Professional Documents
Culture Documents
ARTICLE
Probabilistic identication of underground soil stratication using
cone penetration tests
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by UNIV LIBRARY OF HONG KONG UNIV OF on 04/03/16
Abstract: This paper develops Bayesian approaches for underground soil stratum identication and soil classication using
cone penetration tests (CPTs). The uncertainty in the CPT-based soil classication using the Robertson chart is modeled
explicitly in the Bayesian approaches, and the probability that the soil belongs to one of the nine soil types in the Robertson
chart based on a set of CPT data is formulated using the maximum entropy principle. The proposed Bayesian approaches
contain two major components: a Bayesian model class selection approach to identify the most probable number of
underground soil layers and a Bayesian system identication approach to simultaneously estimate the most probable layer
thicknesses and classify the soil types. Equations are derived for the Bayesian approaches, and the proposed approaches are
illustrated using a real-life CPT performed at the National Geotechnical Experimentation Site (NGES) at Texas A&M
University, USA. It has been shown that the proposed approaches properly identify the underground soil stratication and
classify the soil type of each layer. In addition, as the number of model classes increases, the Bayesian model class selection
approach identies the soil layers progressively, starting from the statistically most signicant boundary and gradually
zooming into less signicant ones with improved resolution. Furthermore, it is found that the evolution of the identied
soil strata as the model class increases provides additional valuable information for assisting in the interpretation of CPT
data in a rational and transparent manner.
Key words: Bayesian system identication, Bayesian model class selection, soil stratigraphy, cone penetration test (CPT), maxi-
For personal use only.
Rsum : Cet article prsente le dveloppement dapproches baysiennes pour lidentication des couches de sol souterrain
et pour la classication des sols a laide dessais de pntration du cne (EPC). Lincertitude lie a la classication base sur
les EPC selon la charte de Robertson est modlise explicitement dans les approches baysiennes, et la probabilit quun sol
appartienne a un des neuf types de sols de la charte de Robertson bas sur une srie de donnes dEPC est formule en
utilisant le principe de lentropie maximale. Les approches baysiennes proposes comprennent deux composantes ma-
jeures : une approche baysienne de slection de classes de modle an didentier le nombre le plus probable de couches
souterraines de sol, et un systme didentication baysien pour estimer les paisseurs les plus probables des couches de
sols et pour classier les types de sols simultanment. Des quations sont drives pour les approches baysiennes, et les
approches proposes sont illustre par un essai rel de pntration du cne ralis au site national dexprimentation
gotechnique (NGES) a luniversit A&M au Texas, tats-Unis. Il a t dmontr que les approches proposes identient
adquatement la stratication du sol et classient le type de sol pour chaque couche. De plus, au fur et a mesure que le
nombre de classes de modle augmente, lapproche baysienne de slection de classes de modle identie les couches de
sol progressivement, en commenant par la frontire la plus signicative statistiquement et dnissant graduellement les
frontires moins signicatives avec une rsolution amliore. Il est aussi dmontr que lvolution des couches de sol
identies quand la classe de modle augmente permet dobtenir de linformation importante pour aider a interprter les
donnes dEPC de faon rationnelle et transparente. [Traduit par la Rdaction]
Mots-cls : systme didentication baysien, slection baysienne de classe de modle, stratigraphie du sol, essais de pntration
du cne (EPC), principe de lentropie maximale.
Can. Geotech. J. 50: 766776 (2013) dx.doi.org/10.1139/cgj-2013-0004 Published at www.nrcresearchpress.com/cgj on 9 May 2013.
Wang et al. 767
shown in Fig. 1 and frequently referred to as the Robertson chart Fig. 1. Robertson soil classication chart based on CPT data (after
(Robertson 1990). The soil is classied based on two parameters Robertson 1990).
(i.e., two axes in Fig. 1): the normalized friction ratio, FR = 100fs/
(q t v0), and the normalized tip resistance, Q t (qt v0)/v0, ob- 1000
tained from CPT testing data, where fs, q t, v0, and v0 are the
sleeve friction, corrected tip resistance, vertical total stress, and 7 8
vertical effective stress, respectively. The chart is divided into nine
areas corresponding to nine different soil types, as described in
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by UNIV LIBRARY OF HONG KONG UNIV OF on 04/03/16
9
Table 1. Based on which area the measured (FR, Q t) combination is
located in, the corresponding soil type is determined accordingly.
6
Normalized Qt = qt v 0
Note that the Robertson chart was developed based on past obser- 100
'v 0
vations and engineering experience, and it inevitably contains
various uncertainties, such as observation scatterness, measure-
ment error, and transformation uncertainty. Such uncertainties
can be properly taken into consideration by integrating the 5
cone resistanc e
Robertson chart with probabilistic approaches.
Several probabilistic soil classication approaches have been
developed in geotechnical literature (e.g., Zhang and Tumay 1999; 10 4
Kurup and Grifn 2006; Jung et al. 2008; Cetin and Ozan 2009).
However, it should be pointed out that these previous approaches, 3
both deterministic (e.g., the Robertson chart) and probabilistic, 1
focused mainly on determining the soil type from a particular data
point (i.e., at a specic depth during CPT). An important question
of how to stratify the underground soil prole and identify differ- 2
ent soil strata from a large number of nearly continuous CPT data
1
points remains open. Consider, for example, measurement data
of a real-life CPT shown in Fig. 2 and performed at the National 0.1 1 10
Geotechnical Experimentation Site (NGES) at Texas A&M Univer- Normalized F = fs
sity (Zhang and Tumay 1999; Briaud 2000). The variation of the R 100
qt v 0
For personal use only.
friction ratio
logarithm of the normalized friction ratio ln(FR) and tip resistance
ln(Q t) with depth are shown in the rst and second column, re-
spectively, in Fig. 2, and the soil types are determined using the Table 1. Description of soil types in the Robertson's soil
Robertson chart and shown in the third column of Fig. 2. If the classication chart (after Robertson 1990).
variation of soil types over depth shown in Fig. 2 is used to identify
Area Soil description
the soil strata, this about 15 m deep soil prole is divided into
about 14 soil layers. This is obviously not practical in engineering 1 Sensitive, ne-grained
applications, and the so-called sound engineering judgment is 2 Organic soils (peats)
frequently exercised in a vague and unquantiable manner to 3 Clays (clay to silty clay)
reduce the number of soil layers. As different engineers might 4 Silt mixtures (clayey silt to silty clay)
have different engineering judgment, the identied soil strata can 5 Sand mixtures (silty sand to sandy silt)
be quite different for each of them. The key issue to this problem 6 Sands (clean sand to silty sand)
is that the uncertainty in the CPT-based soil classication and the 7 Gravelly sand to sand
8 Very stiff sand to clayey sand
spatial distribution of the CPT data (e.g., Fenton 1999a, 1999b;
9 Very stiff, ne-grained
Phoon et al. 2003; Wang et al. 2010) is not considered explicitly,
but treated implicitly and vaguely under the name of engineering
judgement. _h [h1,h2,...,hn,...hN] in a soil prole containing N soil layers, where
This paper integrates the Robertson chart with Bayesian ap- hn is the thickness of the nth layer. The number N in the rst part
proaches to explicitly and properly consider the uncertainty in of this paper (i.e., before the section titled Most probable number
the CPT-based soil classication and spatial distribution of the of soil strata) is considered deterministic, but unknown. It is then
CPT data. Bayesian system identication and model class selection determined by a Bayesian model class selection approach devel-
approaches are developed to integrate engineering judgment as oped in that section of the paper. Let _ [_1,_2,...,_N] be a set of
the so-called prior information with project-specic test data in a ln(FR) and ln(Q t) data obtained from a CPT test in a soil prole
transparent and quantitative manner and to identify the most with N soil layers, in which _n [_n1(d1),_n2(d2),...,_nkn(dkn)], n
probable thicknessesboundaries and number of soil layers in a 1,2,...,N, is a set of ln(FR) and ln(Q t) measured at the kn depths of
systematic and rational manner. The paper starts with the pro- d1,d2,...,dkn within the nth soil layer. As the Robertson chart is
posed probabilistic framework for soil stratum identication and plotted in a loglog scale (see Fig. 1), ln(FR) and ln(Q t) are used
probabilistic characterization of soil types based on the Robertson herein for mathematical convenience.
chart, followed by the Bayesian system identication and model For a given number N of soil layers, the probability P(_nN) that all
class selection approaches. Then, the implementation procedure data points in the nth layer belong to the same soil type (i.e., any
for the Bayesian approaches is described. The proposed ap- one of the nine soil types specied in the Robertson chart, see
proaches are illustrated using a real-life CPT test performed at the Fig. 1 and Table 1) can be expressed as the sum of the probabilities
NGES at Texas A&M University, USA. that all data points belong to each of the nine soil types specied
in the Robertson chart
Probabilistic framework for soil stratum
identification 9
Fig. 2. A set of real CPT results at NGES, Texas A&M University (after Zhang and Tumay 1999).
2
Clay
4
Silty Clay
6
Depth (m)
GWT
12
14
For personal use only.
layer belonging to the soil type J is the key input variable in eqs. (1)
and (2), and a probabilistic model is developed in the next section
to estimate the PSTJ(_n,iN) value.
Layer n : n = [ FR,n , Qt ,n ] hn Probabilistic soil classification based on the
Robertson chart
As the Robertson chart was developed based on past observa-
tions and engineering experience, it inevitably contains various
. . .
Fig. 4. Illustration of the probability density contour and two- Fig. 5. Best-tting of the Robertson soil classication chart by
dimensional joint PDF for soil classication based on a given CPT quadratic functions.
data point.
6.9 Q O N S R
6.9 i i
[ln( F ), ln( Q )] 8
R t
6 7 VIII
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by UNIV LIBRARY OF HONG KONG UNIV OF on 04/03/16
P L
5 9
9 J
5 6
V H
4
4 G
M 5 IV
5 3
3 III
C K 4
4 2
I E
2 1 T 3
1 3 1 II
Robertson (1990)
1 2
A Best-fitted curves F
2 B
0 D
-2.3 -1.3 -0.3 0.7 1.7 2.3
0
Logarithm of normalized friction
For personal use only.
[ln(Qt) ln(Q it)]2
2
2Qt all boundaries (see the labels A, B, C, , T in Fig. 5). These coordi-
nates, together with the quadratic functions, form the boundaries
for the integration, which can be performed numerically and con-
veniently using computer software, such as Matlab (Mathworks,
where _ [FR,Qt], FR and Qt are the standard deviation of the Inc. 2010).
joint Gaussian distribution along the ln(FR) and ln(Q t) axes, respec- Note that in this study, _ [FR,Qt] are the so-called nuisance
tively, and they reect the uncertainties in both the Robertson parameters in the Bayesian analysis, which are quantities neces-
chart and measurement data. Ideally, the values of FR and Qt sarily included in the formulation and analysis, but of no intrinsic
can be estimated based on experimentation and past knowl- interest (Sivia and Skilling 2006). In addition, as illustrated in
edge or experience with CPT measurements, if available. Then, Fig. 3, for a soil prole containing N soil layers, a set of _
the probability PSTJ that the data point i with the measured [FR,Qt] is needed for each soil layer, resulting in N sets of
[ln(F Ri ), ln(Q it)] values belongs to the soil type J is expressed as _ [FR,Qt]. These N sets of model parameters are dened as
_ N [_1,_2,...,_N], in which
_ n [FR,n,Qt,n], n 1,2,...,N, is the
(4) PSTJ P[ln(FR), ln(Qt)|_]d ln(FR)d ln(Qt) standard deviation of the joint Gaussian distribution for the nth
J
layer.
for J 1, 2, ..., 9 The parameters of real interest in the identication of un-
derground soil stratication are the thickness vector h _
The two-dimensional integration in eq. (4) is performed [h1,h2,...,hn,...hN] and its dimension N (i.e., the total number of soil
repeatedly for each soil type (i.e., for [ln(FR), ln(Q t)] soil layers). It is worthwhile to point out that dividing the CPT data _ into
type J when J varies from 1 to 9). For a set of CPT data different soil layers (i.e., dividing _ into _n, n = 1,2,, N) requires
Table 2. Best-tted parameters for the quadratic func- joint posterior distribution (e.g., Ang and Tang 2007; Cao and
tion ln(Q t) = a ln(FR)2 + b ln(FR) + c. Wang 2013)
Function
(5) P(
_ N | _, N) KNP(_ |
_ N, N)P(
_ N |N)
ID a b c
I 0.3707 1.3625 1.0549
II 0.5586 0.5399 0.3049 where KN 1/P(_N) is a normalizing constant; P(_
_ N,N) is the
III 0.5405 0.2739 1.6959 likelihood function that reects the model t with the site obser-
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by UNIV LIBRARY OF HONG KONG UNIV OF on 04/03/16
P(_ |_ , N)
C 2.3026 2.2268
D 2.3026 0 (6) P(_|
_ N, N) n n
E 2.3026 2.0234 n1
F 0.5589 0.1776
P (_n_n,N) is calculated using eqs. (1)(4). Note that, following
G 2.3026 3.9639
H 1.8687 4.0953
J 1.4505 4.5104 the conventional notation of Bayesian framework, P(_nN) in eq. (1)
K 1.3334 2.2126 is re-written as P(_n_n,N) in eq. (6) with the parameters _n for
L 0.9622 5.3534 the given Gaussian PDF in eq. (3). In addition, as the boundaries
M 2.3026 3.4335 of soil layers (layer thicknesses) change, the division of CPT
For personal use only.
1 1
for FR [FR,min , FR,max ], Qt [Qt,min , Qt,max ]
(8) P(_n |N) FR,max FR,min Qt,max Qt,min
0 for others
where FR,min, FR,max; Qt,min, Qt,max are the lower and upper If there is no uncertainty associated with the Robertson chart or
bound values of FR and Qt, respectively. Note that the uniform measurement data, and the soil type can be classied, with full
prior distribution only requires relatively uninformative prior in- certainty, by the measured values of ln(FR) and ln(Q t), the PDF in
formation (i.e., the possible ranges of the model parameters,
_ min Fig. 4 that quanties the occurrence probability of different soil
and _max ). However, it is worthwhile to note that if high-quality
types should be reduced to a single point at [ln(F Ri ),ln(Qit)]. This
prior information is available (e.g., via experimentation and past
knowledge or experience with CPT measurements), more sophis- leads to the lower bound FR and Qt values of FR,min = 0 and
ticated types of prior distributions may be used. The Bayesian Qt,min = 0 for eq. (8). On the other hand, the upper bound values
approach proposed in this paper is general and equally applicable FR,max and Qt,max are considered to occur in an extreme case
for other types of prior distributions. where the Robertson chart provides no valuable information
for soil classication. In such an extreme case, although number k*, therefore, can be determined by comparing the con-
[ln(F Ri ),ln(Qit)] is obtained from the CPT tests, the occurrence prob- ditional probabilities P(Mk_) for all candidate model classes and
ability of different soil types are still similar, and the PDF in Fig. 4
selecting the one with the maximum value of P(Mk_).
can be approximated as a uniform distribution that spreads over
According to Bayes' theorem (Beck and Yuen 2004; Yuen 2010;
the whole Robertson chart. FR and Qt for such a uniform distri-
Cao and Wang 2013), P(Mk_) is written as
bution is taken as the upper bound values FR,max and Qt,max, and
they are expressed as
(11) P(Mk | _) P(_ |Mk)P(Mk)/P(_), k 1, 2, ..., Nmax
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by UNIV LIBRARY OF HONG KONG UNIV OF on 04/03/16
(9)
FR,max [ln(FR,max ) ln(FR,min )]2 /12
where P(_) is the PDF of _ and it is constant and independent of Mk;
Qt,max [ln(Qt,max ) ln(Qt,min )]2 /12
P(_Mk) is the conditional probability of _ given the model class Mk;
P(Mk) is the prior probability of the model class Mk, which reects
where FR,max = 10, FR,min = 0.1, Qt,max = 1000, Qt,min = 1 are the the prior knowledge on the number of the underground soil lay-
maximum and minimum value of FR and Q t in the Robertson
ers. P(_Mk) is frequently referred to as the evidence for the
chart, respectively (see Fig. 1).
model class Mk provided by the CPT data _, and it increases as the
Combining eq. (9) with eq. (8), together with FR,max = 10, FR,min =
plausibility of _ conditional on Mk increases. If no prevailing prior
0.1, Qt,max = 1000, Qt,min = 1, FR,min = 0, and Qt,min = 0, leads to the
knowledge on the number of soil layers is available, the prior
prior distribution for the model parameters _n as
probability for each of the Nmax candidate model classes is the
same and therefore, P(Mk) can be taken as 1/Nmax. Then, according
(10) P(_n |N) ( 1.331 )( 1.991 ) for FR [0, 1.33], Qt [0, 1.99]
to eq. (11), P(Mk_) is proportional to P(_Mk), and the most probable
0 for others number of soil layer can be selected by comparing the value of
P(_Mk) among the Nmax candidate model classes. The model class
Then, the most probable thicknessesboundaries of the soil layers with the maximum value of P(_Mk) is taken as the most probable
are identied by using an asymptotic technique to approximate the model class.
posterior PDF of the model parameters in eq. (5). The asymptotic Using conditional probability and the theorem of total proba-
technique involves approximating the posterior PDF as a Gaussian bility, the evidence P(_Mk) can be expressed as
PDF (Bleistein and Handelsman 1986). The posterior PDF for the
For personal use only.
Yes
ent model classes, the most probable model class Mk and the proba- 4. Compute the likelihood function:
ble number k* of soil layers are determined. Note that the most
Compute the PDF (i.e., eq. (3)) for a CPT data point (i.e., a
probable thicknesses _hk for each model class have been determined
in the Bayesian system identication approach and are used as input combination of [ln(FR), ln(Q t)]) measured at a given depth
in the Bayesian model class selection. Therefore, the determination during CPT tests
Compute the probability that all the data points in the nth
of the most probable number k* of soil layers simultaneously leads to
the determination of the most probable layer thicknesses _hk. layer belong to a certain soil type J by eqs. (2) and (4).
Compute the probability that all the data points in the nth
Implementation procedure layer belong to the same type of soil by eq. (1).
Figure 6 shows a owchart of the proposed approach, which Compute the likelihood function by eq. (6).
involves eight steps in the implementation. Details of each step
and their associated equations are summarized as follows:
_ N_,N)].
5. Construct the objective function fobj ln[P(
1. Obtain a set of CPT data and convert them to the logarithm of the 6. Minimize the objective function (e.g., by a MATLAB function
normalized friction ratio ln(FR) and cone tip resistance ln(Q t).
fminsearch) and determine the most probable thicknesses h_k for
2. Choose an appropriate maximum number Nmax of soil layers
the kth model class. Compare PSTJ(_n_n,N), J 1,2...,9 for the nth
for the CPT data _, resulting in Nmax candidate model classes
and the k value varying from 1 to Nmax. layer and determine the soil type for the nth soil layer.
3. Dene the prior distribution for model parameters 7. Calculate the conditional probability P(_Mk) of _ given the kth
_ N [_1,_2,...,_N] using eqs. (7) and (10). model class using eq. (12).
8. Repeat steps 37 Nmax times to calculate _hk and P(_Mk) for in Fig. 7. Distribution of the CPT data of the illustrative example on
total Nmax candidate model classes. The model class with the the Robertson chart.
maximum value of P(_Mk) is selected as the most probable 6.9
model class Mk and the corresponding _hk values are the most
probable layer thicknessesboundaries.
Although the proposed Bayesian approach involves quite a
6 7 8
number of equations and seems mathematically complicated, it is
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by UNIV LIBRARY OF HONG KONG UNIV OF on 04/03/16
8.2-11m
1999; Briaud 2000). The site comprises a sequence of sandy clay, 2
clay, silty clay, and clay with silt seams extending from the ground 11-15m
surface to a depth of about 15 m. The groundwater table is at about 0
6 m below the ground surface. The distribution of the measured -2.3 -1.3 -0.3 0.7 1.7 2.3
ln(FR) and ln(Q t) data and the soil type of all the CPT data points
Logarithm of normalized friction ratio, ln(FR)
based on the Robertson chart are shown in Fig. 2. As shown in this
gure, the soil prole mainly comprises soil types 3, 4, 5, 6, 8, and
9 (see Table 1 for the description of each soil type). Figure 7 plots
the ln(FR) and ln(Q t) data pairs in the Robertson chart, and the data Table 4. Results of the Bayesian model class selection approach for
points are mainly located within areas 3, 4, and 5, with some the illustrative example.
scattered data points in areas 6, 8, and 9. As a reference, the soil Most probable
stratication based on the boring log at the same site (Zhang and thicknesses, hk (m)
Tumay 1999; Briaud 2000) is also illustrated in Fig. 2. The soil
prole includes a sandy clay layer from about 0 to 1.1 m, a clay Model
layer from about 1.1 to 4.9 m, a silty clay layer from about 4.9 to class Mk
ln[P(_ Mk)] h1 h2 h3 h4 h5 h6
7.1 m, and a clay with silt seams layer from about 7.1 to 15 m. M1 210.4 15
Consider, for example, six candidate model classes (i.e., the M2 147.2 6.9 8.1
maximum number of soil layers Nmax = 6). The possible numbers M3 111.2 1.4 5.4 8.2
(i.e., k) of soil layers are therefore 1, 2, 3, 4, 5, and 6. For all model M4 98.2 1.0 5.8 4.2 4.0
classes, the prior distributions of the model parameters M5 86.2 1.4 5.3 1.5 2.8 4.0
_ N are all
estimated using eqs. (7) and (10). Using the prior knowledge and M6 93.3 1.5 3.8 1.5 1.5 2.9 3.8
the CPT data shown in Fig. 2, the proposed Bayesian system iden-
tication and model class selection approaches are used to pro-
vide the most probable number of soil layers and the most
addition, the soil types in these ve soil layers are classied and
probable boundaries (or thicknesses) of the layers. The results
listed in order from the ground surface as soil type 8 (i.e., very
obtained from the Bayesian approaches are discussed and com-
stiff sand to clayey sand), soil type 3 (clays (clay to silty clay)),
pared with those from other approaches, as described in the fol-
soil type 5 (sand mixtures (silty sand to sandy silt)), soil type 4
lowing subsections.
(silt mixtures (clayey silt to silty clay)), and soil type 5 (sand
Most probable number of soil strata mixtures (silty sand to sandy silt)).
Table 4 summarizes the logarithm of evidence in the second
Most probable thicknesses or boundaries
column for the six candidate model classes. The value of
Table 4 also summarizes the most probable thicknesses (i.e., hk)
ln[P(_Mk)] increases from 210.4 to 86.2 as k increases from 1 to of the soil layers for the six candidate model classes. Using
5, and it then decreases from 86.2 to 93.3 as k further in- hk, k 1, 2, ..., 6 the most probable boundaries of soil layers are
creases from 5 to 6. The model class with ve soil layers has the delineated accordingly for the six model classes, as shown in Fig. 8
maximum value of ln[P(_Mk)] among all the six model classes. by short dashed lines. For the most probable model class M5 , the
Therefore, the most probable number of soil layers is 5. In most probable thicknesses of these ve layers are 1.4, 5.3, 1.5, 2.8,
Fig. 8. Most probable boundaries of soil layers for different model classes in the illustrative example.
4 Soil type
3
6
Depth (m)
Soil type
8 5
Soil type
10 4
12
Soil type
5
14
Note: 3: Clays (clay to silty clay), 4: Silt mixtures (clayey silt to silty clay)
5: Sand mixtures (silty sand to sandy silt), 8: Very stiff sand to clayey sand
For personal use only.
and 4.0 m. The most probable boundaries for M5 are thus at a index prole using the same set of CPT data is included as the
depth of 1.4, 6.7, 8.2, and 11.0 m, respectively. third column in Fig. 9. The prole shows the vertical variation
Figure 8 further shows the evolution of the identied soil strata of the fuzzy soil type indices of the highly probable clayey soil (HPC),
by the Bayesian approaches as the model class varies from M1 to the highly probable mixed soil (HPM), and the highly probable
M6 (i.e., the number of layers increases from 1 to 6). Model class M1 sandy soil (HPS). However, the probabilistic fuzzy subset approach
only has one layer, and therefore it does not have any internal can not directly stratify the soil prole, and sound engineering
boundary. The number of internal boundaries increases from 1 for judgment has to be exercised to identify the underground soil
M2 to 5 for M6. All model classes M2 to M6 share a common internal stratication. The fourth column in Fig. 9 includes an example of
boundary at a depth of about 6.7 m. In addition, model classes M3 a reasonable soil classication and stratication that can be in-
to M6 share one more internal boundary at a depth of about 1.4 m, ferred from the fuzzy soil type index prole in the third column
model classes M4 to M6 share one more internal boundary at a with such sound engineering judgment. The soil prole is divided
depth of about 11.0 m, and model classes M5 to M6 share one more into ve layers, and the boundaries are located at a depth of about
common internal boundary at a depth of about 8.2 m. 1.0, 7.0, 8.0, and 11.0 m. The number of soil layers and their bound-
The rst layer in M2 is further divided into two new layers in aries obtained from the probabilistic fuzzy subset approach and
M3; the third layer in M3 is further divided into two new layers in further interpreted with engineering judgment are in good agree-
M4; the third layer in M4 is further divided into two new layers ment with those from the Bayesian approaches developed in the
in M5; and the second layer in M5 is further divided into two new paper. In addition, the soil types in these ve soil layers are clas-
layers in M6. Apart from those identied in the previous model sied by the probabilistic fuzzy subset approach and listed in
class Mk 1, the most probable boundaries in the model class Mk order from the ground surface as HPM, HPC, HPM with HPS inclu-
include one additional internal boundary that divides one layer in
sions, HPM, and HPS with HPM inclusions. Such classications are
Mk 1 into two new layers in Mk. As the value of k increases, the
also consistent with those obtained from the Bayesian approaches,
Bayesian system identication approach applied in this study
as shown under the column of M5 in Fig. 9. When compared with
identies the soil layers progressively, starting from the statisti-
other probabilistic approaches, the Bayesian approaches developed
cally most signicant boundary and gradually zooming into less
in the paper directly stratify the soil prole. This bypasses the
signicant ones with improved resolution. It is worthwhile to
vague and unquantiable exercise of the engineering judgment
point out that, although the most probable model class (the most
probable number of soil layers) and its corresponding most prob- and allows transparent and quantitative consideration of the en-
able layer thicknesses are of the primary interest in the identi- gineering judgment as prior knowledge in the Bayesian frame-
cation of the underground soil stratication, the evolution of the work in the CPT-based soil classication and identication of the
identied soil strata as the model class increases (see Fig. 8) also underground soil stratication.
provides valuable information for assisting in the interpretation Figure 9 also includes, in the second column, the soil classica-
of CPT data in a rational and transparent manner. tion and stratication results based on the boring log obtained at
the same site. The soil prole is divided into four layers, and the
Result comparisons soil types in these four soil layers are classied in according to the
Figure 9 compares the results from the proposed Bayesian ap- Unied Soil Classication System (USCS (ASTM 2006)) and listed in
proaches with those from other approaches, such as the probabi- order from the ground surface as: sandy clay, clay, silty clay, and
listic fuzzy subset approach developed by Zhang and Tumay clay with silt seams. There are obviously some inconsistencies
(1999). As reported by Zhang and Tumay (1999), the fuzzy soil type between the results from the boring log with USCS and those from
Fig. 9. Comparison among the results of the boundaries of soil layers determined by different approaches in the illustrative example.
The Bayesian approach
of this study
Soil type inferred
Cone Tip Resistance (MPa) Fuzzy Soil Type Index Model Class Soil Type
from Zhang and
0 5 10 15 Boring log 0 0.2 0.4 0.6 0.8 1 Tumay 1999 M1 M2 M3 M4 M 5* M6 3 4 5 6 7 8 9
0
Sandy Clay HPM Soil type
8
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by UNIV LIBRARY OF HONG KONG UNIV OF on 04/03/16
Tip HPS
Resistance
2 HPM
Friction
Ratio HPC
Clay HPC
4 Soil type
3
Silty Clay
6
GWT
Depth (m)
12
HPS with HPM
inclusions Soil type
5
Zhang and Tumay
14 1999
For personal use only.
0 2 4 6 Note: 3: Clays (clay to silty clay), 4: Silt mixtures (clayey silt to silty clay)
5: Sand mixtures (silty sand to sandy silt), 8: Very stiff sand to clayey sand
Sleeve Friction Ratio (%)
the CPT-based approaches, either deterministic (i.e., the Robertson Summary and concluding remarks
chart) or probabilistic (e.g., the probabilistic fuzzy subset ap- This paper integrated the Robertson chart with Bayesian ap-
proach or the Bayesian approaches in this paper) ones. Such in- proaches to explicitly and properly consider the uncertainty in
consistencies have been reported in the literature (e.g., Zhang and CPT-based soil classication and spatial distribution of the CPT
Tumay 1999; Robertson 2009), and they can be attributed to the data. A Bayesian framework for soil stratum identication and
different soil classication systems used. Note that the USCS clas- soil classication based on CPT data has been developed. The
sies the soil based on the grain size distribution and the Atter- probability that the soil belongs to one of the nine soil types in the
berg limits test results, and it mainly reects the compositional Robertson chart for a given set of CPT data has been formulated
types of soil. In contrast, the CPT-based system classies the soil using the maximum entropy principle. The Bayesian framework
based on the mechanical responses of soil during CPT tests, and it contains two major components: a Bayesian model class selection
mainly reects the soil behavior types. Robertson (2009) has approach to identify the most probable number of underground
pointed out that the classication results from the USCS and CPT- soil layers and a Bayesian system identication approach to simul-
based system may not agree with each other, particularly for the taneously estimate the most probable layer thicknesses and clas-
mixed soils region (e.g., the soil type 5 sand mixtures and soil type sify the soil types.
4 silt mixtures) in the Robertson chart. Because a large portion of Equations were derived for the Bayesian approaches, and the
the CPT data points in this illustrative example are located within proposed approaches were illustrated using a real-life CPT test
the areas of soil types 4 and 5 (see the last column in Fig. 9), it is not example performed at the NGES site of Texas A&M University,
surprising to observe such inconsistencies. USA. It has been shown that the proposed approaches properly
identify the underground soil stratication and classify the soil
A closed examination on the evolution of the identied soil
type of each layer. In addition, as the number of model classes
strata as the model class increases in the proposed Bayesian ap-
increases, the Bayesian model class selection approach identies
proaches (see the fth column in Fig. 9), however, shows that all
the soil layers progressively, starting from the statistically most
layer boundaries identied by the USCS with a boring log are all
signicant boundary and gradually zooming into less signicant
included in the results obtained from the proposed Bayesian ap- ones with improved resolution. It is also found that, although the
proaches. There are three internal boundaries for the four layers most probable model class (the most probable number of soil
identied by the USCS with a boring log. The third and rst layers) and its corresponding most probable layer thicknesses are
boundaries are consistent with the internal boundaries identied of primary interest in the identication of the underground soil
in model classes M2 and M3, and the second boundary is consistent stratication, the evolution of the identied soil strata as the
with the new internal boundary identied in model class M6 model class increases also provides valuable information for as-
(see Fig. 9). This further illustrates the value of the evolution of the sisting in the interpretation of CPT data in a rational and trans-
identied soil strata as the model class increases for assisting in parent manner. The proposed Bayesian approach has been shown
the interpretation of CPT data. to perform well at the NGES site of Texas A&M University, and
further development and testing at other sites and case studies Jaynes, E.T. 2003. Probability theory: the logic of science. Cambridge University
will be carried out in future. Press, New York.
Jefferies, M.G., and Davies, M.P. 1993. Use of CPTU to estimate equivalent SPT
N60. Geotechnical Testing Journal, 16(4): 458468. doi:10.1520/GTJ10286J.
Acknowledgements Jung, B.-C., Gardoni, P., and Biscontin, G. 2008. Probabilistic soil identication
The work described in this paper was supported by a Strategic based on cone penetration tests. Gotechnique, 58(7): 591603. doi:10.1680/
Research Grant from City University of Hong Kong (Project Num- geot.2008.58.7.591.
ber 7002838) and a grant from the National Natural Science Foun- Kurup, P.U., and Grifn, E.P. 2006. Prediction of soil composition from CPT data
dation of China (Project Number 51208446). This nancial support using general regression neural network. Journal of Computing in Civil En-
Can. Geotech. J. Downloaded from www.nrcresearchpress.com by UNIV LIBRARY OF HONG KONG UNIV OF on 04/03/16