You are on page 1of 11

Pagination not final (cite DOI) / Pagination provisoire (citer le DOI)

ARTICLE
Interpolating spatially varying soil property values from sparse
data for facilitating characteristic value selection
Tengyuan Zhao, Silvana Montoya-Noguera, Kok-Kwang Phoon, and Yu Wang

Abstract: Limit state design, incorporated into many recent geotechnical design codes, introduces the application of partial or
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 14.136.36.238 on 12/16/17

resistance factors to selected characteristic values. Partial or resistance factors are usually set by national standard organizations,
while characteristic values of geotechnical parameters are selected by engineers, often based on sparse measurement data
combined with subjective engineering experience and judgment. Due to this subjective selection and individual judgment, the
characteristic value derived by different engineers from the same dataset may vary greatly, especially when the test data contain
significant variability. To address this issue, a new method based on Bayesian compressive sampling (BCS) is proposed in this
study. BCS is able to reconstruct a high-resolution geotechnical property profile from sparse measurement data and quantify the
uncertainty, e.g., confidence interval (CI) associated with the interpreted profile. The quantified uncertainty in the BCS has a
clear statistical meaning: the corresponding confidence level for a CI from the BCS is the expected coverage proportion (i.e.,
fraction) of the complete profile that falls within the CI, if all data points along depth can be measured to provide the complete
profile. This statistical meaning can be used to facilitate objective determination of characteristic values for geotechnical
properties.

Key words: reliability-based design, Bayesian compressive sampling, compressive sensing, sparse measurement data, site
investigation.
For personal use only.

Résumé : La conception états limites, intégrés dans de nombreux codes de conception géotechnique, présente l’application de
facteurs de résistance de résistance ou partielle à certaines valeurs de caractéristique. Des facteurs de résistance ou partiels sont
habituellement fixés par les organismes de normalisation, tandis que les valeurs caractéristiques des paramètres géotechniques
sont sélectionnées par les ingénieurs, souvent basées sur des données de mesure combinée à l’expérience en génie et de
jugement subjectif. En raison de cette sélection subjective et ce jugement individuel, la valeur caractéristique dérivée par
différents ingénieurs du même jeu de données peut varier considérablement, en particulier lorsque les données d’essai contien-
nent une variabilité importante. Pour résoudre ce problème, une nouvelle méthode basée sur l’échantillonnage en compression
bayésien (« BCS ») est proposée dans cette étude. BCS est capable de reconstruire une propriété géotechnique à haute résolution
à partir de données de mesure profil clairsemé et de quantifier l’incertitude, par exemple, l’intervalle de confiance (IC) associé
à l’interprétation de profil. L’incertitude quantifiée dans le BCS a une signification statistique claire: le niveau de confiance
correspondant pour un IC de la BCS est la couverture prévue (proportion) de la fraction profil complet qui relève de l’IC, si tous
les points de données le long de la profondeur peuvent être mesurés pour fournir le profil complet. Cette signification statistique
peut être utilisée pour faciliter l’objectif de détermination des valeurs caractéristiques des propriétés géotechniques. [Traduit
par la Rédaction]

Mots-clés : conception basée sur la fiabilité, échantillonnage compression bayésienne, détection de compression, données
clairsemées de mesure, étude de site.

Introduction values by resistance factors. Partial or resistance factors are usu-


ally set by national standard organizations and are used to achieve
Limit state design methods have been recently incorporated
a target level of reliability or safety (although explicit reliability
into many geotechnical codes of practice throughout the world,
calibration may not be conducted), while the characteristic values
e.g., Eurocode 7 (CEN 2004), the AASHTO (1998) bridge code, and of geotechnical parameters are selected by geotechnical engi-
the Canadian highway bridge design code (CSA 2014), among others. neers. In engineering practice, these characteristic values are of-
To achieve a specific target reliability level, the design value is ten selected based on a limited number of test results, therefore
determined by dividing the characteristic strength values by par- engineering judgment and previous relevant experience are
tial factors (e.g., Meyerhof 1995; Fenton and Naghibi 2011; Reddy frequently used to select the characteristic values (e.g., Orr
and Stuedlein 2017) or multiplying the characteristic resistance 2017).

Received 17 April 2017. Accepted 27 June 2017.


T. Zhao, S. Montoya-Noguera,* and Y. Wang.† Department of Architecture and Civil Engineering, City University of Hong Kong, Tat Chee Avenue,
Kowloon, Hong Kong.
K.-K. Phoon.† Department of Civil and Environmental Engineering, National University of Singapore, Singapore.
Corresponding author: Yu Wang (email: yuwang@cityu.edu.hk).
*Present address: Department of Civil Engineering, Universidad EAFIT, Colombia.
†K.-K. Phoon currently serves as an Associate Editor; Y. Wang currently serves as an Editorial Board Member; peer review and editorial decisions regarding

this manuscript were handled by D. Sheng.


Copyright remains with the author(s) or their institution(s). Permission for reuse (free in most cases) can be obtained from RightsLink.

Can. Geotech. J. 00: 1–11 (0000) dx.doi.org/10.1139/cgj-2017-0219 Published at www.nrcresearchpress.com/cgj on xx xxx xxxx.
Pagination not final (cite DOI) / Pagination provisoire (citer le DOI)
2 Can. Geotech. J. Vol. 00, 0000

Because of this subjective selection and individual judgment, the structure at the limit state. For example, some researchers
the characteristic values, derived by different geotechnical engi- have argued that the characteristic value is related to the concept
neers from the same dataset, may vary greatly, especially when of a mobilized strength along the critical slip surface (Ching and
the test data are scarce or contain significant variability. For ex- Phoon 2013a, 2013b; Ching et al. 2014, 2016a) or a mobilized mod-
ample, Bond and Harris (2008) presented three case studies in ulus over a domain influenced by the structure at the limit state
which about 100 engineers were asked to select the characteristic (Ching et al. 2016b). As the limit state of a geostructure is problem-
values on the basis of Eurocode 7 from the same set of test data. dependent, and a realistic assessment of the characteristic value
The case studies dealt with different types of data (standard pen- in the context of spatial variability where nonclassical failure
etration test blow counts, field vane tests, and triaxial tests), dif- mechanisms can emerge is less straightforward, the extent of the
ferent soil types (clays and gravels), and different numbers of data failure zone governing the behavior of the structure at the limit
points (from 25 to above 100 points for profiles 10–30 m deep). The state is not considered in this paper.
selected characteristic values varied greatly, and the maximum
characteristic value obtained was about 3–5 times greater than Coverage proportion of confidence interval profiles
the minimum one. Orr (2017) suggested that more guidance is
The CI, which may be used to quantify uncertainty, is more
needed to select characteristic values in an objective manner to
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 14.136.36.238 on 12/16/17

informative than simply reporting a point estimate (e.g., Phoon


reduce and properly account for this broad range of interpretation.
and Ching 2014). CI is an interval estimation of a parameter of
Statistical analyses of laboratory and in situ test results to de-
interest that gives a confidence level that the true parameter falls
termine geotechnical parameters for reliability-based design ap-
within the estimated CI. To evaluate the confidence interval, an-
plications have been discussed broadly and recommended in
alytical equations can be used when the distribution of the data is
the literature (e.g., Vanmarcke 1977; Phoon and Kulhawy 1999;
known. For example, for normally distributed data with a known
Baecher and Christian 2003; Fenton and Griffiths 2008; Becker
2010; Gong et al. 2014; Li et al. 2016; Phoon et al. 2016). Although mean (␮) and standard deviation (␴), the CI for a confidence level
statistical methods are explicitly recommended in some design ␣, denoted as CI␣, is expressed as
guides (e.g., Det Norske Veritas 2010), currently the use of such
analyses has not been included in some existing design codes, (1) CI␣ ⫽ ␮ ± z(1⫺␣)/2␴
such as Eurocode 7, partly because of the limited number of site-
specific measurement data and the inherent variability encoun- where z(1−␣)/2 = −⌽−1[(1 − ␣)/2] and ⌽−1(·) is the inverse standard
tered in natural soil deposits (Orr 2017). Furthermore, most of the normal cumulative distribution function. Note that the lower
available statistically based methods for determining the charac- bound of eq. (1) or its variants (i.e., ␮ minus a factored ␴) has been
For personal use only.

teristic values of geotechnical parameters focus on point statistics proposed in literature, e.g., Schneider and Schneider (2013) and
(e.g., mean and coefficient of variation for a previously defined Orr (2017), as a characteristic value at a given depth or for a homo-
homogeneous soil layer) and hence ignore the spatially varying geneous soil layer in Eurocode 7.
pattern of soil properties (e.g., Cao and Wang 2014; Wang et al. For spatially varying data, such as random field samples (RFSs),
2016a; Wang and Aladejare 2016; Wang and Cao 2013). the values vary along a spatial dimension (e.g., depth), hence the
This paper aims to address these two issues — limited data and mean and CI␣ also vary along this spatial dimension, e.g., profiles
inherent spatial variability — and to provide a statistical proce- varying with depth. For a given random field with known ␮ and ␴,
dure for an “objective” determination of characteristic value from CI␣ profiles (i.e., variations of CI␣ with depth) can be generated
spatially varying, but “sparsely measured,” data. It uses compres- analytically by applying eq. (1) to different depths. For example,
sive sampling theory to reconstruct the best estimate of a soil the CI90% profiles can be obtained by substituting ␣ = 90% to eq. (1)
property profile from sparse measurement data points (Wang and
for different depths of interest.
Zhao 2016) and Bayesian theory to estimate the statistical uncer-
tainty associated with the interpreted profile (Wang and Zhao Coverage proportion (CP␣) of CI␣ profiles for random field
2017). Use of the Bayesian framework acknowledges the critical data
role of engineering judgment, but reduces the subjective interpre- The coverage proportion (CP␣) of a soil property profile f, e.g., a
tation uncertainty by quantitatively representing it as prior RFS that falls within a CI␣ profile with a confidence level (␣), is
knowledge (e.g., Cao et al. 2016; Vick 2002; Wang and Aladejare defined as (e.g., Marra and Wood 2012; Nychka 1988; Wahba 1983)
2015; Wang et al. 2016b).
An interpretation of the statistical meaning of the confidence N
interval for random field data is first presented. Then the formu-
lation of Bayesian compressive sampling (BCS) is reviewed and
(2) CP␣ ⫽
1
N 兺 [I(f 僆 CI )]
k⫽1
k ␣

used to provide average and confidence interval profiles given


only sparsely measured, but spatially varying, geotechnical data.
Note that the measurement data of soil properties obtained in where N is the total number of data points in the soil property
geotechnical engineering are usually sparse and limited, particu- profile f and I共 · 兲 is the indicator function. I(·) equals unity if a data
larly for small- or medium-sized projects. An important question point fk (k = 1, 2, …, N) is within the upper and lower bounds of CI␣;
when interpreting sparse data in geotechnical practice is: How otherwise, it is zero. Note that the expected value of CP␣ is equal to
does the profile interpreted from sparse data compare with the ␣ (Wahba 1983). For example, CI95% implies that 95% of all data
measured profile, if it is possible to measure the geotechnical data points are expected to fall within the upper and lower bounds
with a small interval and a high resolution? This paper shows that given by CI95%, i.e., CP95% = 95%. The coverage proportion has been
the confidence interval (CI) profile quantified in BCS has a clear evaluated for CI␣ profiles obtained from smoothing functions
statistical meaning, which may be used to address this question (e.g., Nychka 1988; Wahba 1983) and generalized additive models
and to facilitate determination of characteristic values in engi- (e.g., Marra and Wood 2012). In this section, the procedure to
neering practice. For illustration, the proposed BCS procedure is evaluate CP␣ of a RFS that falls within a CI␣ profile is firstly ex-
applied to a real case of cone penetration test (CPT) data and the plained. Then the procedure is illustrated with simulated random
selection of the characteristic value of effective friction angle. field data. This section is meant to explain the definition and
This paper addresses the characteristic value only from a purely evaluation of CP␣, and it paves the way for the next section where
“statistical” perspective, although characteristic value may be re- CP␣ will be evaluated for the CI␣ profiles obtained from the BCS
lated to the extent of the failure zone governing the behavior of method with a limited number of measurement data as input.

Published by NRC Research Press


Pagination not final (cite DOI) / Pagination provisoire (citer le DOI)
Zhao et al. 3

Fig. 1. Ns = 1000 sets of random field samples (RFSs) generated for soil property X: 95% coverage proportion (CP95%) of (a) 94.5%; (b) 94.9%;
(c) 96.1%. [Color online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 14.136.36.238 on 12/16/17
For personal use only.

Note that if the random field has no correlation (i.e., the data fore, it accurately represents the prescribed random field (e.g.,
points over the depth are independent), the probability distribu- Phoon et al. 2002). The number of RFSs (Ns) generated is 1000. All
tion of CP␣ follows a binomial distribution because of the indica- RFSs generated are shown in Fig. 1 in light gray. Additionally, the
tor function in eq. (2) (e.g., De Veaux et al. 2014; Efron and profile of the mean values, evaluated as the average of all Ns values
Tibshirani 1993). Hence, the average CP␣ is equal to ␣ and the at each depth, is also shown.
variance is equal to ␣(1 – ␣)/N. The shape of the distribution of CP␣
is symmetric (i.e., zero skewness) for ␣ = 0.5 and it is negatively Probability distribution of CP␣ for CI␣ profiles
skewed as ␣ approaches 1 (De Veaux et al. 2014). The distribution The CI␣ profiles over depth are constructed using eq. (1) with a
of CP␣ tends to the normal distribution as N increases and ␣ ap- given confidence level ␣. For example, given ␣ = 95%, the CI95%
proaches 0.5. De Veaux et al. (2014) suggest that a normal distri- profiles are obtained with the 2.5th and 97.5th percentiles from
bution gives a good approximation of the binomial distribution if eq. (1) and are shown in Fig. 1 by two dotted lines. In each subplot
N␣ ≥ 10 and N(1 − ␣) ≥ 10. of this figure, one RFS is shown in a black solid line to illustrate
the evaluation of CP95%. For each one of these samples, the CP95%
Simulation of random field examples of the RFS profile that is within the CI95% profiles (i.e., between the
For illustration, a one-dimensional stationary Gaussian random two dotted lines in Fig. 1) is evaluated. The CP95% values for the
field is used to represent a soil property X profile. RFSs are gener- three RFSs presented in Figs. 1a, 1b, and 1c are 94.5%, 94.9%, and
ated using a truncated Karhunen–Loève (KL) expansion. Trun- 96.1%, respectively.
cated KL expansion has been increasingly studied and used in The CP␣ values are evaluated for all Ns = 1000 RFSs and for
simulating one-dimensional random processes in recent years different ␣ values ranging from 50% to 95%. Statistical analysis is
(e.g., Zhang and Ellingwood 1994; Phoon et al. 2002, 2005; Li et al. performed for the CP␣ values obtained, and the results are shown
2014). The following parameters were used: mean ␮X = 30; stan- as box-and-whiskers plots for different ␣ values in Fig. 2b. The box
dard deviation ␴X = 2; an exponential correlation function, i.e., is constructed with the interquartile range, IQR = 25%–75% per-
␳i,j ⫽ exp ⫺ 冉 ␭c 冊
2|dXi⫺dXj|
, where dXi and dXj are the depths of two X
data points Xi and Xj, respectively; and ␭c is the correlation length
centiles, and the whiskers show the minimum and maximum
values within 1.5IQR. The maximum and minimum CP␣ values
among all Ns RFSs are shown by crosses in Fig. 2, and the mean CP␣
taken as 2 m in this example. The soil layer thickness (h) is taken values are shown with circles. Figure 2 also includes a 1:1 line in
as 20.44 m and the profile has a resolution of 0.04 m. Hence, there each subplot. The mean CP␣ values for the ␣ value varying from
are N = 512 points for each RFS. Note that only one homogeneous 50% to 95% all plot along the 1:1 line, and the average CP␣ is equal
soil layer (rather than several soil layers) with a thickness of to ␣. The CI␣ profiles can be statistically interpreted as the upper
20.44 m is considered in this illustrative example, and therefore and lower bounds of an interval where the expected coverage
stratification is not needed. For the truncated KL expansion, 200 proportion (i.e., fraction) of a RFS (i.e., a spatially variable soil
KL terms are used. Note that 200 terms are able to preserve 98.2% property X profile) that falls within the interval is ␣.
of the total variance of the random field in terms of the sum of 200 Note that, although the expectation of CP␣ is ␣, the CP␣ value
eigenvalues and the sum of all eigenvalues in this example. There- for each RFS may vary significantly, as shown in Fig. 2. When the

Published by NRC Research Press


Pagination not final (cite DOI) / Pagination provisoire (citer le DOI)
4 Can. Geotech. J. Vol. 00, 0000

Fig. 2. Box-and-whiskers plot for the coverage proportion (CP␣) as a function of the confidence level (␣) for different correlation lengths (␭c):
(a) 0.5 m, (b) 2 m, and (c) 5 m. (Mean values are shown with circles; median values are shown with a line inside the box; minimum and
maximum values are shown with crosses.) [Color online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 14.136.36.238 on 12/16/17

CP␣ values are greater than the ␣ values, a relatively large propor- ence between mean and median values increases with the corre-
tion of the RFS values fall within the interval. In contrast, when lation length. Furthermore, the size of the boxes in Fig. 3a
the CP␣ values are smaller than the ␣ values, a relatively large increases with ␭c, suggesting that the variability of the CP␣ results
proportion of the RFS values fall outside the CI␣ profiles. For ex- increases with ␭c.
ample, for the worst case in Fig. 2b, the smallest CP50% value is To visualize the effect of the correlation length and the confi-
close to 20%. This means that, besides the 50% expected, an addi- dence level on the variation of CP␣ results, Fig. 3b shows the
tional 30% of the RFS values fall outside the CI50% profiles. For all standard deviation evaluated for CP␣ (␴CP␣) as a function of ␣.
Ns = 1000 RFSs, less than 3% of RFSs present a CP95% lower than When there is no correlation in the random field, the standard
85.5% (i.e., a relative difference of 10% on the expected value). In deviation is evaluated as ␴CP␣ ⫽ 兹␣共1 ⫺ ␣兲/N, as for the binomial
For personal use only.

contrast, for CP50%, about 30% of RFSs are below 45% (i.e., the same distribution, and shown by a solid line in Fig. 3b. A random field
10% difference). As the confidence level increases, the variability with no correlation was tested and the results obtained for ␴CP␣
of the CP values decreases, as shown by the decreasing size of the agree well with the analytical solution. When there is no correla-
box and whiskers as the ␣ value approaches unity. In addition, the tion, the ␴CP␣ reaches its maximum for ␣ = 0.5. However, as ␭c
mean and median CP␣ values (shown with a line inside the boxes increases, the maximum value of ␴CP␣ occurs at a relatively large
in Fig. 2) are similar for ␣ = 50%, but the median CP␣ value is ␣ value. For ␭c = 10 m, a maximum ␴CP␣ of about 0.18 is found at ␣ =
slightly larger than the mean CP␣ value for high ␣ values. In other 0.65. In general, ␴CP␣ increases as the ␭c value increases. But this
words, the CP␣ results are symmetric, i.e., present zero skewness, increase is more pronounced for relatively small ␣ values. For ␣ =
for ␣ close to 0.5, but develop a negative or left skewness as ␣ 0.95, ␴CP␣ is quite small, even for high correlation lengths because
approaches unity. This is similar to the effect of small N values the upper bound of CP95% = 1 is reached in many cases.
when the random field has no correlation and the CP␣ values It is worth noting that the aforementioned CI␣ profiles are ob-
follow a binomial distribution (De Veaux et al. 2014). tained from a prescribed random field with known parameters. In
geotechnical engineering practice, if a soil property profile is rep-
Effect of different correlation length in random field
resented by a random field, its random field parameters — such as
The same procedure as described previously was used for gen-
erating RFSs with various ␭c values. Figures 2a and 2c show the mean, standard deviation, and correlation function — are often
box-and-whiskers plots of the corresponding CP␣ results for ␭c = difficult to estimate from measurement data, especially the last
0.5 and 5 m, respectively. For all the cases, the mean CP␣ is equal two parameters. This is in part because the measurement data of
to ␣, although the results present more variability for the case of soil properties are usually limited and sparse. To address this
␭c = 5 m. The variability of CP␣ increases as the correlation length difficulty, a BCS method has been developed recently to statisti-
increases, as shown by the size of the boxes and the length of the cally interpret the sparse measurement data points for providing
whiskers in Fig. 2. Compared to the case of ␭c = 2 m, the CP␣ results the best estimate and CI␣ profiles of the soil properties (Wang and
for the case of ␭c = 0.5 m are more concentrated around ␣. The Zhao 2017), as briefly reviewed in the next section.
decrease in the CP␣ variability as ␣ increases is less visible for the
case of ␭c = 0.5 m. In addition, for the case of ␭c = 0.5 m, the median
Review of BCS
CP␣ is also equal to ␣, and all CP␣ distributions appear to be BCS is a coupling of compressive sampling or sensing (CS) and
symmetric. In contrast, for the case of ␭c = 5 m shown in Fig. 2c, the Bayesian method to reconstruct the average and standard
the variability of CP␣ increases, and distributions are markedly deviation profiles of a soil property profile from only partial in-
skewed for high ␣ values. It can be seen that, for ␣ greater than formation of the profile, i.e., sparse measurement data points
0.7, the median CP␣ values are greater than ␣. Thus more than 50% (e.g., Ji et al. 2008; Wang and Zhao 2017; Wang et al. 2017a). CS,
of RFSs have a CP␣ equal to or greater than ␣. mainly applied in electrical engineering and computer science,
A total of eight different ␭c values were tested ranging from 0.1 exploits sparsity, or compressibility, in many real-world signals
to 10 m. This range of correlation length is consistent with those (e.g., Candès et al. 2006; Candès and Wakin 2008). A signal, de-
of geotechnical properties reported in literature (e.g., Phoon and noted as a column vector f with a length of N, is defined as the
Kulhawy 1999). The box plots of the CP␣ results are shown in variation of a physical quantity with time or space (e.g., a soil
Fig. 3a for three ␣ values: 50%, 80%, and 95%. For all results, the property profile in this study). “Compressibility” means that a
CP␣ mean values are equal to the corresponding ␣ values, inde- signal f can be represented concisely as a weighted summation of
pendent of either ␭c or ␣. However, the median values are greater a proper type of basis function, such as wavelet functions. In the
than the mean values for high confidence levels, and the differ- mathematical formulation of CS, f is expressed as follows:

Published by NRC Research Press


Pagination not final (cite DOI) / Pagination provisoire (citer le DOI)
Zhao et al. 5

Fig. 3. Effect of correlation length (␭c) on CP␣: (a) box plots for CP50%, CP80%, and CP95%; (b) standard deviation of CP␣ (␴CP␣) as a function of ␣.
(Analytical solution for the case with no correlation (i.e., ␭c = 0) is shown with a solid line.) [Color online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 14.136.36.238 on 12/16/17

(3) f ⫽ B␻ The number S of coefficients needed is obtained by an iteration


procedure using cosine similarity (Wang and Zhao 2017).
where B is a N × N orthonormal matrix composed of columns of Because ␻S follows a multivariate Student t distribution and
pre-specified basis functions, and ␻ is the corresponding weight eq. (5), f̂ is also derived as a random vector following a multivariate
coefficient vector with a length of N. Because of the compressibil- Student t distribution (e.g., Ang and Tang 2007; Fenton and
ity of signals, most entries in ␻ are near to zero. Thus, f can be Griffiths 2008), with 2cn degrees of freedom, mean ␮f̂, and scale
reconstructed by identifying and estimating the weight coeffi- matrix (dn/cn)BHBT. The mean and covariance of f̂ (i.e., ␮f̂ and
For personal use only.

cients with significant value using the sparse measurement data COVf̂, respectively) are derived as
vector y that has a length of M, where M < N, as follows:
dn
(7) ␮f̂ ⫽ B␮␻S COVf̂ ⫽ BCOV␻SBT ⫽ BHBT
(4) y ⫽ ⌿f ⫽ A␻ cn ⫺ 1

where ␺ is a M × N matrix and represents the locations of compo- The BCS procedure has been implemented by a package of user
nents of y in f. A = ␺B is also a M × N matrix (Wang and Zhao 2016). functions in MATLAB (Mathworks 2016). Only the sparse measure-
Exploiting sparsity, the resulting underdetermined system of lin- ment data from site characterization are required to obtain the
ear equations, i.e., eq. (4), can be solved by various existing effi- mean and CI profiles of soil properties of interest.
cient algorithms (e.g., Foucart and Rauhut 2013). For example,
In the following section, the BCS method is used to provide the
Wang and Zhao (2017) used a Bayesian method to statistically
best estimate of a complete soil property profile from sparse mea-
reconstruct the signal f̂, which is an approximation of f. Mathe-
surement data and construct the associated CI profiles. A statisti-
matically, f̂ is defined by
cal meaning of the CI profiles obtained from BCS is proposed: the
corresponding confidence level for a CI profile from BCS is the
(5) f̂ ⫽ B␻S expected coverage proportion (i.e., fraction) of the complete pro-
file that falls within the CI, if all data points along the depth can be
where ␻S is the approximate weight coefficient vector with a measured to provide the complete profile. The statistical meaning
length of N, and all components are set to zero except for the S of the BCS CI profiles is similar to that of the CI profiles for ran-
nontrivial components (S ⬍⬍ N). Following the Bayesian frame- dom field data shown in the section “Coverage proportion of con-
work (Wang and Zhao 2017), the posterior marginal distribution fidence interval profiles”. The statistical meaning of the BCS CI
of ␻S derived from y follows a multivariate Student t distribution profiles will be evaluated systematically in the next section.
with a degree of freedom equal to 2cn and a scale matrix of (dn/cn)H.
The mean vector and covariance matrix of ␻S (i.e., ␮␻S and COV␻S, Coverage proportion of the BCS confidence
respectively) are expressed as interval profiles
Construction of CI profiles obtained from BCS with sparse
dnH
(6) ␮␻S ⫽ HATy COV␻S ⫽ measurement data as input
cn ⫺ 1 As f̂ follows a multivariate Student t distribution (Wang and
Zhao 2017), the upper and lower bounds of the confidence interval
where H = (ATA + D)−1, cn = M/2 + c0, and dn ⫽ 共yTy ⫺ are defined as (e.g., Taboga 2012)
␮␻T
S
H⫺1␮␻S兲/2 ⫹ d0. c0 and d0 are small non-negative constants, e.g.,
c0 = d0 = 10−4, D is a N × N diagonal matrix with components Di,i = (8) CI␣ ⫽ ␮f̂ ± t(1⫺␣)/2,2cn兹(2cn ⫺ 2)/2cn兹diag(COVf̂)
␣i, and ␣i are unknown non-negative coefficients. Note that eq. (6)
only depends on ␣i and requires an iterative algorithm (e.g., the
maximum likelihood estimation) to obtain the most probable where t共1⫺␣兲/2,2cn is the Student t factor for a confidence level ␣ and
value of ␣i. Only the ␣i values corresponding to the S (S ⬍⬍ N) a degree of freedom 2cn, 兹共2cn ⫺ 2兲/2cn is a scaling factor, and
nontrivial coefficients in ␻S need to be estimated in BCS, hence 兹diag共COVf̂兲 is a column vector composed of the square root of
bypassing the possible problem caused by high dimensionality. the diagonal elements of COVf̂. Note that CI␣ is composed of two

Published by NRC Research Press


Pagination not final (cite DOI) / Pagination provisoire (citer le DOI)
6 Can. Geotech. J. Vol. 00, 0000

Fig. 4. Three simulated X profiles and those reconstructed from BCS using M = 20 measurement data points y for CP95% of (a) 94.9%; (b) 94.1%;
(c) 96.3%. [Color online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 14.136.36.238 on 12/16/17
For personal use only.

column vectors, which correspond to the upper and lower bounds flected by the best estimate from only 20 measured data points, on
of the two-tailed data distribution at various depths (e.g., for ␣ = average around 95% of all local variations of the original profile
90%, the lower and upper bounds are the 5th and 95th percentiles, fall within the CI95% upper and lower bounds. For example, the
respectively). CP95% values of the three original RFSs shown in Figs. 4a, 4b, and 4c
Consider, for example, using BCS to reconstruct the three RFSs are 94.9%, 94.1%, and 96.3%, respectively. Hence, about 95% of the
shown in Fig. 1 with only M = 20 measurement data points from three original RFSs shown in Fig. 4 fall within the CI95% upper and
each RFS as input (i.e., the sparse measurement data y). Figure 4 lower bounds. Recall that, in the subsection “Probability distribu-
shows the M = 20 measurement data points by open circles. The tion of CP␣ for CI␣ profiles”, the similar CP95% values evaluated for
best estimate (i.e., mean) of the complete soil property X profile the full set of random field data were 94.5%, 94.9%, and 96.1%,
and 95% CI profiles are shown in Fig. 4 as dashed and dotted lines, respectively. The difference is less than 1% and quite minor.
respectively. Additionally, the original and complete RFS profile is
shown by a solid line in Fig. 4 for comparison. Figure 4 shows that, Probability distribution of CP␣ for BCS CI␣ profiles
although the best estimate (i.e., the dashed line) does not go ex- To evaluate the probability distribution of CP␣ for the BCS CI␣,
actly through the measurement data points (i.e., open circles), it the BCS method is used to construct each of the Ns = 1000 RFSs
follows a trend similar to that of these data points and the original shown in Fig. 1, using a limited number, e.g., M = 20, of measure-
and complete RFS profiles. This suggests that the soil property X ment data points from each RFS as input (i.e., yi, i = 1, 2, …, Ns). This
profile reconstructed from BCS is consistent with the original leads to a total of 1000 best estimates and CI profiles obtained
variation of X with depth, even when only a limited number of from BCS. Each of the 1000 RFSs is used as the original and com-
measurement data points (e.g., 20 from a total of 512 data points) plete soil property X profile, which is compared with the corre-
are used as input. sponding best estimate and CI profiles obtained from each BCS
Some local variations of the original profile are not recon- interpretation. The coverage proportion (CP␣) of each of the 1000 RFSs
structed in the best estimate profile. This is because the number of that fall within the corresponding BCS CI␣ profiles was evaluated for
measurement data is too limited (i.e., M = 20 ⬍⬍ N = 512); hence, different ␣ values ranging from 50% to 95%.
the statistical uncertainty is quite significant. The statistical un- Figures 5a–5c show histograms of CP50%, CP80%, and CP95%, re-
certainty can be explicitly and objectively quantified by the cova- spectively, when M = 20 (i.e., 20/512 = 3.9% or less than 4% of the
riance calculated in eq. (7) and the CI profiles calculated with complete profile is measured). The mean CP50%, CP80%, and CP95%
eq. (8). For example, the bounds of the CI95% profiles are shown in values are shown in Figs. 5a–5c as 0.54, 0.79, and 0.92, respectively.
Fig. 4 by two dotted lines. Similar to the CI profiles for random These mean CP values are quite close to their respective ␣ values
field data discussed previously, the BCS CI profiles have a statisti- (i.e., 0.5, 0.8, and 0.95, respectively). Therefore, the confidence
cal meaning: the corresponding confidence level for a BCS CI level may be interpreted as the expected coverage proportion of
profile is the expected coverage proportion (i.e., fraction) of the the original and complete profile that falls within the correspond-
complete profile that falls within the CI, if all data points over ing BCS CI profiles, if the original and complete profile can be
depth can be measured to provide the complete profile. In other measured. Additionally, similar to the random field data dis-
words, although some details of the original profile are not re- cussed in the section “Coverage proportion of confidence interval

Published by NRC Research Press


Pagination not final (cite DOI) / Pagination provisoire (citer le DOI)
Zhao et al. 7

Fig. 5. Histograms of the coverage proportion (CP␣) for 50%, 80%, and 95% confidence levels. (Red vertical lines show the confidence level.)
[Color online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 14.136.36.238 on 12/16/17
For personal use only.

profiles”, the CP probability distribution is close to symmetric Figure 5 shows histograms of CP50%, CP80%, and CP95% for three
when ␣ = 50% (see Fig. 5a). As ␣ increases and approaches unity, M values (i.e., M = 20, 40, and 60). The mean CP50%, CP80%, and
the distribution becomes less symmetric and presents a negative CP95% values are shown in Figs. 5d–5f for M = 40 as 0.54, 0.8, and
skewness (see Figs. 5b and 5c). Similar to Fig. 2, Fig. 6a shows 0.93, respectively. In Figs. 5g–5i, the mean CP50%, CP80%, and CP95%
box-and-whiskers plots for CP␣ with various ␣ values when M = 20. values for M = 60 are shown to be 0.48, 0.77, and 0.92, respectively.
A 1:1 line is also included in Fig. 6a. All mean CP␣ values at various For M = 20 and 40, the mean CP50% is slightly greater than 50%
␣ levels plot close to the 1:1 line. This demonstrates again that the while the mean value for CP95% is slightly less than 95%. Nonethe-
confidence level may be interpreted as the expected coverage pro- less, these mean CP␣ values are quite close to their respective ␣
portion of the original and complete profile that falls within the values. On average, a proportion ␣ of all local variations of the
corresponding CI profiles obtained from BCS, if this original pro- original profile fall within the corresponding CI␣ profiles, even
file can be measured. Additionally, for relatively large ␣ values, when as few as M = 20 points are used to reconstruct the mean and
the data are negatively skewed, similar to the random field data CI␣ profiles, which is less than 4% of the total data. In Fig. 5 it can
shown in Fig. 2b, and the mean CP␣ value is slightly below ␣. be seen that the variability of the CP␣ values decreases as the M
Effect of the number of measurement data points (M) on CP␣ value increases. In addition, as previously shown for M = 20 and
The BCS method was repeated with different numbers of mea- for the full set of random field data, as ␣ approaches unity the
surement data points (M); namely, M = 10–60, with an increment distribution develops a negative skewness. However, as can be
of 10 points. These values correspond to measurement spacing seen for CP80% and CP95%, as M increases, the distribution ap-
between 30 cm and 2 m, and of fractions equal to 2% to about 12% proaches a normal distribution.
of the complete profile. As M increases, more local variations of Similar to Fig. 6a, Figs. 6b and 6c show box-and-whiskers plots
the original profile are captured by the reconstructed BCS mean for CP␣ with various ␣ values when M = 40 and 60, respectively.
profile, as shown by Wang and Zhao (2017). Additionally, as M 1:1 lines are also included in these figures. As seen for M = 20 in
increases, the bounds of the CI␣ profiles become narrow and ap- Fig. 6a, the mean CP␣ values at the ␣ levels evaluated are close to
proach the mean profile. This reflects that the statistical uncer- the 1:1 line. In Fig. 6 it is also evident that the variability of CP␣
tainty is effectively reduced when more data points are available. values decreases as M increases, e.g., see the size of the boxes in
In this section, the effect of M is evaluated on the coverage pro- Fig. 6. Figure 7 shows the mean values of CP50%, CP80%, and CP95%
portion of the original and complete profile that falls within the for all values of M tested. Even though the statistical uncertainty is
corresponding CI profiles obtained from BCS. reduced with increasing M, the mean CP␣ is not greatly affected

Published by NRC Research Press


Pagination not final (cite DOI) / Pagination provisoire (citer le DOI)
8 Can. Geotech. J. Vol. 00, 0000

Fig. 6. Box-and-whiskers plot for coverage proportion (CP␣) of the original RFS profile within the given BCS CI under different M scenarios:
(a) M = 20, (b) M = 40, and (c) M = 60. [Color online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 14.136.36.238 on 12/16/17

Fig. 7. Effect of number of measurement data points (M) on the Fig. 8. Effect of correlation length (␭c) on the mean CP␣ values.
mean CP␣ values for confidence levels 50%, 80%, and 95%. [Color online.]
For personal use only.

and fluctuates around the ␣ value. The relative difference between method is robust and performs satisfactorily for the possible
the average CP␣ and ␣ is less than 15% in all the cases tested. This range of ␭c values for soil properties reported in the literature. It
demonstrates once again that the confidence level may be inter- is worth noting that it is very difficult to determine the correla-
preted as the expected coverage proportion of the original and tion length (␭c) in engineering practice due to the limited amount
complete profile that falls within the corresponding BCS CI pro- of measurement data. Using the BCS method enables the need to
files. determine the ␭c value to be bypassed and provides the best esti-
mate and CI profiles for soil properties. In addition, note that
Effect of correlation length on CP␣
although an exponential correlation function is used for illustra-
To analyze the effect of the correlation length (␭c) on CP␣ re-
tive examples in this paper, the method proposed in the paper is
sults, new sets of RFSs are generated using a truncated KL expan-
general and equally applicable to other types of auto-correlation
sion with different ␭c. Eight ␭c values were tested ranging from 0.1
functions, and the BCS method performs well for other types of
to 10 m to consider possible values of ␭c for soil properties re-
auto-correlation functions (e.g., Wang et al. 2017b).
ported in the literature (e.g., Phoon and Kulhawy 1999). For each
RFS set, the BCS method was repeated for three values of M, Illustrative example: selection of effective friction
namely M = 20, 40, and 60. Figure 8 shows the mean CP␣ results for
all ␭c and M values tested. When ␭c is large (i.e., when there is a angle profile
smoothly varying random field), the average CP␣ tends to be In this section, the BCS method is demonstrated using a set of
greater than ␣, indicating that the statistical uncertainty reflected real CPT data for selection of the characteristic value of the effec-
in the profiles for the bounds of CI␣ is greater than the variations tive friction angle (␾ ). BCS provides the best-estimate profile of ␾ 
with depth. In other words, a relatively large proportion of many and various CI profiles associated with various confidence levels.
RFSs tested falls inside the corresponding CI␣ profiles. In contrast, These CI profiles may be used by engineers to facilitate determi-
with small values of ␭c, which implies a very variable field, the nation of the characteristic value profile in reliability-based
average CP␣ tends to be smaller than ␣, indicating that a relatively design. To illustrate the proposed method, only some of the nor-
small proportion of the original profile falls inside the corre- malized tip resistance (q) values measured from CPT are used.
sponding CI␣ bounds. The difference between the average CP␣ and First, the BCS method is applied to provide the best estimate and
␣ decreases as M increases, irrespective of whether the ␭c values CI profiles of q. The coverage proportion of the BCS CI profiles is
used in the simulation are small or large. For ␭c values between 0.5 evaluated using the original and complete set of CPT data. Then a
and 2 m, which are common values for soil properties, the relative transformation model is used to relate the q profiles to ␾  profiles.
difference between the average CP␣ and ␣ is less than 15%. The BCS The effect of the transformation model uncertainty on the ␾ 

Published by NRC Research Press


Pagination not final (cite DOI) / Pagination provisoire (citer le DOI)
Zhao et al. 9

Fig. 9. Results of illustrative example: estimation of effective friction angle (␾ ) from normalized CPT tip resistance (q): (a) comparison
between original q profile and q profile reconstructed from BCS; best estimate and CI90% profiles of ␾  (b) without and (c) with consideration of
model uncertainty ␧m. [Color online.]
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 14.136.36.238 on 12/16/17
For personal use only.

profiles is also considered using Monte Carlo simulations. Note original CPT data within the CI90% profiles is CP90% = 85%, which is
that the ␾  profiles at various CI levels may be used by geotechni- close to the expected value of 90%.
cal engineers to facilitate selection of the ␾  characteristic value in The effective friction angle (␾ ) is estimated using a correlation
reliability-based design. model from Kulhawy and Mayne (1990) as follows:
The CPT was performed on the Piedmont soils at the Georgia
Tech campus, Atlanta, in which an extensive program of in situ (9) ␾  ⫽ 17.6 ⫹ 11 logq
and laboratory tests have been carried out for soil property deter-
mination (Mayne and Harris 1993). The BCS method is applied to
the CPT data in the residual silty sand layer between the depths of Figure 9b shows as a solid line the ␾  profile obtained when
3.8 and 19.2 m, approximately (Mayne and Harris 1993). Note that applying eq. (9) to the original and complete set of CPT data.
only one soil layer of residual silty sand is considered in this Similarly, shown as dashed and two dotted lines are the results
illustrative example. Stratification therefore is not needed here. when applying eq. (9) to the profiles of the best estimate and the
However, for a profile with different soil layers, before application bounds of the 90% CI obtained from the BCS procedure, respec-
of the method proposed in this paper, stratification of soil layers tively. For comparison, Figs. 9b also includes lab test results from
shall be performed (if possible) using, e.g., the Bayesian method consolidated undrained triaxial compression tests that were per-
(Cao and Wang 2013; Wang et al. 2013, 2014). In this layer, the cone formed using soil samples at different depths from this site
tip resistance (qc) ranges between 3.3 and 7.3 MPa, and the soil has (Mayne and Harris 1993). The mean ␾  from the 13 triaxial test data
a loose to medium-dense relative density according to Meyerhof points is about 35°, which is similar to the value obtained from
(1956). The soil has an average of 33% fines content, 8% clays, and BCS (i.e., 35.39°). However, the triaxial data present more variabil-
a median grain size (D50) of 0.14 mm. Note that although CPT data ity than that shown by the BCS CI90% profiles.
are used here for illustration and validation, the BCS method Note that eq. (9) was obtained by a semi-log regression on
really aims at the typical situation of sparsely measured data (e.g., 20 data sets from different sites, which cover a site condition similar
standard penetration test data or laboratory test data) in engineer- to that at Georgia Tech campus, Atlanta. A total of 633 data points
ing practice. was used in the regression. Significant residual error over eq. (9)
The cone tip resistance is normalized by the square root of the was reported, and the corresponding standard deviation of the
vertical effective stress (␴v0) as follows: q ⫽ 共qc/pa兲/兹␴v0/pa, where residual error is 2.8° (Kulhawy and Mayne 1990). This residual
pa is the atmospheric pressure. The normalized tip resistance is error can be treated as the model uncertainty of eq. (9), and it may
shown in Fig. 9a as a solid line. The BCS procedure is applied to be included in eq. (9) as an additive zero-mean normally distrib-
15 data points (i.e., M = 15), which represent a sampling interval of uted random variable (␧m). The ␧m in this example is modelled as
about 1 m. These points are also shown in Fig. 9a as open circles. a single random variable to represent a perfectly correlated na-
The best estimate and the bounds of the 90% confidence interval ture of ␧m over depth. To account for both the model uncertainty
(CI90%) obtained from BCS are shown in Fig. 9a by a dashed and and the BCS statistical uncertainty, Monte Carlo simulations were
two dotted lines, respectively. The coverage proportion of the carried out to provide the best estimate and various CI profiles for

Published by NRC Research Press


Pagination not final (cite DOI) / Pagination provisoire (citer le DOI)
10 Can. Geotech. J. Vol. 00, 0000

the effective friction angle. Five thousand random samples of ␧m this, BCS was used to estimate an effective friction angle profile
were generated in the simulations. Each ␧m sample was used to- from sparse CPT data points in a real case history. Furthermore,
gether with eq. (9), and a complete q profile is reconstructed from the uncertainty in the transformation model that relates CPT data
BCS to generate a ␾  profile, leading to 5000 friction angle pro- to effective friction angle can also be considered in the proposed
files. Then, the CI90% ␾  profiles with model uncertainty was eval- method. It is shown that the effective friction angle CI profiles
uated and is shown in Fig. 9c by two gray lines, together with those from the proposed method using sparse CPT data points are con-
without model uncertainty using the same symbols in Fig. 9b. The sistent with those from triaxial tests. Hence, the best estimate and
interval given by the two gray lines (i.e., with model uncertainty) CI profiles from the proposed method may be used to facilitate an
is obviously much bigger than that given by two dotted lines (i.e., objective determination of geotechnical property characteristic
without model uncertainty). It is evident that the model uncer- values from sparse measurement data.
tainty has significant effect on the ␾  profiles. The 13 ␾  data
points from triaxial tests are also included in Fig. 9c. Eleven out of Acknowledgements
13 data points (i.e., 11/13 = 85%) fall within the CI90% ␾  profiles. This The work described in this paper was supported by grants from
is quite consistent with the statistical meaning of CI90% ␾  profiles the Research Grants Council of the Hong Kong Special Adminis-
trative Region, China (Project No. 9042331 (CityU 11225216) and
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 14.136.36.238 on 12/16/17

that about 90% of data points are expected to fall within the cor-
responding bounds. The lower bound of CI90% with both statistical Project No. 8779012 (T22–603/15N)). The financial support is grate-
and model uncertainty represents a 5% fractile of the ␾  profile fully acknowledged.
and might be selected as the characteristic value profile of ␾  in
reliability-based design, if the characteristic value is defined as the References
5% fractile. AASHTO. 1998. LRFD bridge design specifications. American Association of State
Highway and Transportation Officials (AASHTO), Washington, D.C.
It is worth noting that the interpreted profile from the BCS Ang, A.H.-S., and Tang, W.H. 2007. Probability concepts in engineering: empha-
method in this paper has meaning similar to the local estimation sis on applications to civil and environmental engineering. John Wiley and
of a geotechnical property of interest at the location where a Sons, New York.
borehole was drilled (e.g., Honjo and Setiawan 2007; Honjo 2008). Baecher, G.B., and Christian, J.T. 2003. Reliability and statistics in geotechnical
engineering. Wiley, UK.
If global estimation of a geotechnical property (i.e., the geotech- Becker, D.E. 2010. Testing in geotechnical design. Geotechnical Engineering,
nical property within the whole site) is of interest (e.g., Honjo and 41(1).
Setiawan 2007; Honjo 2008), a perfect correlation in the horizon- Bond, A., and Harris, A. 2008. Decoding Eurocode 7. Taylor & Francis, London
tal direction may be assumed. Alternatively, the BCS method may and New York.
Candès, E.J., and Wakin, M.B. 2008. An introduction to compressive sampling.
be extended from one to two dimensions, and the perfect corre-
For personal use only.

IEEE Signal Processing Magazine, 25(2): 21–30. doi:10.1109/MSP.2007.914731.


lation assumption is not needed for the two-dimensional BCS Candès, E.J., Romberg, J.K., and Tao, T. 2006. Stable signal recovery from incom-
method, which is currently under development and beyond the plete and inaccurate measurements. Communications on Pure and Applied
scope of this study. Mathematics, 59(8): 1207–1223. doi:10.1002/cpa.20124.
Cao, Z., and Wang, Y. 2013. Bayesian approach for probabilistic site character-
ization using cone penetration tests. Journal of Geotechnical and Geoenvi-
Conclusions ronmental Engineering, 139(2): 267–276. doi:10.1061/(ASCE)GT.1943-5606.
This paper developed a statistical procedure to facilitate objec- 0000765.
Cao, Z., and Wang, Y. 2014. Bayesian model comparison and characterization of
tive selection of a geotechnical property characteristic value from
undrained shear strength. Journal of Geotechnical and Geoenvironmental
spatially varying, but sparsely measured, data. The proposed pro- Engineering, 140(6). doi:10.1061/(ASCE)GT.1943-5606.0001108.
cedure is based on the BCS method, which is not only able to Cao, Z., Wang, Y., and Li, D. 2016. Quantification of prior knowledge in geotech-
reconstruct the best-estimate profile of a geotechnical property nical site characterization. Engineering Geology, 203: 107–116. doi:10.1016/j.
enggeo.2015.08.018.
from sparse measurement data, but also able to provide CI pro-
CEN. 2004. Eurocode 7: Geotechnical design — Part 1: General rules. EN 1997-1:
files for quantifying the statistical uncertainty associated with the 2004. European Committee for Standardization (CEN), Brussels, Belgium.
interpretation. The quantified uncertainty in BCS has a clear sta- Ching, J.Y., and Phoon, K.-K. 2013a. Mobilized shear strength of spatially variable
tistical meaning: the corresponding confidence level for BCS CI is soils under simple stress states. Structural Safety, 41: 20–28. doi:10.1016/j.
the expected coverage proportion (i.e., fraction) of the complete strusafe.2012.10.001.
Ching, J.Y., and Phoon, K.-K. 2013b. Probability distribution for mobilized shear
profile that falls within the CI, if all data points over the depth can strengths of spatially variable soils under uniform stress states. Georisk, 7(3):
be measured to provide the complete profile. 209–224.
The statistical meaning of CI was firstly illustrated using ran- Ching, J.Y., Phoon, K.-K., and Kao, P.H. 2014. Probability model for overall shear
dom field data. When a large number of complete sets of RFSs are strength of spatially variable soils under uniform stress states. Journal of
Engineering Mechanics, ASCE, 140(3): 487–501. doi:10.1061/(ASCE)EM.1943-
used, the expected coverage proportion (CP␣) for a confidence 7889.0000667.
interval with a confidence level ␣ (CI␣) is equal to ␣. In addition, Ching, J.Y., Hu, Y.G., and Phoon, K.-K. 2016a. On characterizing spatially variable
when only a limited number of data points from the RFS are shear strength using spatial average. Probabilistic Engineering Mechanics,
measured, the proposed BCS method can be used to reconstruct 45: 31–43. doi:10.1016/j.probengmech.2016.02.006.
Ching, J.Y., Lee, S.W., and Phoon, K.-K. 2016b. Undrained strength for a 3D
the best estimate and CI␣ profiles of the complete set of RFS val- spatially variable clay column subjected to compression or shear. Probabilis-
ues. It is shown that on average, the BCS CP␣ is close to ␣ even if tic Engineering Mechanics, 45: 127–139. doi:10.1016/j.probengmech.2016.03.
only about 2% of the data points from the original and complete 002.
RFS are measured. As more data points are available, the statisti- CSA. 2014. Canadian highway bridge design code. CAN/CSA S614:2014. Canadian
Standards Association, Mississauga, Ont.
cal uncertainty is reduced and the variability in CP␣ also reduces, De Veaux, R.D., Velleman, P.F., and Bock, D.E. 2014. Intro stats. 4th ed. Pearson.
but the average value is only slightly affected. In addition, the Det Norske Veritas. 2010. Statistical representation of soil data. DNV-RP-C207.
effect of the correlation length (␭c) of the random field on the Det Norske Veritas.
average CP␣ was also investigated. It is shown that the proposed Efron, B., and Tibshirani, R.J. 1993. An introduction to the bootstrap. Chapman
& Hall/CRC, Boca Raton, Fla. doi:10.1111/1467-9639.00050.
method is robust and performs satisfactorily for the typical range Fenton, G.A., and Griffiths, D.V. 2008. Risk assessment in geotechnical engineer-
of ␭c values for soil properties reported in the literature. ing. John Wiley and Sons, Inc. doi:10.1002/9780470284704.
For geotechnical engineering applications, the complete set of Fenton, G.A., and Naghibi, M. 2011. Geotechnical resistance factors for ultimate
data (i.e., a high-resolution measurement data profile over depth) limit state design of deep foundations in frictional soils. Canadian Geotech-
nical Journal, 48(11): 1742–1756. doi:10.1139/t11-068.
is often not available, and the BCS method can be used to not only Foucart, S., and Rauhut, H. 2013. A mathematical introduction to compressive
provide the best-estimate profiles from sparse measurement data, sensing. Applied and numerical harmonic analysis book series. Birkhäuser,
but also offer various confidence interval profiles. To illustrate New York. doi:10.1007/978-0-8176-4948-7.

Published by NRC Research Press


Pagination not final (cite DOI) / Pagination provisoire (citer le DOI)
Zhao et al. 11

Gong, W., Khoshnevisan, S., and Juang, C.H. 2014. Gradient-based design robust- Reddy, S.C., and Stuedlein, A.W. 2017. Ultimate limit state reliability-based de-
ness measure for robust geotechnical design. Canadian Geotechnical Jour- sign of augered cast-in-place pile considering lower-bound capacities. Cana-
nal, 51(11): 1331–1342. doi:10.1139/cgj-2013-0428. dian Geotechnical Journal, 54(12): 1693–1703. doi:10.1139/cgj-2016-0145.
Honjo, Y. 2008. General vs. local reliability based design in geotechnical engi- Schneider, H.R., and Schneider, M.A. 2013. Dealing with uncertainties in EC7
neering. In Proceedings of the 4th Asian-Pacific Symposium on Structural with emphasis on determination of characteristic soil properties. In Modern
Reliability and its Applications (APSSRA’08), Hong Kong, 19–20 June. geotechnical design codes of practice. Edited by P. Arnold et al., ISO Press.
pp. 41–52. pp. 87–101.
Honjo, Y., and Setiawan, B. 2007. General and local estimation of local average Taboga, M. 2012. Lectures on probability theory and mathematical statistics.
and their application in geotechnical parameter estimations. Georisk, 1(3): 2nd ed. CreateSpace Independent Publishing Platform.
167–176. doi:10.1080/17499510701745960. Vanmarcke, E.H. 1977. Probabilistic Modeling of Soil Profiles. Journal of the
Ji, S., Xue, Y., and Carin, L. 2008. Bayesian compressive sensing. IEEE Transac- Geotechnical Engineering Division, ASCE, 103(11): 1227–1246.
tions on Signal Processing, 56(6): 2346–2356. doi:10.1109/TSP. 2007.914345. Vick, S. 2002. Degrees of belief - subjective probability and engineering judg-
Kulhawy, F.H., and Mayne, P.W. 1990. Manual on estimating soil properties for ment. ASCE Press.
foundation design. Ithaca, New York. Wahba, G. 1983. Bayesian “confidence intervals” for the cross-validated smooth
Li, D.Q., Qi, X.H., Phoon, K.K., Zhang, L.M., and Zhou, C.B. 2014. Effect of spatially spline. Journal of the Royal Statistical Society. Series B: Statistical Methodol-
variable shear strength parameters with linearly increasing mean trend on ogy, 45(1): 133–150.
reliability of infinite slopes. Structural Safety, 49: 45–55. doi:10.1016/j.strusafe. Wang, Y., and Aladejare, A.E. 2015. Selection of site-specific regression model for
2013.08.005. characterization of uniaxial compressive strength of rock. International
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by 14.136.36.238 on 12/16/17

Li, D.Q., Shao, K.B., Cao, Z.J., Tang, X.S., and Phoon, K.K. 2016. A generalized Journal of Rock Mechanics & Mining Sciences, 75: 73–81. doi:10.1016/j.ijrmms.
surrogate response aided-subset simulation approach for efficient geotech- 2015.01.008.
nical reliability-based design. Computers and Geotechnics, 74: 88–101. doi: Wang, Y., and Aladejare, A.E. 2016. Evaluating variability and uncertainty of
10.1016/j.compgeo.2015.12.010. Geological Strength Index at a specific site. Rock Mechanics and Rock Engi-
Marra, G., and Wood, S.N. 2012. Coverage properties of confidence intervals for neering, 49(9): 3559–3573. doi:10.1007/s00603-016-0957-5.
generalized additive model components. Scandinavian Journal of Statistics, Wang, Y., and Cao, Z. 2013. Probabilistic characterization of Young’s modulus of
39(1): 53–74. doi:10.1111/j.1467-9469.2011.00760.x. soil using equivalent samples. Engineering Geology, 159: 106–118. doi:10.1016/
Mathworks. 2016. MATLAB – The language of technical computing. j.enggeo.2013.03.017.
Mayne, P.W., and Harris, D.E. 1993. Axial load-displacement behavior of drilled Wang, Y., and Zhao, T. 2016. Interpretation of soil property profile from limited
shaft foundations in Piedmont Residuum. Report prepared for FHWA, measurement data: a compressive sampling perspective. Canadian Geotech-
McLean, Va. nical Journal, 53(9): 1547–1559. doi:10.1139/cgj-2015-0545.
Meyerhof, G.G. 1956. Penetration tests and bearing capacity of cohesionless Wang, Y., and Zhao, T. 2017. Statistical interpretation of soil property profiles
soils. Journal of the Soil Mechanics and Foundations Division, ASCE, 82(SM1): from sparse data using Bayesian compressive sampling. Géotechnique, 67(6):
1–19. 523–536. doi:10.1680/jgeot.16.P.143.
Meyerhof, G.G. 1995. Development of geotechnical limit state design. Canadian Wang, Y., Huang, K., and Cao, Z. 2013. Probabilistic identification of under-
Geotechnical Journal, 32(1): 128–136. doi:10.1139/t95-010. ground soil stratification using cone penetration tests. Canadian Geotechni-
Nychka, D. 1988. Bayesian confidence intervals for smoothing splines. Journal of cal Journal, 50(7): 766–776. doi:10.1139/cgj-2013-0004.
the American Statistical Association, 83(404): 1134–1143. doi:10.1080/01621459. Wang, Y., Huang, K., and Cao, Z. 2014. Bayesian identification of soil strata in
For personal use only.

1988.10478711. London clay. Géotechnique, 64(3): 239. doi:10.1680/geot.13.T.018.


Orr, T.L.L. 2017. Defining and selecting characteristic values of geotechnical Wang, Y., Akeju, O.V., and Cao, Z. 2016a. Bayesian Equivalent Sample Toolkit
parameters for designs to Eurocode 7. Georisk: Assessment and Management (BEST): an Excel VBA program for probabilistic characterisation of geotech-
of Risk for Engineered Systems and Geohazards, 11(1): 103–115. doi:10.1080/ nical properties from limited observation data. Georisk: Assessment and
17499518.2016.1235711. Management of Risk for Engineered Systems and Geohazards, 10(4): 251–268.
Phoon, K.-K., and Ching, J. 2014. Risk and reliability in geotechnical engineering. Wang, Y., Cao, Z., and Li, D. 2016b. Bayesian perspective on geotechnical vari-
CRC Press, Taylor & Francis Group, Boca Raton, Fla. ability and site characterization. Engineering Geology, 203: 117–125. doi:10.
Phoon, K.-K., and Kulhawy, F.H. 1999. Characterization of geotechnical variability. 1016/j.enggeo.2015.08.017.
Canadian Geotechnical Journal, 36(4): 612–624. doi:10.1139/t99-038. Wang, Y., Akeju, O.V., and Zhao, T. 2017a. Interpolation of spatially varying but
Phoon, K.-K., Huang, S.P., and Quek, S.T. 2002. Implementation of Karhunen- sparsely measured geo-data: a comparative study. Engineering Geology, 231:
Loeve expansion for simulation using a wavelet-Galerkin scheme. Probabilistic 200–217. doi:10.1016/j.enggeo.2017.10.019.
Engineering Mechanics, 17(3): 293–303. doi:10.1016/S0266-8920(02)00013-9. Wang, Y., Zhao, T., and Phoon, K.-K. 2017b. Direct simulation of random field
Phoon, K.-K., Huang, H.W., and Quek, S.T. 2005. Simulation of strongly non- samples from sparsely measured geotechnical data with consideration of
Gaussian processes using Karhunen-Loeve expansion. Probabilistic Engineer- uncertainty in interpretation. Canadian Geotechnical Journal. [Published on-
ing Mechanics, 20(2): 188–198. doi:10.1016/j.probengmech.2005.05.007. line ahead of print 27 October 2017.] doi:10.1139/cgj-2017-0254.
Phoon, K.-K., Prakoso, W.A., Wang, Y., and Ching, J. 2016. Uncertainty represen- Zhang, J., and Ellingwood, B. 1994. Orthogonal series expansions of random
tation of geotechnical design parameters. Chapter 3. In Reliability of geotech- fields in reliability analysis. Journal of Engineering Mechanics, 120(12): 2660–
nical structures in ISO2394. pp. 49–88. doi:10.1201/9781315364179-4. 2667. doi:10.1061/(ASCE)0733-9399(1994)120:12(2660).

Published by NRC Research Press

You might also like