite element analysis, we verified the consistency between the
Monte Carlo method
whiplash test, the calibration test and the simulation results of them. We also studied the influential factors and
their probability density functions and presented the sensitivity analysis of whiplash test model. Based on the
Latin hypercube sampling, we utilized the back propagation neural network (BPNN) and the least squares
support vector machine (LS-SVM) to establish the mathematical models. Furthermore, the accuracies of two
models are validated. Comparing with the results acquired by the guidance of uncertainty measurement and the
Bayesian method, we demonstrate that the LS-SVM-based Monte Carlo method is the most appropriate technique
for the evaluation of whiplash test uncertainty.
⁎
Corresponding authors.
E-mail addresses: shenlongwang@usst.edu.cn (S. Wang), dingxh@usst.edu.cn (X. Ding).
https://doi.org/10.1016/j.measurement.2018.01.065
Received 9 August 2017; Received in revised form 31 October 2017; Accepted 30 January 2018
Available online 01 February 2018
0263-2241/ © 2018 Elsevier Ltd. All rights reserved.
S. Wang et al. Measurement 119 (2018) 229–245
[11] expounded the Bayesian probabilistic approach to select the most The paper is organized as follows: Section 2 describes the whiplash
plausible class of dynamic models for a structural or mechanical system. test model, and Section 3 reviews the uncertainty quantification tech-
Higdon et al. [12] utilized a Bayesian formulation and Gaussian process nique. Section 4 reports the sensitivity analysis of inputs and the ver-
models to quantify uncertainties in both calibrations and predictions. ification of the consistence with FEA and physical test. Section 5 pre-
Jiang and Mahadevan [13] presented a Bayesian structural equation sents the mathematical modeling between inputs and outputs by NN
modeling technique to address uncertainty quantification in hier- and SVM. Section 6 provides the measurement uncertainty evaluation
archical model development. and compares the results of different approaches. Finally, Section 7
Considering the high costs of whiplash modeling, the prior sample concludes the paper.
becomes scarcity, i.e., complexity and computational resources for
implementing the Bayesian method increase rapidly. To circumvent
2. Whiplash test model
these disadvantages, scholars presented an evaluation approach con-
sidering the propagation of probability density functions via the Monte
In this paper, we carried out the whiplash test according to the C-
Carlo method (MCM). Advantages of this technique include its simpli-
NCAP (China-New Car Assessment Program) [35]. The seat and re-
city of implementation and explicitly of distribution function of the
straint systems are modeled on the original vehicle structure and fixed
output quantity. Doubilet et al. [14] proposed a MCM for the prob-
on the mobile sled. The sled is fired at a specific acceleration waveform
abilistic sensitivity analysis of uncertain data in medical decision ana-
and the low-speed rear-end collision is simulated. The BioRID II
lyses. Wübbeler et al. [15] illustrated the evaluation of measurement
Dummy is assembled on the seat, and then we measured the whiplash
uncertainties using the MCM and described its relation to the GUM, and
injuries during the rear-end collision to evaluate the protection effect of
Hölle et al. [16] conducted the similar researches by applying the MCM
head restraint on occupant’s neck. The whiplash test device is shown in
to the uncertainty quantification of pneumatic multihole probes. Pa-
Fig. 1.
lenčár et al. [17] employed the MCM for the determination of un-
Based on the C-NCAP, we set the parameters as follows: the time
certainties of temperature scale measurement by standard platinum
interval of the acceleration waveform is 0 ms to 150 ms, the velocity
resistance thermometer. Furthermore, the Bayesian approach based on
variation of the accelerating trolley is controlled at ΔV = 15.65 km/
Markov chain Monte Carlo (MCMC) sampler combining both the ac-
h ± 0.8 km/h, and the duration of the waveform is
curacy of Bayesian method and the effective sampling of Monte Carlo
ΔT = 91 ms ± 3 ms. The full score of the whiplash test is 4 points,
technique, has been applied extensively in uncertainty quantification of
which can be obtained from the injury index measured by the sensor
simulation predictions [18], phylogenetic trees [19], probabilistic
mounted on the dummy. Three groups of measurement data formed the
modeling [20],rail-sleeper-ballast system [21], identification of vibra-
score of whiplash test, including the neck injury criterion (NIC), the
tion sources [22] and many other fields.
load and torque of the upper neck and which of the lower neck. For
In this research, we analyzed the uncertainty in the measurement
each injury index, the high-performance limit and low-performance
and modeling of whiplash test. Taking into account the high costs of the
limit are set corresponding to the highest score and 0 points, respec-
human and financial resources, as well as the non-repeatability of the
tively. We utilized the linear interpolation method to evaluate the score
uncalibrated device in whiplash test, it is hard to gain plenty of his-
for case where the injury index is between high-performance limit and
torical data and sufficient samples. Namely, apart from the above stu-
low-performance limit. In addition, 2 points, 2 points and 4 points
dies, the traditional Bayesian approach and MCM are not applicable for
penalty are given for the situation that the maximum dynamic angle of
the uncertainty quantification. To resolve this problem, we combined a
the backrest, the head restraint interference space, the seat rail dynamic
Monte Carlo method with the finite element analysis (FEA), the ma-
displacement does not meet the requirements, respectively. Then the
chine learning [30–32] and the Latin hypercube sampling [33,34]. The
overall scoring principle of whiplash test is shown in Table 1.
procedures of the proposed techniques are:
(1) Replacing whiplash test by FEA after the verification of the con- 3. Measurement uncertainty evaluation techniques
sistence between them;
(2) The Latin hypercube sampling is utilized and the FEA is carried out Over the past years, scholars are devoted to develop the research
to obtain the inputs and outputs respectively; methods of measurement uncertainty and compiled the guidance to the
(3) We applied the neural network (NN) and support vector machine expression of uncertainty in measurement (GUM). According to the
(SVM) to build the mathematical model between the inputs and GUM, the uncertainty of the measurement results generally includes
outputs; several components, which can be classified into two categories based
(4) The uncertainty evaluation in whiplash test model is accomplished on the numerical evaluation methods. Type-A uncertainty is evaluated
via NN and SVM-based MCM. by statistical approach and characterized by the experimental standard
deviation, which is expressed as uA . Type-B uncertainty is obtained by
230
S. Wang et al. Measurement 119 (2018) 229–245
Table 1
The overall scoring principle of whiplash test according to C-NCAP Management Regulation (2015 edition) [35].
n
Start ∑ (x i−x )2
s (xk ) i=1
uA = u (x ) = s (x ) = =
n n (n−1) (1)
Define the output Y and determine the inputs
where, x represents the mean-value, n is the test times, s (xk ) is the
X 1 , X 2 , ..., X N
standard deviation of tests, and x i denotes the result of the i th test.
In general, the Type-A evaluation method based on the GUM is most
Research the PDFs of the inputs X i commonly utilized for estimating the uncertainty of measurement. For
i 1, 2, ..., N cases where the measurement model is linear and the probability dis-
tribution of the output is normal distribution, the accurate results can
be acquired. However, when the physical model is complex or the
Build a mathematical model between the
output and the inputs probability distribution of the output is significantly asymmetric, the
Y f ( X 1 , X 2 ,..., X N ) suggested method to evaluate the measurement uncertainty will result
in the inaccurate results.
It is known from Section 2 that the model of whiplash test is highly
Verification and validation of the model nonlinear so that the above-mentioned technique is not applicable. As a
result, the Monte Carlo method (MCM), which is a stochastic simulation
Choose the sample size M; method, was proposed. This approach is especially applicable to the
Extract sample values from the input PDFs following situation: (a) the test model show distinct nonlinearity; (b)
xi , r , i 1, 2, ..., N , r 1, 2, ..., M the probability density functions (PDFs) of the inputs are obviously
non-linear; (c) the PDFs of the outputs deviate largely from the normal
Calculate the model values distribution and t distribution, especially for PDFs presenting obvious
yr f ( x1, r , x 2 , r , ..., x N , r ), r 1, 2, ..., M nonlinearity.
The flowchart of the MCM for measurement uncertainty evaluation
of complex tests is shown in Fig. 2. For case of the whiplash test, it
N
yr yr 1
includes the following steps:
Interexchange
y r and y r (1) The score of whiplash is defined in Section 2. Several injury indexes
1 Y
affecting the test as well as the relationship between it and the
Gain the distribution function GY ( ) injury indexes are put forward. Furthermore, the injury indexes are
r r 1 treated as outputs, then the inputs related to the outputs are con-
sidered and the PDFs of them are studied in Section 4.
(2) We established and verified a mathematical model between the
Evaluate the standard uncertainty u(y) inputs and the outputs in Section 5. Choose the sample size and
extract sample values from the input PDFs so that the outputs and
the scores of whiplash test are calculated.
End (3) Sort the calculated scores in a strictly increasing order and then we
can acquire the distribution function of them. Therefore, we ob-
Fig. 2. Flowchart of the MCM for measurement uncertainty evaluation. tained the standard uncertainty in Section 6 and compared it to the
results evaluated by Type-A method based on the GUM.
non-statistical approach and characterized by empirical or standard
deviation estimated from the assumed probability distribution, re- In addition, based on the statistical theory, the Bayesian method
presented by uB . making full use of the existing information, sample information, em-
For the situation with sufficient and independent test data, the pirical information and statistical inference of the experimental dis-
Type-A uncertainty is more objective than other techniques for un- tribution, are also been widely used in the uncertainty quantification of
certainty quantification. A series of measurements are acquired and the many important fields. For whiplash test model, the posterior dis-
standard deviation of tests are obtained, then the Type-A uncertainty tribution of the Bayesian method is a combination of the prior dis-
can be represented as tribution and the measured samples, which is more accurate than the
Type-A uncertainty of the current sample information. Thus, it can re-
place the GUM, especially for the small sample measurements. In the
case of less empirical information, the accuracy of the Bayesian method
231
S. Wang et al. Measurement 119 (2018) 229–245
Head Clearance
Pelvic Position Prestress and
Indicator 20.8 Bumper
Head restraint
Center
Pelvic Angle
H Point
Backrest Angle
Table 2 In the third step, the prior distribution and the measured samples
The requirements and tolerances of the dummy setting according to the C-NCAP are combined to the posterior distribution, and then the estimate of the
Management Regulation (2015 edition) [35]. measurement and the standard uncertainty can be acquired by
Position Requirement Tolerance n μ0
x + s 2τ 2
s2 τ2
μ ̂ = E [h (μ |x )] = n 1
, σ 2 = D [h (μ |x )] =
HX XHPM forward 20 mm ± 5 mm + nτ 2 + s 2 (4)
s2 τ2
HZ ZHPM ± 10 mm
Pelvic angle 26.5° ± 2.5° Above all, in case of the physical models are linear and the prob-
Head angle 0° (Horizontal) ± 0.5°
Head clearance Bref + 15 mm ± 2 mm
ability distributions of outputs follow normal distribution, the mea-
surement uncertainty can be evaluated by the Type-A evaluation
method based on the GUM. For situation that the prior distribution and
is not as good as that of the MCM. Therefore, the Bayesian method is the empirical information are definite, or for the small sample mea-
limited to case where the measured sample approximates the normal surements, we can utilize the Bayesian method to replace the GUM.
distribution. It can be regarded as an improvement of the GUM, three However, for the complex physical model, such as the whiplash test in
steps of which are summarized here illuminating the idea of this the present manuscript, the model shows distinct nonlinearity and the
method: PDFs of the inputs and outputs deviate largely from the normal dis-
In the first step, the prior mean μ0 and the prior standard deviation tribution and t distribution, the results acquired by the MCM is more
of the mean τ are obtained by the historical data and the empirical accurate.
information. Suppose that the prior information satisfies the following
normal distribution: 4. Sensitivity analysis of inputs for whiplash test model
p (μ) ∼ N (μ0 ,τ 2) (2)
4.1. The main influence factors
In the second step, we calculated the mean value x and the standard
deviation s of the current sample by Whiplash test is a dynamic complex test of the seat. Many factors
influenced the test, including the dummy’s state parameters, the seat’s
n n
1 1 state parameters and the installation parameters.
x =
n
∑ x i, s =
n−1
∑ (x i−x )2
Before whiplash test, the dummy is required to implement a cali-
n=1 i=1 (3)
bration test to detect whether the movement state of the dummy’s neck
Table 3
Comparison of finite element analysis and test installation parameters.
Dummy installation parameters Coordinate of H point Pelvic angle (°) Head clearance (mm) Head angle (°)
X (mm) Z (mm)
Test FEA Test FEA Test FEA Test FEA Test FEA
3050.3 3050.8 469.4 470.3 27.4 25.86 27 27 0.1 0
Seat installation parameters Coordinate of rotating axis point Head restraint center Backrest angle (°)
X (mm) Z (mm) X (mm) Z (mm)
Test FEA Test FEA Test FEA Test FEA Test FEA
3257.4 3267.2 423.1 421.9 3436.7 3443.6 1011.1 1010 16.96 16.69
232
S. Wang et al. Measurement 119 (2018) 229–245
125 15
FEA FEA
100 Test 10 Test
NIC (m2 s 2 )
75 5
a (m s2 )
50 0
25 -5
0 -10
-25 -15
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
t(s) t(s)
(a) Comparison of the sled pulse (b) Comparison of the NIC
100 300
FEA FEA
Test Test
Upper neck Fx (N )
Upper neck Fz (N )
150
0
-100
-150
-200 -300
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
t(s) t(s)
(c) Comparison of the upper neck shear force Fx (d) Comparison of the upper neck tension Fz
100 300
FEA FEA
Test Test
Lower neck Fz (N)
50
Lower neck Fx (N)
-300
-50
-100 -600
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
t(s) t(s)
(e) Comparison of the lower neck shear force Fx (f) Comparison of the lower neck tension Fz
20 4
FEA FEA
Test Test
Lower neck M y (N. m)
Upper neck My(N. m)
2
10
0
-2
-10 -4
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
t(s) t ( s)
(g) Comparison of the upper neck torque M y (h) Comparison of the lower neck torque M y
Fig. 4. The sled pulse and the associated injury indexes of the FEA and the test.
is within the range specified by the standard. If the calibration result is neck of dummy are different in each formal whiplash test, which will
unqualified, it is necessary to adjust the dummy, of which the main resulted in the uncertainty.
parameter is the preload of the neck of dummy wire. As long as the After adjusting the relevant parameters of the dummy and the seat,
acceleration, speed, torque and other parameters of the upper and the dummy is required to be installing in the seat according to relevant
lower neck are controlled within a certain range, the calibration of the standards, which means to modulate the location and posture, in-
dummy is considered qualified. Thus, the preload force (F1,F2 ) of the cluding the dummy’s H-point position, pelvic angle, head angle, head
233
S. Wang et al. Measurement 119 (2018) 229–245
Table 4
Score of FEA and test.
Injury index High performance Low performance Value of test Value of FEA Individual score of Individual score of Score of Score of
limit limit test FEA test FEA
NIC 8 m2/s2 30 m2/s2 8.15 m2/s2 7.2 m2/s2 1.99 2 3.92 3.85
Upper neck Fx 340 N 730 N 0.26 N 21.72 N 0.94 0.85
Upper neck Fz 475 N 1130 N 189.97 N 215.2 N
Upper neck My 12 N·m 40 N·m 13.73 N·m 16.18 N·m
Lower neck Fx 340 N 730 N 85.35 N 58.3 N 1 1
Lower neck Fz 257 N 1480 N 104.26 N 29.15 N
Lower neck My 12 N m 40 N m 2.62 N m 2.33 N m
15 0.4 15 0.12
Frequency
0.2
5 5 0.04
0.1
0 0 0 0
-4 -2 0 2 4 -20 -15 -10 -5 0 5
Hx Hz
(a) PDF of H x (b) PDF of Hz
40 0.6 25 0.15
Probability density function
15
20 0.3
10
0.05
10 0.15 5
0 0 0 0
-2 -1 0 1 2 15 17.5 20 22.5 25
HC BA
(c) PDF of HC (d) PDF of BA
Fig. 5. The probability density distributions of the X coordinate of H point, the Z coordinate of H point, the head clearance and the backrest angle.
POT A
Rod POT B
Potentiometer POT C
234
S. Wang et al. Measurement 119 (2018) 229–245
20 10
FEA FEA
Test 0 Test
POT A (degree/°)
POT B (degree/°)
10
-10
-20
0
-30
-10 -40
0 40 80 120 160 200 0 40 80 120 160 200
t (ms) t (ms)
(a) Comparison of the POT A (b) Comparison of the POT B
Fig. 7. The consistence verification of the FEA and the calibration test.
clearance and other related position posture, as shown in Fig. 3. seat cannot be re-utilized, and the replace of the seat would increase the
However, due to the human factors and so forth, the adjustment results impact of uncontrollable factors. The finite element simulation is in-
cannot be precisely the same in the installing process of the dummy. It troduced to replace the physical test, which makes the adjustment of
is shown in Table 2 that the requirements and tolerances of the dummy the inputs easier, is a perfect solution to the above problems. Hence, it
setting, which means the error between the relevant position and the is necessary to verify the consistency of the FEA results and the physical
standard value is allowed. Consequently, the whiplash uncertainty test results. The actual model and finite element model of the whiplash
happens. test device are shown in Fig. 1(a) and (b), respectively. In addition, the
To establish the evaluation of whiplash uncertainty by the MCM, it comparison of the installation parameters of the two corresponding
is necessary to study the main influence factors and the PDFs of these models is shown in Table 3.
input parameters by large numbers of repeated tests. In general, the Fig. 4 represents the sled pulse and the injury indexes corresponding
finite element analysis (FEA) and the test are combined, which can help to the FEA and the whiplash test. From the curve of each output, the
us to obtain large amounts of samples as well as saving the time and changing trend is basically the same. The FEA curves generally lag
costs. As the head clearance is associated with the pelvic angle in the behind the test, while the peaks are roughly in synchronization. For
simulation of whiplash test, the head clearance is considered instead of Fig. 4(e) and (f) corresponding to the largest relative error, whose peaks
both. are far below the high-performance limits, we found that they do not
In summary, the factors affecting the measurement uncertainty of affect the results.
the whiplash test are: the preload force F1 of the neck of dummy, pre- Shown in Table 4 is the score of whiplash test compared with which
load force F2 of the neck of dummy, X coordinate of H point (HX ), Z of the FEA, it can be seen that the score of each injury index and the
coordinate of H point (HZ ), backrest angle and head clearance. overall score corresponding to the FEA meet well with which of the test.
Hence, we concluded that the simulation accuracy meets the require-
4.2. Consistency verification of finite element analysis and whiplash test ments and the FEA model can be used for further study and research.
results
4.3. PDFs of main inputs
It is elaborated in Section 4.1 that great deals of repeated tests are
required in the evaluation of whiplash uncertainty, which helps us to According to the historical data, the input parameter statistics are
study the relationship between the score of whiplash and the sensitivity acquired, which reveal that the X coordinate and the Z coordinate of H
of the influencing factors. However, each test requires plenty of human, point and the head clearance follow the normal distribution, while the
material and financial resources. Furthermore, the accomplished-test backrest angle obeys the uniform distribution. We show the PDFs of
235
S. Wang et al. Measurement 119 (2018) 229–245
them in Fig. 5 as follows: of slide and the torque of upper neck respectively by the acceleration
Consequently, the data are processed so that the PDFs of the X co- sensor and torque sensor in the dummy model.
ordinate of H point, the Z coordinate of H point, the head clearance and Before analyzing the PDFs of the preload forces, we can verify the
the backrest angle are obtained and expressed as Eqs. (5)(8). consistence of the calibration test and FEA. The preload forces are set as
the default value F1 = 68.7 N and F2 = 119.7 N , then we obtained the
Hx ∼ N (0.1498,1.23782) (5)
comparison of the calibration and FEA results and showed it in Fig. 7,
Hz ∼ N (−7.4775,4.05272) (6) which demonstrates that the FEA results meet well with which of the
calibration test. Thus, we can utilize the FEA model to study the PDFs of
HC ∼ N (−0.1098,0.73772) (7) the preload of the neck of dummy. It is worth noting that the POT C, the
acceleration of slide and the torque of upper neck changed little as the
BA ∼ U (17.5,22.5) (8)
change of F1 and F2 . Above all, in this section, the ranges of the preloads
Since there is no corresponding sensor to measure the magnitude of depend only on whether the upper and lower neck angles (POT A and
the preload force F1 and F2 , the probability density distribution function POT B) meet the requirements of the calibration test.
of them cannot be gained from the historical data. In addition, after a Keep F2 constantly, we obtained the curve of the upper neck angle
certain number of whiplash tests, we need to calibrate for the dummy. change with F1 and showed it in Fig. 8(a), it indicates that the POT A
Judging from the experience, the unqualified calibration phenomenon shifted upward as F1 increased. Note that in the period of 125 ms to
sometimes happened, thus, we need to adjust the preload force for new 135 ms, the curve may exceeded the upper limit as F1 changes, which
whiplash test. Therefore, the results of the calibration test can be uti- can be utilized to determine the maximum value of the preload force.
lized in the FEA to determine the range of the preload in which the Based on the standard of the calibration test, the upper neck angle must
calibration will be qualified, namely, the distributions of the preloads be within the range of −9° to 2°between 125 ms and 135 ms, thus, from
are obtained. After the dummy split, we obtained the dummy model for Fig. 8(b), we can figure out that the maximum value of F1 is 90 N.
the calibration and connected it to the test slide. The connection of the Similarly, the dependence of the lower neck angle on the preload
potentiometer with the rod is simulated by the kinematic pair of sliding force of front neck (F1) can be obtained and shown in Fig. 9(a), the POT
pommel (shown in Fig. 6), and then upper neck angle (POT A), the B also shifted upward as F1 increased. We investigated the variations of
lower neck angle (POT B) and the angle of the lower neck relative to the the minimum value of POT B and the value of which at 108 ms with F1
slide (POT C) can be acquired. Moreover, we acquired the acceleration and showed them in Fig. 9(b) and (c), respectively. As the lower neck
236
S. Wang et al. Measurement 119 (2018) 229–245
20 10
F2 =60N F2 =60N
F2=90N F2 =90N
F2=119.7N
0 F2=119.7N
POT A (degree/°)
POT B (degree/°)
10 F2 =150N F2 =150N
F2 =180N -10 F2 =180N
-20
0
-30
-10 -40
0 50 100 150 200 0 50 100 150 200
t (ms) t (ms)
(a) Variation of POT A with F2 (b) Variation of POT B with F2
16 7
FEA FEA
Maximum POT A (degree/°)
Y -0.0366X +17.5771
14 6.5
Y -0.0094X +7.4351
(99.68, 6.5)
12 6
(166.04, 11.5)
10 5.5
50 100 150 200 50 100 150 200
F2 (N) F2 (N)
(c) Relationship between maximum (d) Relationship between maximum
POT A and F2 POT B and F2
-29.5
FEA
(76.29, -30) Fitting Curve
-30
POT B (degree/°) at 98 ms
-30.5
-31
Y -0.012X -29.0845
-31.5
50 100 150
F2 (N)
angle must be less than −30° between 98 ms and 108 ms, and the Above all, in order to ensure the qualification of the calibration test,
minimum value must be larger than −36°, the range of F1 can be de- we must control the preload forces of the dummy neck in a certain
termined as [33.51 N, 82.78 N]. range. The PDFs of the calibration outputs are obtained by the historical
Meanwhile, F1 is kept constantly, we can get the relationship be- statistics and shown in Fig. 11, which all obey the normal distribution.
tween POT A and POT B with F2 and show it in Fig. 10(a) and (b). In Then, we inferred that the adjustable preload forces F1 and F2 .also
addition, the dependence of the maximum value of POT A, the max- follow the normal distribution. As the confidence interval of F1 and F2
imum value of POT B and the value of POT B at 98 ms on F2 are ac- are [33.51 N, 82.78 N] and [76.29 N, 99.48 N], we acquired the PDFs of
quired and respectively shown in Fig. 10(c)–(e). Thus, the range of F2 F1 and F2 in Eqs. (9) and (10) and shown it in Fig. 12.
can be determined as [76.29 N, 99.48 N].
237
S. Wang et al. Measurement 119 (2018) 229–245
8 0.8 8 1.2
Frequency
Frequency
4 0.4 4 0.6
2 0.2 2 0.3
0 0 0 0
10 11 12 13 14 15
4 4.5 5 5.5 6 6.5
Maximum POT A (degree/°) Maximum POT B (degree/°)
(a) PDF of maximum POT A (b) PDF of maximum POT B
8 0.8 6 0.4
Frequency
Frequency
4 0.4 3 0.2
0 0 0 0
-20 -19 -18 -17 -16 -15 10 12 14 16 18
0.04 0.03
Probability density function
0.03
0.02
0.02
0.01
0.01
0 0
20 40 60 80 100 80 100 120 140 160 180
F1 (N) F2 (N)
238
S. Wang et al. Measurement 119 (2018) 229–245
239
S. Wang et al. Measurement 119 (2018) 229–245
Table 5
50 groups of whiplash test results acquired by FEA.
Hx Hz Head clearance Backrest angle F1 F2 NIC Upper Fx Upper Fz Upper My Lower Fx Lower Fz Lower My Score
1 217.525 48.452 27.268 17.719 42.6 122.3 6.28 28.68 187.1 15.22 50.96 44.1 2.53 3.89
2 216.650 42.437 26.288 17.257 57.8 116.1 7.65 21.6 188.5 15.61 54.44 26.87 2.23 3.87
3 216.358 41.340 26.614 16.050 27.2 145.1 6.96 20.78 214.7 15.35 54.33 50.42 2.51 3.88
4 216.899 49.483 27.000 18.023 62.0 125.5 7.07 29.57 190.4 15.48 52.7 21.85 2.53 3.88
5 214.688 48.068 27.591 14.877 46.4 148.5 6.61 26.15 217.7 17.62 64.44 19.19 2.28 3.8
6 219.552 42.741 26.203 20.722 49.7 151.4 7.56 25.71 139.6 12.83 30.46 13.95 2.7 3.97
7 216.075 41.990 28.124 18.743 60.9 132.0 8.57 26.88 189.7 15.41 55.9 16.61 2.28 3.83
8 216.864 49.307 27.754 15.948 50.2 103.3 6.15 24.54 226.4 17.44 66.82 59.46 2.26 3.81
9 214.382 44.426 26.064 19.066 51.2 133.3 7.28 32.83 179.2 14.8 48.84 17.9 2.41 3.9
10 214.649 51.762 24.900 19.581 53.5 121.5 6.6 27.15 172.7 14.02 43.9 29.8 3.03 3.93
11 215.912 43.184 25.998 19.955 63.3 130.0 8.3 29.13 169.5 14.03 45.02 8.49 2.58 3.9
12 216.719 43.782 27.864 14.798 65.9 134.7 7.61 20.26 226.6 17.88 68.67 15.14 2.23 3.79
13 214.857 47.210 26.542 16.464 70.3 140.3 7.2 20.52 225.1 16.75 62.93 14.3 2.19 3.83
14 215.709 50.784 27.645 14.120 76.2 126.1 7.28 17.05 263.8 18.6 82.28 41.34 2.24 3.76
15 214.459 52.558 28.490 18.452 58.2 136.3 7.22 26.14 204.1 15.9 57.49 29.19 2.49 3.86
16 216.308 46.317 26.666 18.137 66.4 127.9 7.5 22.33 184.8 15.43 52.71 13.91 2.41 3.88
17 215.014 42.933 27.017 14.566 64.5 136.2 7.83 20.32 205 17.81 67.44 8.05 2.22 3.79
18 215.741 45.708 27.148 14.985 52.9 120.5 6.72 21.15 230.3 18.02 69.09 36.92 2.29 3.78
19 215.440 48.781 25.861 17.495 35.8 115.1 6.38 26.9 186.9 15.15 48.9 51.26 2.57 3.89
20 215.557 45.526 26.476 20.642 71.2 129.2 8.33 27.61 165.5 13.63 42.55 8.54 2.86 3.91
21 217.210 45.323 26.347 18.283 53.8 169.1 7.21 24.01 173.7 14.87 41.26 −16 2.6 3.9
22 215.606 46.907 26.632 19.382 73.9 156.1 7.92 30.59 183.7 14.8 48.65 −25 2.64 3.9
23 218.043 40.707 27.316 15.539 85.3 109.9 7.86 16.33 213.4 17.69 66.7 17.82 2.3 3.8
24 217.447 45.201 26.792 15.168 36.6 131.8 5.85 20.85 210.5 16.86 55.6 46.11 2.38 3.83
25 215.079 47.048 26.100 20.013 52.1 113.3 7.05 28.46 166.4 13.89 48.44 36.86 2.8 3.93
26 215.826 44.721 27.095 20.279 40.1 134.1 7.12 30.99 163.1 13.55 40.72 37.53 2.75 3.94
27 218.193 43.618 26.572 17.026 57.0 123.2 7.12 26.35 188.7 15.43 53.19 24.74 2.19 3.88
28 214.306 45.019 26.402 17.666 44.7 99.9 9.63 18.86 232.4 16.27 73.94 28.78 3.16 3.7
29 215.498 47.350 26.880 19.769 75.6 118.2 7.93 28.44 177.7 14.47 52.98 4.2 2.68 3.91
30 214.057 44.075 27.787 14.341 54.3 146.2 7.64 14.41 245.9 18.07 70.43 32.85 2.29 3.78
31 216.605 41.659 27.372 16.787 83.5 141.7 8.28 19.78 208.8 16.57 60.38 −7.9 2.38 3.81
32 215.358 47.952 27.440 15.234 48.7 139.0 6.67 19.27 227.4 17.57 67.83 24.62 2.27 3.8
33 216.140 40.030 25.749 16.901 56.0 164.4 7.92 19.89 190.9 15.91 51.58 −7.3 2.33 3.86
34 214.929 48.944 26.432 17.385 69.1 156.8 7.58 26.09 208.4 16.38 54.21 3.64 2.37 3.84
35 213.256 46.256 27.047 18.590 61.4 119.1 7.79 26.58 190.1 15.41 59.04 25.49 2.35 3.88
36 215.873 50.451 28.259 15.798 60.5 124.8 6.87 19.35 229 17.44 69.65 35.23 2.22 3.81
37 216.531 50.111 27.352 16.422 62.8 93.1 6.81 27.45 247.2 16.5 67.42 73.18 2.25 3.84
38 217.082 48.249 26.720 19.200 41.7 130.4 6.86 26.77 162.9 14.01 38.63 27.61 2.8 3.93
39 215.993 44.267 27.216 17.923 67.7 108.5 7.91 28.26 189.9 15.8 59.72 27.06 2.34 3.86
40 213.693 42.352 27.518 15.754 68.5 144.0 8.31 14.45 225.5 17.96 71.37 5.25 2.24 3.76
41 217.656 38.289 27.168 14.216 80.0 149.8 8.12 16.5 192.3 17.51 61.48 −22.6 2.29 3.79
42 217.002 49.946 26.747 20.198 55.3 153.3 7.08 33.61 157.4 13.4 35.02 7.07 3.06 3.95
43 216.446 47.544 27.480 16.187 59.3 141.4 7.04 20.36 251.4 16.72 65.38 49.34 2.31 3.83
44 213.964 39.602 25.414 19.006 65.1 143.6 8.68 27.55 184 15.35 50.92 −0.4 2.29 3.82
45 215.285 53.186 26.258 14.440 56.2 147.3 6.04 19.29 269.8 18.24 72.19 60.68 2.27 3.78
46 215.139 46.025 26.915 18.809 47.4 127.2 7.58 24.23 178 14.83 49.87 28.1 2.46 3.9
47 217.308 46.612 25.610 16.670 44.2 138.5 6.87 23.57 193.7 15.57 48.91 28.76 2.49 3.87
48 215.204 53.540 26.846 19.455 47.7 137.7 6.38 31.9 185 14.37 44.94 29.13 2.94 3.92
49 216.184 54.769 27.943 20.485 59.9 112.6 6.55 33.07 183.7 13.9 46.96 44.82 3.08 3.93
50 216.280 51.180 26.932 15.444 72.2 162.3 7 23.69 250.3 18.26 64.53 11.67 2.28 3.78
240
S. Wang et al. Measurement 119 (2018) 229–245
Start Start
Y Y
End
End
Fig. 15. Flowchart of the BPNN.
Fig. 17. Flowchart of the LS-SVM.
Bias b
Input X 1 Output Y1 then we regard the established model as an applicable model for sub-
sequent uncertainty quantification.
K ( X , X 1) Analogously, the mathematical model established by LS-SVM is
verified and validated, of which we select the kernel function as the
radial basis function. The prediction results of LS-SVM compared with
Input X 2 K(X , X 2) Output Y2 which obtained by FEA, the mean relative error (MRE) and the mean
square error (MSE) of the proposed technique are shown in Fig. 19.
Fig. 19(a)–(g) depicts that the results predicted by LS-SVM meet
well with the simulation results. Furthermore, the MRE of the NIC and
the upper neck torque My for the whiplash scoring are less than 3% and
the MSE is relatively small, then we verified and validated the estab-
Input X n lished model, which is more accurate than the BPNN model for the
Output Ym
K(X , X m) analysis of whiplash uncertainty. Then, the estimated values of the
score of whiplash, the standard uncertainty as well as the inclusion
The kernel functions interval corresponding to the probability p are similarly acquired by
MCM as which from the BPNN model.
Fig. 16. The architecture of the SVM.
torque My are the most likely deductions for the score of whiplash test. 6. Evaluation of the whiplash uncertainty in whiplash test
Therefore, the parameters adjustment in the neural network modeling
is mainly based on the relative error of these two items. The prediction Firstly, M (=106 ) groups of input data are randomly generated by
results of BPNN compared with which obtained by FEA, the mean re- Monte Carlo sampling method, including the preload force F1 of the
lative error (MRE) and the mean square error (MSE) of BPNN are shown neck of dummy, preload force F2 of the neck of dummy, X coordinate of
in Fig. 18. H point (HX ), Z coordinate of H point (HZ ), backrest angle and head
It is shown in Fig. 18(a)–(g) that the results predicted by BPNN meet clearance. Import them into the BPNN model established in Section 5.2,
well with the simulation results. Moreover, the MRE of the NIC and the and M groups of output data are then obtained. Secondly, we analyzed
upper neck torque My are less than 10% and the MSE is relatively small, the scores of whiplash test corresponding to the outputs according to
241
S. Wang et al. Measurement 119 (2018) 229–245
10
FEA MRE=9.84%
Prediction MSE=1.11
9
NIC (m2 s 2 )
8
5
1 4 7 10
Test points
(a) NIC
20 40
FEA MRE=6.03% FEA MRE=15.7%
Prediction MSE=1.56 Prediction MSE=25.02
Upper neck M y (N.m)
Upper neck Fx ( N )
18
30
16
20
14
12 10
1 4 7 10 1 4 7 10
Test points Test points
(b) Upper neck M y (c) Upper neck Fx
300 3.5
FEA MRE=11.53% FEA MRE=8.34%
Upper neck Fz ( N )
Prediction MSE=0.099
Lower neck M y ( N .m )
Prediction MSE=1379.91
250 3
200 2.5
150 2
1 4 7 10 1 4 7 10
Test points Test points
(d) Upper neck Fz (e) Lower neck M y
80 100
FEA MRE=6.37% FEA MRE=42.47%
Prediction MSE=30.07 Prediction MSE=463.66
Lower neck Fz ( N )
Lower neck Fx ( N )
60 50
40 0
20 -50
1 4 7 10 1 3 5 7 9
Test points Test points
(f) Lower neck Fx (g) Lower neck Fz
Fig. 18. Comparison of the BPNN prediction results and simulation results, the MRE and the MSE.
the scoring criteria in the C-NCAP specification. Finally, we obtained The scores of whiplash test corresponding to M samples are sorted in
the distribution function GY (η) , the estimated value y and the standard strict increment, and the distribution function GY (η) is obtained and
uncertainty u (y ) of the score of whiplash test and the shortest inclusion shown in Fig. 20.
interval [ylow ,yhigh ] corresponding to different inclusion probabilities Similarly, the measurement uncertainty quantification of whiplash
through the flowchart of measurement uncertainty evaluation de- test based on the LS-SVM model and the MCM, as well as the dis-
scribed in Fig. 2. The analysis results are shown in Table 6 as: tribution function of the score of whiplash test are obtained and shown
242
S. Wang et al. Measurement 119 (2018) 229–245
10
FEA MRE=2.34%
Prediction MSE=0.043
9
NIC (m2 s2 )
7
5
1 4 7 10
Test points
(a) NIC
21 50
FEA MRE=2.06% FEA MRE=9.58%
Prediction MSE=0.20 Prediction MSE=6.48
Upper neck My(N. m)
Upper neck Fx (N )
40
18
30
15
20
12 10
1 4 7 10 1 4 7 10
Test points Test points
(b) Upper neck M y (c) Upper neck Fx
400 3.5
FEA MRE=6.75% FEA MRE=7.19%
Prediction MSE=430.03 Prediction MSE=0.055
Lower neck M y (N. m)
Upper neck Fz (N )
300 3
200 2.5
100 2
1 4 7 10 1 4 7 10
Test points Test points
(d) Upper neck Fz (e) Lower neck M y
90 100
FEA MRE=4.83% FEA MRE=47.51%
Prediction MSE=13.14 Prediction MSE=308.87
Lower neck Fz (N)
Lower neck Fx (N)
50
60
30 -50
1 4 7 10 1 3 5 7 9
Test points Test points
(f) Lower neck Fx (g) Lower neck Fz
Fig. 19. Comparison of the LS-SVM prediction results and simulation results, the MRE and the MSE.
in Table 7 and Fig. 21, respectively. method to evaluate the measurement uncertainty will result in the in-
As the whiplash test model is highly nonlinear and mathematical accurate results. Thus, the MCM is the primary technique to evaluate
model between the score of whiplash test and the measurable quantity the uncertainty of the whiplash test. However, the results of the GUM
cannot be expressed by a simple polynomial or basic function, the GUM can also be used as a comparison and reference due to its universality,
243
S. Wang et al. Measurement 119 (2018) 229–245
Table 6 Table 9
Measurement uncertainty quantification of whiplash test based on the BPNN model and Measurement uncertainty evaluation of whiplash test based on the Bayesian method and
the MCM. the simulation.
Estimate value Standard Inclusion Shortest inclusion Estimate value Standard Inclusion Shortest inclusion
y uncertainty u (y ) probability p interval [ylow ,yhigh ] y uncertainty u (y ) probability p interval [ylow ,yhigh ]
3.84 0.044 90% [3.77, 3.91] 3.85 0.042 90% [3.78, 3.92]
95% [3.75, 3.93] 95% [3.77, 3.93]
99% [3.73, 3.95] 99% [3.74, 3.96]
value x and the standard deviation s of the prior and current samples by
the Bayesian approach expounded in Section 3. Then we estimated the
33.8 value of the whiplash test and the standard uncertainty of the posterior
distribution, as well as the shortest inclusion interval [ylow ,yhigh ] corre-
sponding to the different inclusion probabilities and showed it in
33.7 Table 9.
In conclusion, the estimate values of the score of whiplash test are
all about 3.85, the standard uncertainty is around 0.04, and the in-
3.6
0 2 4 6 8 10 clusion interval corresponding to inclusion probability p = 95% is
5 about [3.77, 3.93]. It can be easily seen that there is a subtle difference
Number of sample ( 10 )
between the results obtained by the BPNN model and the MCM, the LS-
Fig. 20. The distribution function GY (η) of whiplash test based on the BPNN model and SVM model and the MCM and the Bayesian method. However, the gap
the MCM. between them is within the allowable range of error. Thus, we can
assume that these results are highly reliable, which proves the accuracy
Table 7 of the uncertainty quantification and verifies the reasonable of the es-
Measurement uncertainty quantification of whiplash test based on the LS-SVM model and tablished model. According to Section 5.2, we obtained a more accurate
the MCM. model based on the LS-SVM and the MCM, which means the results of
the final uncertainty evaluation are priority refer to Table 7. Moreover,
Estimate value Standard Inclusion Shortest inclusion
y uncertainty u (y ) probability p interval [ylow ,yhigh ]
the results acquired by Type-A uncertainty evaluation are reliable only
on the mean value estimation. The error of the standard uncertainty,
3.85 0.039 90% [3.79, 3.91] the inclusion interval and other results are relatively large, which
95% [3.77, 3.93] means the results are for reference only. Finally, we further confirmed
99% [3.75, 3.95] that the GUM method is not applicable to the measurement uncertainty
evaluation of nonlinear models, which is consistent with the provisions
4 of the related technical specification.
7. Conclusions
3.9
Score of whiplash test
244
S. Wang et al. Measurement 119 (2018) 229–245
mathematical model established based on the LS-SVM and MCM possess sensitivity analysis using Monte Carlo simulation, Med. Decis. Making 5 (1985)
a relatively small error, which can be used to guide the practice in the 157–177.
[15] G. Wübbeler, M. Krystek, C. Elster, Evaluation of measurement uncertainty and its
future. Moreover, the evaluation results by the BPNN and MCM or the numerical calculation by a Monte Carlo method, Meas. Sci. Technol. 19 (2008)
Bayesian method are also accurate, while the GUM method is not sui- 084009.
table for the uncertainty quantification for the highly nonlinear whi- [16] M. Hölle, C. Bartsch, P. Jeschke, Evaluation of measurement uncertainties for
pneumatic multihole probes using a Monte Carlo method, J. Eng. Gas Turbines
plash test model. (d) This paper only analyzes the inclusion interval Power 139 (2017) 072605.
corresponding to the inclusion probability p = 90%, 95% and 99%, it [17] R. Palenčár, P. Sopkuliak, J. Palenčár, S. Ďuriš, E. Suroviak, M. Halaj, Application of
can be extended to other probabilistic cases. Monte Carlo method for evaluation of uncertainties of ITS-90 by standard platinum
resistance thermometer, Meas. Sci. Rev. 17 (2017) 108–116.
The proposed method provides a powerful tool to evaluate the [18] K.M. Hanson, A framework for assessing uncertainties in simulation predictions,
measurement uncertainties associated with the system parameters in Physica D 133 (1999) 179–188.
whiplash test. Nevertheless, this article has its limitations, for example, [19] B. Larget, D.L. Simon, Markov chain Monte Carlo algorithms for the Bayesian
analysis of phylogenetic trees, Mol. Biol. Evol. 16 (1999) 750–759.
the variations in the sled pulse are not considered, the whiplash injury
[20] M. Filippone, M. Girolami, Pseudo-marginal Bayesian inference for Gaussian pro-
mechanism are not studied and the uncertainty evaluation for larger cesses, IEEE Trans. Pattern Anal. Mach. Intell. 36 (2014) 2214–2226.
head clearance are not discussed. In future, the proposed technique may [21] H.F. Lam, S.A. Alabi, J.H. Yang, Identification of rail-sleeper-ballast system through
be extended to include these situations for different uncertainties of time-domain Markov chain Monte Carlo-based Bayesian approach, Eng. Struct. 140
(2017) 421–436.
different NCAP whiplash test scores. Furthermore, we may promote it [22] C. Faure, F. Ablitzer, J. Antoni, C. Pézerat, Empirical and fully Bayesian approaches
to other fields for measurement uncertainty quantification, such as the for the identification of vibration sources from transverse displacement measure-
artificial intelligence, the structural engineering and the vibration ments, Mech. Syst. Sig. Process. 94 (2017) 180–201.
[23] D. Xiu, G.E. Karniadakis, Modeling uncertainty in steady state diffusion problems
pattern recognition, and will report results in future publications. via generalized polynomial chaos, Comput. Methods Appl. Mech. Eng. 191 (2002)
4927–4948.
Acknowledgements [24] A. Petrocchi, D. Spina, D.M.M.-P. Schreurs, Measurement uncertainty propagation
in transistor model parameters via polynomial chaos expansion, IEEE Microwave
Wirel. Compon. Lett. 27 (2017) 572–574.
The work reported in this paper was supported by the National [25] J.C. Helton, J.D. Johnson, C.J. Sallaberry, C.B. Storlie, Survey of sampling-based
Natural Science Foundation of China under Grant No. 11602142. methods for uncertainty and sensitivity analysis, Reliab. Eng. Syst. Saf. 91 (2006)
1175–1209.
[26] C.J. Roy, W.L. Oberkampf, A comprehensive framework for verification, validation,
References and uncertainty quantification in scientific computing, Comput. Methods Appl.
Mech. Eng. 200 (2011) 2131–2144.
[27] S. Boyaval, A fast Monte-Carlo method with a reduced basis of control variates
[1] M. Moog, J. Quintner, T. Hall, M. Zusman, The late whiplash syndrome: a psy-
applied to uncertainty propagation and Bayesian estimation, Comput. Methods
chophysical study, Eur. J. Pain 6 (2002) 283–294.
Appl. Mech. Eng. 241–244 (2012) 190–205.
[2] Y. Sterner, B. Gerdle, Acute and chronic whiplash disorders-a review, J. Rehabil.
[28] N. Chen, D. Yu, B. Xia, M. Beer, Uncertainty analysis of a structural-acoustic pro-
Med. 36 (2004) 193–210.
blem using imprecise probabilities based on p-box representations, Mech. Syst. Sig.
[3] B.D. Stemper, N. Yoganandan, F.A. Pintar, Effect of head restraint backset on head-
Process. 80 (2016) 45–57.
neck kinematics in whiplash, Accid. Anal. Prev. 38 (2006) 317–323.
[29] X. Chen, J. Li, A subset multicanonical Monte Carlo method for simulating rare
[4] P.C. Ivancic, D. Sha, M.M. Panjabi, Whiplash injury prevention with active head
failure events, J. Comput. Phys. 344 (2017) 23–35.
restraint, Clin. Biomech. 24 (Nov 2009) 699–707.
[30] G. Huang, S. Song, C. Wu, K. You, Robust support vector regression for uncertain
[5] D.W. Mang, G.P. Siegmund, H.J. Brown, S.C. Goonetilleke, J.S. Blouin, Loud pre-
input and output data, IEEE Trans. Neural Networks Learn. Syst. 23 (2012)
impact tones reduce the cervical multifidus muscle response during rear-end col-
1690–1700.
lisions: a potential method for reducing whiplash injuries, Spine J. 15 (2015)
[31] H. Quan, D. Srinivasan, A. Khosravi, Incorporating wind power forecast un-
153–161.
certainties into stochastic unit commitment using neural network-based prediction
[6] M. Özdemir, S.K. İder, M.İ. Gökler, Experimental and numerical investigation of
intervals, IEEE Trans. Neural Networks Learn. Syst. 26 (2015) 2123–2135.
comparability of whiplash sled test results, J. Braz. Soc. Mech. Sci. Eng. 38 (2016)
[32] F.R. Cecconi, M. Manfren, L.C. Tagliabue, A.L.C. Ciribini, E. De Angelis,
395–402.
Probabilistic behavioral modeling in building performance simulation: a Monte
[7] M. Van Ratingen, J. Ellway, M. Avery, P. Gloyns, V. Sandner, T. Versmissen, The
Carlo approach, Energy Build. 148 (2017) 128–141.
Euro NCAP Whiplash Test, 2009, pp. 09–0231.
[33] J.C. Helton, F.J. Davis, Latin hypercube sampling and the propagation of un-
[8] Joint Committee for Guides in Metrology, Evaluation of Measurement Data-An
certainty in analyses of complex systems, Reliab. Eng. Syst. Saf. 81 (2003) 23–69.
Introduction to The “Guide to The Expression of Uncertainty in Measurement” and
[34] S. Marino, I.B. Hogue, C.J. Ray, D.E. Kirschner, A methodology for performing
Related Documents, JCGM 104:2009.
global uncertainty and sensitivity analysis in systems biology, J. Theor. Biol. 254
[9] M.A. Navacerrada, A. Pedrero, C. Díaz, Study of the uncertainty of façade sound
(2008) 178–196.
insulation measurements: analysis of the ISO 12999–1 uncertainty proposal, Appl.
[35] China Automotive Technology & Research Center, C-NCAP Management
Acoust. 114 (2016) 1–9.
Regulation, ed, 2015.
[10] A. Onatski, N. Williams, Modeling model uncertainty, J. Eur. Econom. Assoc. 1
[36] V. Vapnik, The Nature of Statistical Learning Theory, Springer-Verlag, New York,
(2003) 1087–1122.
1995.
[11] J.L. Beck, K.-V. Yuen, Model selection using response measurements: Bayesian
[37] C. Cortes, V. Vapnik, Support-vector networks, Machine Learning 20 (1995)
probabilistic approach, J. Eng. Mech. 130 (2004) 192–203.
273–297.
[12] D. Higdon, M. Kennedy, J.C. Cavendish, J.A. Cafeo, R.D. Ryne, Combining field data
[38] V. Vapnik, Statistical Learning Theory, John Wiley, New York, 1998.
and computer simulations for calibration and prediction, SIAM J. Scien. Comput. 26
[39] V. Vapnik, An overview of statistical learning theory, IEEE Trans. Neural Networks
(2004) 448–466.
10 (1999) 988–999.
[13] X. Jiang, S. Mahadevan, Bayesian hierarchical uncertainty quantification by
[40] J.A.K. Sutkens, J. Vandewalle, Least squares support vector machine classifiers,
structural equation modeling, Int. J. Numer. Meth. Eng. 80 (2009) 717–737.
Neural Process. Lett. 9 (1999) 293–300.
[14] P. Doubilet, C.B. Begg, M.C. Weinstein, P. Braun, B.J. McNeil, Probabilistic
245