You are on page 1of 12

Computers and Geotechnics 122 (2020) 103537

Contents lists available at ScienceDirect

Computers and Geotechnics


journal homepage: www.elsevier.com/locate/compgeo

Research Paper

Differentiation of discrete data with unequal measurement intervals and T


quantification of uncertainty in differentiation using Bayesian compressive
sampling
Tengyuan Zhaoa, Yu Wangb,

a
School of Human Settlements and Civil Engineering, Xi’an Jiaotong University, Shaanxi Province, China
b
Department of Architecture and Civil Engineering, City University of Hong Kong, Tat Chee Avenue, Kowloon, Hong Kong

ARTICLE INFO ABSTRACT

Keywords: Calculation of derivatives on discrete measurement data with unequal intervals is often required in geotechnical
Compressive sensing engineering, such as interpretation of stiffness reduction curve of soil from pressuremeter test data, pile lateral
Sparse measurements responses from inclinometer data. Such a task is however tricky and challenging, because a small error or noise
Data interpretation in the measurements may amplify and lead to huge fluctuations in the derivatives obtained. The amplification
Bayesian methods
becomes increasingly significant as the order of derivative increases. It is therefore of great importance to
evaluate reliability of the derivatives obtained and quantify the uncertainty associated with the derivative
calculation. A Bayesian compressive sampling-based method is proposed in this paper to address this problem. It
not only provides high-order derivatives on discrete measurement data, even at un-sampled locations, but also
quantifies the uncertainty associated with the derivatives obtained and offers an index to evaluate reliability of
the derivatives obtained. The proposed approach is illustrated using both real-life pressuremeter data and nu-
merical example of pile lateral responses. A comparison is also made between the proposed method and several
existing methods in geotechnical literature. It shows that the proposed method performs better than existing
methods and it is applicable to problems with both elastic and plastic soil responses.

1. Introduction polynomial function, a spline function), followed by differentiation of


the fitted function (e.g., [21,5,29,9,19]). Although the approach is
In geotechnical engineering practice, measurements of quantities of simple for implementation, the calculated derivatives might be sig-
interest are often measured in a discrete manner with unequal intervals, nificantly affected by the predetermined function form (e.g.,
such as load-displacement data, deflection data along piles or drilled [31,47,16]). Li et al. [31] developed a digital filter approach to dif-
shaft, among others (e.g., [32,48]). Interpretation of the measurements ferentiate discrete and noisy measurements, which has been used to
obtained, however, may involve differentiation of the discrete mea- interpret secant shear modulus degradation curves from pressuremeter
surements (e.g., [36,10,31]). For example, in a pressuremeter test, it is tests (e.g., [41]). This method does not predetermine the function form
required to differentiate pressure data with respect to cavity strain to when modelling the discrete measurements, but it is only applicable to
derive stiffness reduction curve of soil (e.g., [10,31,41]). In a pile lat- measurements with equal intervals, which are often not available in
eral load test, calculation of derivatives for inclinometer deflection data geotechnical engineering practice. A weighted residual method (e.g.,
may be needed to derive pile lateral response, such as the distribution [46,3]) is also used to calculate derivatives of discrete measurements,
of bending moment, shear force and soil resistance along piles (e.g., but it cannot provide the derivatives at the locations where no mea-
[38,48]). surement data are available. In addition, compressive sampling/sen-
Calculation of derivatives for discrete data with unequal measure- sing, a new sampling strategy in signal processing and image analysis, is
ment intervals, however, is tricky and challenging. A number of recently adopted to interpret pile lateral response from inclinometer
methods have been proposed for calculating derivatives of discrete data, which involves calculation of derivatives of discrete measure-
measurements in literature. A commonly used method is to fit the ments (e.g., [48]).
discrete measurements with a predetermined function form (e.g., a Note that all the methods mentioned above can only be used to


Corresponding author.
E-mail addresses: tyzhao@xjtu.edu.cn (T. Zhao), yuwang@cityu.edu.hk (Y. Wang).

https://doi.org/10.1016/j.compgeo.2020.103537
Received 6 August 2019; Received in revised form 12 February 2020; Accepted 8 March 2020
Available online 20 March 2020
0266-352X/ © 2020 Elsevier Ltd. All rights reserved.
T. Zhao and Y. Wang Computers and Geotechnics 122 (2020) 103537

calculate derivatives of discrete measurements in a deterministic a bold italic symbol denotes a vector from here after.
manner, and the uncertainty associated with the derivatives obtained To obtain the non-trivial elements of ω from y, a relation between
cannot be evaluated or quantified. It is well recognized in literature that them may be established and expressed as
a small error in the discrete measurements may lead to huge errors in y= f = B = A = i = 1 Ai i . A = ΨB is a matrix with a dimen-
N

the calculated derivatives (e.g., [34,45,14]). As a result, the differ- sion of K × N; Ψ is a matrix with a dimension of K × N, reflecting the
entiation of discrete measurements might be inaccurate and contain locations of y in f. Using y= A , an approximation of ω, denoted as s ,
significant uncertainty. The uncertainty associated with calculated de- can be obtained deterministically (e.g., [13,11,42,12,48] or probabil-
rivatives may further propagate to and affect design or analysis when istically (e.g., [24,25;1,20,43,49]. s has the same length with ω, and
the calculated derivatives are used as input for the design or analysis its elements are all zero except the S non-trivial ones. Then, f can be
desired. This underscores the significance in proper differentiation of ^
approximated as f^ , which is expressed as f = B s .
discrete data with unequal measurement intervals and quantification of In this study, s is estimated under a Bayesian framework (e.g.,
the associated uncertainty. [30,44,40,2,28,8]), which is capable of quantifying uncertainty and
This paper aims to address this problem using Bayesian compressive dealing with the measurement error in the discrete measurements y. In
sampling (BCS) method (e.g., [24,43]), which is a novel method for the context of BCS, measurement error is modelled as a Gaussian
image analysis that is capable of not only recovering a complete image random variable ε with zero mean and unknown variance. Specifically,
from a remarkably small number of discrete measurements on the the discrete measurements y is expressed as y= A s + IK , where IK is a
image, but also quantifying the associated uncertainty. Using BCS, a K-length column vector with all elements being one. It has been shown
novel method is proposed in this paper, which not only provides high- that s follows a multivariate Student’s t distribution, which has the
order derivatives on discrete measurement data, even at un-sampled mean µ s and covariance matrix COV s (e.g., [43,50]):
locations, but also quantifies the uncertainty associated with the deri-
vatives in an objective manner. The uncertainty associated with the µ s
= HAT y= (ATA + D) - 1ATy
derivatives can be used as an index to evaluate reliability of the high- dn H dn (ATA + D) - 1
COV s = = (1)
order derivatives obtained from discrete measurement data, which can cn 1 cn 1

never be achieved by existing methods (e.g., [48]). In this paper, the


H = (ATA + D)−1; cn = K/2 + c; dn = d+( yT y - µ s T H - 1µ s )/2; c and
BCS is firstly reviewed and used to reconstruct data of interest with a
d are small non-negative constants (e.g., c = d = 10−4). D is a diagonal
high resolution from discrete measurements. Then, the derivative(s) of
matrix with Di,i = αi (i = 1, 2, …, N), which can be determined by
interest is (or are) calculated and the uncertainty associated with dif-
maximizing the likelihood of y, as detailed in literature (e.g.,
ferentiation is quantified properly. A step by step implementation
procedure is also provided. Both real-life and numerical examples are [25,20,43]). Then, the mean and covariance matrix of f^ can be ob-
used to illustrate and validate the proposed approach. tained using Eq. (1), f^ = B s, and expressed as (e.g., [26]:

2. Data Reconstruction, Differentiation and Its Uncertainty µ f^ = E (f^ ) = E (B s) = Bµ s


Quantification
COV f^ = E [(f^ µ f^ ) (f^ µ f^ )T] = BCOV s BT (2)
In this section, high-resolution data of a quantity of interest is first
where “E(∙)” means expectation operation; µ f^ represents the mean of
reconstructed from discrete measurements with unequal intervals using
BCS (e.g., [20,43]). The reconstructed data are represented analytically reconstructed f^ , which is denoted as the “best estimate” in this study;
by a weighted summation of a series of basis functions (e.g., cosine whereas diagonal elements of COV ^ represent the variances of f^ ,
f
functions). Therefore, derivatives of the reconstructed data, together which quantify the uncertainty associated with the reconstruction. It is
with its uncertainty, can be calculated properly, as discussed in the worthwhile to point out that Eq. (2) not only provides the best estimate
following two subsections. and quantifies the uncertainty associated with f^ , but also paves the way
to quantify the uncertainty associated with the derivatives of f^ , as
2.1. Data reconstruction using Bayesian compressive sampling
discussed in the next subsection.
Bayesian compressive sampling or sensing (BCS) is a probabilistic
extension of compressive sampling or sensing (CS) that is originated 2.2. Differentiation of discrete measurements and quantification of
from the field of signal processing and image analysis after the seminal uncertainty in differentiation
work by Candès et al. [7] more than ten years ago. It is capable of
reconstructing a signal, i.e., variations of a quantity with time, space, or In this study, discrete cosine transform is used to construct the
another quantity, from sparse measurements on that signal (e.g., matrix B. Columns of B, i.e., Bj (j = 1, 2, …, N), are a series of cosine
[6,24,25]). During the reconstruction, the BCS exploits the “compres- functions with different frequencies (e.g., [37]). The larger the j value,
sibility” nature of signals. “Compressibility” means that a signal, de- the higher frequency of the cosine function. This property can be uti-
noted as an N-length column vector f, can be represented concisely by a lized to improve the performance of BCS (See [42]). Note that the i-th
weighted summation of a limited number of basis functions (e.g., or- element of Bj, i.e., Bi,j (i = 1, 2, …, N) in terms of a coordinate z (e.g.,
thonormal cosine functions). Mathematically, f is expressed as depth), is expressed as (e.g., [48]):
f = i = 1 Bi i = B . B is an N × N orthonormal matrix, whose columns
N

Bi (i = 1, 2, …, N) represent a series of basis functions. ωi is the i-th (j 1)(z i z1 ) j 1


element of ω and represents the weight of Bi contributing to f. In the Bi, j = cj cos 2 +
2N 4N (3)
context of CS or BCS, most elements of ω are negligible except several
(e.g., S≪N) non-trivial ones that have significantly large magnitude. where cj = 1/ N for j = 1 and cj =2/ N for j ≠ 1. z1 and zi are z
Therefore, f can be reconstructed if the non-trivial elements of ω are
coordinates, corresponding to the first and the i-th element of f^ , re-
identified and estimated using the sparse and discrete measurements on
spectively. η represents a resolution that an engineer would like to have
f. The discrete measurements are denoted as a K-length column vector
in the reconstructed data f^ . Using f^ = B = and Eq. (3),
N
B
y= [y (z1), y (z2), , y (zK )]T , where z represents the spatial coordinate s j=1 j sj

(e.g., depth) and the superscript “T” represents transpose operation in the i-th element of f^ , i.e., fi (i = 1, 2, …, N), can be expressed ana-
linear algebra. Note that a bold upright symbol denotes a matrix, while lytically as:

2
T. Zhao and Y. Wang Computers and Geotechnics 122 (2020) 103537

N N
(j 1)(z i z1 ) j 1 COV s , which remains unchanged with the increase of n. Bi(,nl ) and Bi(,nm)
fi = Bi, j = sj cj cos 2 +
sj
2N 4N represent elements of B(n) . Therefore, when the magnitude of elements
j=1 j=1 (4) (n )
of B(n) increases with n, the variance of f^ , i.e., Eq. (10), also in-
sj in Eq. (4) is the j-th element of s. Because Eq. (4) is an analytical creases. This indicates that, as n increases, the uncertainty in the deri-
and explicit function of zi, differentiation of fi with respect to zi can be (n ) (n )
vatives f^ increases, and f^ becomes less reliable. This observation is
performed analytically. Consider, for example, the first order derivative consistent with common sense and results reported in literature (e.g.,
of fi with respect to zi. It is expressed as: [39]). The derivations in this subsection not only provide mathematical
dfi
N
dBi, j N
(j 1) (j 1)(z i z1 ) j 1 proof to support the observation, but also offer a rational way to
= = s j cj sin 2 +
quantify the increase of uncertainty as the order of derivative increases.
sj
dz i
j =1
dz i
j=1
N 2N 4N
(5)
Eq. (5) suggests that the first order derivative of fi with respect to zi is
3. Implementation procedure
equal to the weighted summation of the first order derivative of all
elements of the i-th row of B (denoted as Bir) with respect to zi, which is
In this section, the proposed method is summarized as a six-step
denoted as an N-length row vector Bir(1) (i = 1, 2, …, N). The j-th ele-
dBi, j implementation procedure. Each step and the key equations involved
ment of Bir(1) is expressed as Bi(1)
, j = dz i . The superscript “(1)” represents are detailed as below:
the first order derivative. Note that Eq. (5) can be rewritten in a more
concise form as: Step 1: Obtain the discrete measurements as
y= [y (z1), y (z2), , y (zK )]T and the range that z covers, e.g., h = zK-
df^i
= Bir(1) s z1. Determine a resolution η that an engineer would like to have in
dz i (6) the reconstructed data. Calculate N as N = h/η + 1;
The first order derivative of all elements of f^ with respect to z, i.e., Step 2: Decompose the discrete measurements y into two compo-
(1) nents: a linear component yL and a nonlinear component yNL, i.e.,
f^ , can be obtained using Eq. (6), and they are expressed in a matrix
y = yL + yNL;
form as:
Step 3: Calculate the function for the linear component using the
^
B1(1)
r
first and last data points of y as fL = a z+ b . The coefficients “a”
(1) B2(1) and “b” are calculated as a = [y (zK ) y (z1)]/(z k z1) and
f^ = r
s = B(1) s b = y (z1) az k , respectively. z is a vector representing z1, z2, …,
(1) with a resolution of η;
BNr (7) Step 4: Convert the nonlinear component yNL into a periodic data
where is the i-th (i = 1, 2, …, N) row of matrix B , which is the
Bir(1) (1) set by alternatively cloning itself and its folded and inverted copies
first order derivative of matrix B with respect to z. Following a similar for multiple (e.g., NT = 20) times. This technique has been used in
procedure, the n-th order derivatives of f^ with respect to z, denoted as literature to improve the performance of the proposed method when
(n ) calculating the derivatives at the boundaries (e.g., [31,41,48]);
f^ , is derived as: Step 5: Take the periodic data set as input for the BCS approach.
(n ) Both the best estimate and quantified uncertainty of the nonlinear
f^ = B(n) s (8)
component with a high resolution are obtained using Eqs. (2). Note
where B(n) represents the n-th order derivative of B with respect to z. that the reconstructed f^NL with a high resolution is represented
The element at the i-th row and the j-th column of B(n) is expressed as analytically by a weighted summation of cosine functions;
dnBi, j
Bi(,nj ) = dzin
. Step 6: Obtain the reconstructed data with a high resolution as a
(n ) summation of two components as f^ = f^L + f^NL . Calculate the n-th
Once Eq. (8) is obtained, the mean and covariance matrix of f^ can
dnf^ dnf^L dnf^NL
be derived in a manner similar to Eq. (2): order derivative with respect to z using dz n
= dz n
+ dz n
. The mean
dnf^
^ (n) and quantified uncertainty associated with can be obtained
µ ^(n) = E (f ) = E (B(n) s) = B (n ) µ s
dz n
f
dnf^L
(n ) (n ) using Eq. (9) when n > 1. Note that the first term in the right-
COV ^ (n) = E [(f^ µ ^(n) ) (f^ ) ]=
µ ^(n) T B(n) COV [B(n) ]T dz n
f f f
s
(9) hand side of
dnf^ dnf^L dnf^NL
= dz n + dz n disappears when n > 1 because
dz n
where µ ^ (n)
(n )
represents the mean of f^ , i.e., the best estimate of the n- high-order derivatives of a linear term are zero. When n = 1,
f df^L
= aIN is a constant, where IN is an N-length column vector with
th order derivative of f^ with respect to z; whereas diagonal elements of dz
all elements being one.
(n )
COV (n) represent the variance associated with f^ , quantifying the
f^
(n ) 4. Illustrative Example I: Pressuremeter data involving first-order
uncertainty associated with the n-th order derivative f^ . It is worth-
derivative
while to point out that as n increases, the magnitude of elements of B(n) ,
dnBi, j
i.e., Bi(,nj ) = dzin
may increase too because it contains a factor of In this section, the proposed method and the implementation pro-
( (j
N
1) n
) due to repeated differentiation of fi . Note that each element cedure are illustrated using a set of real pressuremeter test data, i.e.,
cavity strain (εc) vs pressure (P) data pairs, from Prust et al. [35], as
(n )
of B(n) is explicitly involved in the estimated variance of f^ , i.e., the shown in Fig. 1 by open circles. The pressuremeter test was carried out
diagonal elements of COV ^(n) , which is derived as Eq. (10) from Eq. (9): in a site in Thailand. The site is comprised of a top layer of a made
f
N N ground extending from the ground surface to a depth of around
2
(n) = Bi(,nl ) Bi(,nm) COV sl,m 0.1–4 m, a Bangkok soft clay layer which extends to a depth of around
f^i l=1 m=1 (10) 12.9–17.9 m, a first stiff clay layer which further extends to a depth of
(n ) around 20.4–29.3 m, underlain by a Bangkok aquifer layer which
where 2 is the estimated variance for the i-th element of f^ ;
fi
( n)
mainly contains sand. In this study, the data in Fig. 1 are from a
COV sl, m is the element at the l-th row and the m-th column of the matrix pressuremeter test in sands within the Bangkok aquifer layer.

3
T. Zhao and Y. Wang Computers and Geotechnics 122 (2020) 103537

1200 that, convert the nonlinear component into a periodic signal by alter-
natively rotating itself and its folded and inverted copies along log10(εr)
axis for multiple (e.g., NT = 20) times (Step 4). The periodic data are
1000
then taken as input to BCS. Subsequently, the periodic data of the
nonlinear component f^NL with a high resolution are obtained and the
800 uncertainty in terms of two times standard deviation (SD) is calculated
Pressure, P (kPa)

(Step 5). Using the best estimate of f^NL and 2 SD associated with f^NL ,
600 the upper and lower bounds of f^ can be calculated as µ ± 2
NL fi fi
(i = 1, 2, …, N), where µ fi and fi are the i-th element of µ f^ and the
square root of the i-th diagonal element of COV f^ . Then, adding the
400
linear function with f^ = 3.7841log ( ) + 19.0149 back to the non-
L 10 r
linear component reconstructed leads to the best estimate (P, εr) data
200 curve with a high resolution and the upper and lower bounds of the (P,
εr) curve, as shown in Fig. 3a by a dashed and a pair of two dotted lines.
0 For comparison, Fig. 3a also includes the original (P, εr) data pairs by
0 2 4 6 8 10 open circles. It is evident in Fig. 3a that the dashed line is very con-
Cavity strain, c
(%) sistent with the open circles, indicating the (P, εr) data curve re-
constructed by BCS is reasonable. In addition, the lower and upper
Fig. 1. Pressuremeter test result from Bangkok, Thailand (after [35]). bounds in Fig. 3a are almost overlapped with the dashed line, in-
dicating that the reconstructed (P, εr) curve is reliable and confident
As the final unloading portion of a pressuremeter test is generally (e.g., [50]). Following Step 6 in Section 3, dP/log10(εr) is calculated as:
considered as relatively reliable (e.g., [36,41]), it is used in this study to
derive the secant shear modulus Gsec (e.g., [10]): dP df^L df^NL df^NL
= + = 3.7841 +
d (log10 r ) d (log10 r ) d (log10 r ) d (log10 r ) (13)
1 dP
Gsec = (1 + curr )
2 d curr (11) Then, the best estimate and SD of dP / d (log10 r ) are expressed as two
column vectors, i.e., µdP / d (log10 r ) = 3.7841 + B(1) µ s and
where curr = 1c + max is the current cavity strain and εmax is the max-
dP / d (log10 r ) = diag ( B COV s [B ] ) , respectively, using Eqs. (9) and
max (1) (1) T
imum cavity strain for the unloading portion of a pressuremeter test. As
(13). “diag(∙)” represents the diagonal elements of a matrix. Although
shown in Eq. (11), derivation of Gsec requires the differentiation of
both the mean and SD of dP / d (log10 r ) are obtained, dP/dεr is unknown
pressure P with respect to the current cavity strain εcurr. The P vs εcurr
but required in Eq. (12). Therefore, it is necessary to relate
data pairs of the unloading portion in Fig. 1 can be calculated readily
dP / d (log10 r ) to dP/dεr (e.g., [31]):
using curr = 1c + max . Note that P vs εcurr data pairs are preferably pre-
max
sented in a logarithm scale (e.g., [10,31]). Because the εcurr data are dP
=
1 dP
=
1 dP
negative during unloading, εcurr data are transformed into positive ra- dr r ln(10) d (log10 r ) 2.3026 r d (log10 r ) (14)
dial strain εr by setting εr = −εcurr. Fig. 2a plots the P vs εr data pairs by
Substituting the mean and SD of dP / d (log10 r ) obtained into Eq.
open circles in a logarithm scale. Eq. (11) is then rewritten as:
(14) leads to the mean and SD of dP/dεr as:
1 dP
Gsec = ( r 1) µdP / d r =
1
( 3.7841 + B(1) µ s )
2 dr (12) 2.3026 r
1
=
In such a case, the discrete measurements with unequal measure- dP / d r 2.3026 r dP / d (log10 r ) (15)
ment intervals are obtained (see open circles in Fig. 2a). Then, de-
The best estimate of dP/dεr (e.g., µdP / d r ) is shown in Fig. 3b by a
termine a resolution η (in a logarithm scale) in the reconstructed data
dashed line. Similarly, and µdP / d r ± 2 dP / d r are shown in Fig. 3b by a
such as η = 0.01 and calculate the range h that log10(εr) spans as pair of dotted lines. Then, using Eqs. (15) and (12), the best estimate
h = [log10(εr)]max-[log10(εr)]min = 2.06. Compute the length N as and SD of secant shear modulus Gsec are derived as:
N = h/ η + 1 = 207 (Step 1). In this example, z represents log10(εr).
1
Then, decompose the discrete measurements as a combination of linear µ Gsec = 2 ( r 1) µdP / d r
component yL and a nonlinear component yNL as shown in Fig. 2b and 1
Gsec = (
2 r
1) dP / d r (16)
c, respectively (Step 2). The function of the linear component is cal-
culated as f^L = alog10 ( r ) + b , where “a = -3.7841” and “b = 19.0149” Using Eq. (16), µ Gsec and µ Gsec ± 2 Gsec are computed and shown in
are obtained using the equations in Step 2 in Section 3 (Step 3). After Fig. 3c using the same legends as those in Fig. 3a and b. Note that the

1400 1400 400


(a) Original measurement (b) Linear component of y, (c) Nonlinear component
1200 data y 1200 i.e., yL of y, i.e., yNL
300
Pressure, P (kPa)

1000 1000
800 800
600 =600 + 200
400 400 100
200 200
0 0 0
0.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10
Radial strain, r (%) Radial strain, r (%) Radial strain, r (%)

Fig. 2. Decomposition of discrete pressuremeter data y with unequal measurement intervals.

4
T. Zhao and Y. Wang Computers and Geotechnics 122 (2020) 103537

1200 0
Transformed-strain

Secant shear modulus Gs (MPa)


(a) (b) (c)
150 approach (Jardine [36])
1000
The proposed method

Pressure, P (kPa)
800 -60 Best estimate
100 Best estimate 2SD
600

r
dP/d
400 -120 50
Measurements y Best estimate
200 Best estimate Best estimate
Best estimate 2SD 2SD
0 -180 0
0.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10
Radial strain, r (%) Radial strain, r (%) Shear strain, (%)

Fig. 3. (a) Reconstructed pressuremeter data; (b) derivatives and (c) secant shear modulus.

shear strain γ in Fig. 3c is approximately two times of the radial strain, is elastoplastic, which is reflected by the p-y curve shown in Fig. 4c.
i.e., γ ≈ 2εr (e.g., [41]). Fig. 3c shows that as the shear strain γ in- Fig. 4c plots the p-y curves at depths of 0.5 m, 1.0 m and 3.0 m, re-
creases, the secant shear modulus Gsec gradually decreases. Further- spectively, by solid, dashed and dotted lines. Fig. 4c shows that the soil-
more, when γ is relatively small, the uncertainty in the Gsec interpreted pile system tends to be elastic when deflections are less than 1 mm. In
by the proposed method is relatively large, and vice versa. This may be this example, the deflections at the depth of 0.5 m, 1.0 m and 3.0 m are
attributed to the fact that, when γ is small, the pressuremeter data are around 19.0 mm, 13.8 mm and 5.2 mm, respectively, which are much
affected more significantly by measurement errors, and hence, less re- larger than the elastic threshold of 1 mm, as shown in Fig. 5a. For the
liable. These observations are consistent with what have been reported pipe pile at the depths less than z = 0.5 m, the soil-pile system exists
in literature (e.g., [17,23,36]). In addition, Fig. 3c also shows the Gsec high levels of plasticity.
obtained from the transformed-strain approach [23] by a solid line with Generally speaking, once the underlying deflection measurements f
open triangles. The dashed line from the proposed method is very (z) and the flexural rigidity EI of the pile are available, the pile lateral
consistent with the solid line with open triangles, indicating that the responses, i.e., distribution of M, V and p along piles can be mathe-
best estimated Gsec curve from the proposed method is reasonable. matically expressed as below in accordance with beam theory (e.g.,
[18,33]):

5. Illustrative Example II: Pile lateral response data involving high d 2f (z )


M = EI
order derivatives dz 2 (17a)

Note that Section 4 only explores the first order derivative of dis- d3f (z ) dM
V = EI =
crete measurements, while the proposed method is applicable to high dz 3 dz (17b)
order differentiation. In this section, this point is systematically in-
vestigated using a pile example from LPILE package [15]. The pile d 4 f (z ) dV
p = EI =
lateral responses, including distribution of bending moment (M), shear dz 4 dz (17c)
force (V), and soil resistance (p), are interpreted from deflection mea-
surements. This involves calculation of a second, third and fourth order z in this example is used to represent the depth coordinate along piles.
derivatives with respect to depth, respectively. In this example, the pile In this example, the complete f(z) along depth is unknown and needs to
is installed within a sand layer, which has a thickness of around be reconstructed from the K = 11 discrete measurements on f(z), i.e.,
10.67 m, as shown in Fig. 4a. The layer has a friction angle of the open circles in Fig. 5a. Specify a resolution η in the reconstructed f
ϕ′silt = 38°, and effective unit weight of γsilt = 17.28 kN/m3. The pile (z) data such as η = 0.09144 m and calculate the range h that an in-
used in this example is a pipe pile which has a length of 9.144 m, an clinometer records as h = zmax-zmin = 9.144 m. Compute the length N
outside diameter of 406.4 mm, a wall thickness of 12.7 mm (see as N = h/ η + 1 = 101 (Step 1). Decompose the discrete inclinometer
Fig. 4b), an elastic modulus of around E = 199.95 GPa, and an area data y (i.e., the open circles in Fig. 5a) into two components: a linear
moment of inertia of I = 3.05 × 10−4 m4. Note that although a con- component yL and a nonlinear component yNL (Step 2). Calculate the
stant EI = E × I = 60.92 N·m2 is used in this example, the method is function for the linear component using the first and last data points of
^
equally applicable to incorporate non-elastic modulus, which may be y as fL = a z+ b = -2.7650 z + 25.28 (Step 3). Convert the nonlinear
obtained through the methods given in literature (e.g., [51,52]). Given component yNL into a periodic dataset by alternatively cloning itself
the above conditions, some specified loads, such as a lateral shear load and its folded and inverted copies for multiple (e.g., NT = 20) times
of Q = 200.0 kN and an axial load of 133.45 kN at the pile head as (Step 4). Then, the periodic data are taken as input to the BCS method
shown in Fig. 4a, and built-in p-y criterions in LPILE package, dis- and the nonlinear component of the inclinometer data with a high re-
tribution of deflection, shear force and soil resistance along the pile solution, i.e., f^NL is obtained probabilistically, such as a best estimate
with an equal interval of 9.144 cm are obtained from the LPILE and ± 2SD (Step 5), as shown in Fig. 5a by a dashed line and two
package, as shown in Fig. 5a, 5b, 5c and 5d, respectively by a solid line. dotted lines, respectively. It is evident from Fig. 5a that both the dashed
In this example, K = 11 deflection data (i.e., the open circles in Fig. 5a) line and the dotted lines are almost overlapped with the solid line and
are taken as discrete data measured from inclinometer. They are used open circles, indicating that the deflection data reconstructed by BCS is
together with EI of the steel pipe pile as input to the proposed method reasonable, realistic and confident. Then, calculate the 2nd, 3rd and 4th
for interpretation of the pile lateral responses (i.e., distribution of M, V, order derivative of f^ = f^L + f^NL with respect to depth z using
and p along the pile). Note that the pile lateral responses from LIPILE dnf^ dnf^L dnf^NL dnf^NL dnf^
= + (n = 2, 3, and 4). Note that the term nL
=
are only used for comparison and validation of the results from the dz n dz n dz n dz n
dz
proposed method. In addition, the soil-pile interaction in this example disappears for the linear function f^L = a z+ b when n ≥ 2 (Step 6).

5
T. Zhao and Y. Wang Computers and Geotechnics 122 (2020) 103537

(a) loads and configuration of the pipe pile (b) section of the steel pipe pile
300
p-y curves at depths z
z = 0.5m
z = 1.0m
Soil resistance, p(kN/m)

z = 2.0m
200

100

0
0.000 0.005 0.010 0.015 0.020
Deflection y (m)
(c) Non-elastic p-y relationships at several depths
Fig. 4. A pipe pile example from LIPLE package (after [15]).

Deflection (mm) Bending moment (kN m) Shear force (kN) Soil resistance (kN/m)
-10 0 10 20 30 0 100 200 300 -200 0 200 -400 -200 0 200 400
0
(a) (b) (c) (d)
2

4
Depth, z(m)

Deflection data
10 from LPILE
Sparse deflection Bending moment Shear force from Soil resistance from
data y from LPILE LPILE LPILE
12 Best estimate Best estimate Best estimate shear Best estimate soil
deflection data bending moment force resistance
Best estimate SD Best estimate SD Best estimate SD Best estimate SD
14
Fig. 5. Interpretation of pile lateral response from discrete inclinometer data.

6
T. Zhao and Y. Wang Computers and Geotechnics 122 (2020) 103537

Since f^NL is obtained probabilistically, NL


dnf^
can also be evaluated sta- pile lateral responses (i.e., M, V and p). In this section, predetermined
dz n
tistically with mean and covariance matrix shown in Eq. (9). The mean function forms, such as polynomials, splines and cosine-exponential
and covariance of bending moment M, V and p are then obtained as combined model are used to fit deflection data and interpret pile lateral
µM = EIµ ^(2) and COVM = EI COV ^ (2) , µ V = EIµ ^ (3) and responses (i.e., M, V and p). The results from these methods are com-
f f f pared with those from the proposed method. Note that the same de-
COVV = EI COV ^(3) , and µp = EIµ ^ (4) and COVp = EI COV ^(4) , respec-
f f f flection data as those in Section 5 (i.e., the open circles in Fig. 5a) are
tively. μM, μV and μp represent the best estimate of pile responses M, V used for all methods to ensure a fair comparison.
and p; diagonal elements of COVM, COVV, and COVp represent the For the polynomial method, the measured deflection data are used
quantified variance for M, V and p, respectively, from which the SD is to fit the function f(z) = a0 + a1z + a2z2+…+ aqzq (q is a positive
obtained. Fig. 5b, 5c and 5d summarize the best estimate and un- integer). In this study, q is taken as q = 7 as it is often recommended in
certainty associated with M, V and p using the same legends as those in literature when using the polynomial for differentiation (e.g., [22], and
Fig. 5a. The dashed lines in all subplots of Fig. 5 are generally consistent the coefficients ai (i = 0, 1, 2, …, 7) are obtained through a least-square
with the corresponding solid lines, and most variations of the solid lines method. With the data shown in Fig. 5a (i.e., the open circles), it is
fall within the corresponding regions defined by best estimate ± 2SD, obtained that f(z) = 25.28–11.20z-1.58z2 + 1.69z3-0.37z4 + 0.035z5-
although there are slight differences between the dashed and solid lines 1.4 × 10−3z6 + 1.6 × 10−5z7. For the spline method, the measured
in Fig. 5c and 5d. Note that a third and fourth order (i.e., high-order) deflection data are used for fitting piecewise polynomials in a format f
derivatives of deflection data with respect to depth are calculated when (z) = a0 + a1(z-zi) + a2(z-zi)2 + a3(z-zi)3 + a4(z-zi)4, where
computing the distribution of shear force in Fig. 5c and soil resistance zi ≤ z ≤ zi+1. zi and zi+1 are the depths in an ascending order corre-
in Fig. 5d, respectively (see Eqs. (17b and c)). As discussed under the sponding to measured deflection data. Using the algorithm by Jonas
Section 2.2, uncertainty in the derivative results increases as order of [27], the coefficients for the piecewise polynomials are obtained, as
the derivative increases. This point is further supported by the SD as- summarized in Table 1. For the cosine-exponential model, the measured
sociated with the second, third and fourth order derivatives of re- deflection data are used for fitting the function
constructed data with respect to depth, as shown in Fig. 6 by a dotted, f (z ) = a exp( z / b)cos(z / b) (e.g., [4]). The parameters “a” and “b” are
solid, and dash-dotted lines, respectively. Fig. 6 also includes the SD obtained as a = 25.18 and b = 2.04 through a least square method.
associated with the reconstructed data by a dashed line, where no The three fitted functions f(z) for polynomial, spline and cosine-ex-
differentiation is involved. Fig. 6 shows that, as the order of derivative ponential methods are shown in Fig. 7a by a dotted line with diamonds,
(i.e., n) increases, the corresponding SD also increases. This implies that crosses, and open squares, respectively. Fig. 7a also includes the ori-
the differentiation obtained from the proposed method becomes less ginal deflection data from LPILE by a bold solid line. It shows that the
accurate and reliable as the order of derivative increases, as explained dotted lines with open diamonds, crosses and open squares are very
in the Section 2.2. The proposed method not only provides a rational consistent with the solid line. This suggests that all the above methods
and proper way for calculating high order derivatives of discrete data provide a reasonable and realistic fitting to the measured deflection
with unequal measurement intervals, but also provides a quantitative data. Then, bending moment M, shear force V and soil resistance p are
index on the reliability of the differentiation results obtained. This is calculated using Eq. (17) together with these three functions f(z), as
different from Zhao and Wang [48], where differentiation of discrete shown in Fig. 7b, 7c and 7d, respectively, by the same legends as those
data can only be obtained deterministically. in Fig. 7a. Fig. 7b–d also include M, V and p from LPILE by a bold solid
line. It is seen from Fig. 7b to 7d that there are large discrepancies
6. Comparison with existing methods between the dotted lines with open squares, crosses and open squares
and the bold solid lines. Moreover, Fig. 7 shows that as the order of
As mentioned in Section 1, a commonly used method in literature to differentiation increases, the discrepancies increase due to the ampli-
interpret the inclinometer data is to fit the discrete deflection mea- fication of error during differentiation. These results indicate that M, V
surements first with a predetermined function form, followed by dif- and p interpreted from the polynomial, spline and cosine-exponential
ferentiation of the fitted function with respect to depth to obtain the method may be unrealistic although all the three methods provide a
reasonable fitting to the measured deflection data.
Standard deviation (SD) In contrast, the results from the proposed method (i.e., the dashed
0.0 0.5 1.0 line in Fig. 7) are in good agreement with those from the LPILE (i.e., the
0
bold solid line), indicating that the proposed method not only provides
a reasonable fitting to the deflection data but also a realistic inter-
2 pretation of M, V and p. The proposed method therefore is more ac-
curate and robust than the existing methods. In addition, note that the
proposed method also quantifies the uncertainty associated with cal-
4
culation of the high-order derivatives (see Figs. 5 and 6) and offers an
index to evaluate reliability of the derivatives obtained. In contrast, the
Depth, z(m)

6 uncertainty cannot be quantified by either the polynomials, spline or


cosine-exponential methods.

8 7. Effect of number of measurements and measurement error

10 In this section, the effect of number of measurements and mea-


surement errors on the performance of the proposed method are ana-
SD of the n-th derivative
lyzed, which are discussed in detail in the following two subsections.
12 n=0 n=2
n=3 n=4 7.1. Effect of number of measurements
14
In this subsection, different number of measurements scenarios, e.g.,
Fig. 6. Evolution of the quantified uncertainty associated with the derivatives K = 20, 30, 40 together with the K = 11 scenario explored in Section 5
in pile lateral response interpretation. are utilized to explore the effect of number of measurements on the

7
T. Zhao and Y. Wang Computers and Geotechnics 122 (2020) 103537

Table 1
Summary of piecewise polynomials coefficients for a 4th order spline method.
Depth z (m) Coefficients for polynomials “a0 + a1(z − zi) + a2(z − zi)2 + a3(z − zi)3 + a4(z − zi)4″

a0 a1 a2 a3 a4

0.0000 ≤ z ≤ 0.9144 25.2800 −6.0961 −14.5730 12.6554 −3.4032


0.9144 ≤ z ≤ 1.8288 13.8000 −10.8544 2.9565 −1.0364 0.3889
1.8288 ≤ z ≤ 2.7432 6.4700 −7.2269 1.9780 0.2438 −0.1480
2.7432 ≤ z ≤ 3.5662 1.9300 −3.8057 1.9785 −0.2433 −0.0451
3.5662 ≤ z ≤ 4.4806 −0.1900 −0.7477 0.9707 −0.4247 0.0726
4.4806 ≤ z ≤ 5.4864 −0.3360 0.1843 0.1700 −0.1591 0.0370
5.4864 ≤ z ≤ 6.4008 −0.1210 0.2096 −0.0805 −0.0236 0.0160
6.4008 ≤ z ≤ 7.3152 0.0007 0.0411 −0.0548 0.0406 −0.0140
7.3152 ≤ z ≤ 8.3210 0.0137 0.0023 −0.0114 −0.0055 0.0050
8.3210 ≤ z ≤ 9.1440 0.0041 −0.0168 0.0025 0.0147 −0.0074

Deflection (mm) Bending moment (kN m) Shear force (kN) Soil resistance (kN/m)
-10 0 10 20 30-100 0 100 200 300 400 -200 0 200 -400 -200 0 200 400
0
(a ) (b) (c ) (d)

4
Depth, z(m)

10
Sparse deflection data y Original data (i.e., deflection, bending moment, shear force and soil resistance) from LPILE
Fitted deflection function and pile lateral responses interpreted from different methods
The proposed method Polynomical method with an order of seven
Spline method with an order of four Cosine-exponential method

Fig. 7. Comparison of pile lateral responses from the proposed method and that from polynomial, spline and cosine-exponential method.

Deflection (mm) Deflection (mm) Deflection (mm) Deflection (mm)


-10 0 10 20 30 -10 0 10 20 30 -10 0 10 20 30 -10 0 10 20 30
0
(a) K = 11 (b) K = 20 (c) K = 30 (d) K = 40

4
Depth, z(m)

10
Deflection data Deflection data Deflection data Deflection data
12 from LPILE from LPILE from LPILE from LPILE
Sparse deflection Sparse deflection Sparse deflection Sparse deflection
data y data y data y data y
14
Fig. 8. Different number of discrete deflection data points.

8
T. Zhao and Y. Wang Computers and Geotechnics 122 (2020) 103537

Deflection (mm) Bending moment (kN m) Shear force (kN) Soil resistance (kN/m)
-10 0 10 20 30 -100 0 100 200 300 400 -200-100 0 100 200 300 -400 -200 0 200 400
0
(a) (b) (c) (d)

4
Depth, z(m)

10 Deflection data Bending moment Shear force from Soil resistance from
from LPILE from LPILE LPILE LPILE
Best estimate under Best estimate under Best estimate under Best estimate under
12 different K scenarios
different K scenarios different K scenarios different K scenarios
K = 11 K = 20 K = 11 K = 20 K = 11 K = 20 K = 11 K = 20
K = 30 K = 40 K = 30 K = 40 K = 30 K = 40 K = 30 K = 40
14
Fig. 9. Best estimate of pile lateral responses derived by the proposed method under different number of measurements.

SD (mm) SD (kN m) SD (kN) SD (kN/m)


0.00 0.02 0.04 0 2 4 6 0 5 10 15 20 0 20 40 60 80
0
(a) (b) (c) (d)

4
Depth, z(m)

10
Quantified SD under Quantified SD under Quantified SD under Quantified SD under
different K scenarios different K scenarios different K scenarios different K scenarios
12
K = 11 K = 20 K = 11 K = 20 K = 11 K = 20 K = 11 K = 20
K = 30 K = 40 K = 30 K = 40 K = 30 K = 40 K = 30 K = 40
14
Fig. 10. Standard deviation of pile lateral responses quantified by the proposed method under different number of measurements: (a) for deflection data; (b) for
bending moment; (c) for shear force and (d) for soil resistance.

proposed method. The four measurements K scenarios, i.e., K = 11, 20, the results obtained become increasingly reliable and confident. This is
30, and 40, are shown as open circles in Fig. 8a to 8d, respectively. Each reasonable because more information is adopted in the reconstruction
subplot in Fig. 8 also includes the complete deflection profile from and differentiation, and the associated uncertainty is effectively re-
LPILE by a solid line. Then, following the procedure illustrated in the duced as K increases (e.g., [43,50]).
Section 5, the best estimate of reconstructed data, distribution of M, V
and p are obtained, as shown in Fig. 9a–d. In each subplot of Fig. 9, the
results corresponding to K = 11, 20, 30, and 40 scenarios are re- 7.2. Effect of measurement error
presented by a solid line with open circles, crosses, open diamonds and
open stars, respectively. Fig. 9 shows that all lines are generally con- In this subsection, the performance of the proposed method is ex-
sistent with the solid line in each subplot, suggesting that the best es- amined when the discrete measurements contain noise. Two types of
timate of reconstructed data, distribution of M, V and p are reasonable noises are added to the sparsely measured deflection data: (1) Gaussian
and realistic even when K is relatively small. In addition, the un- white noise; and (2) uniform white noise. In this example, a peak-to-
certainties (in terms of SD) of reconstructed data, distribution of M, V peak noise amplitude of 1% of the maximum deflection in Fig. 5a is
and p, are greatly influenced by the number of measurements K, as added. The deflection data with the Gaussian and uniform white noises
shown in Fig. 10a–d, respectively. The legends used in Fig. 10 are the are shown in Fig. 11a and 11b, respectively, where the open circles
same as those in Fig. 9. Fig. 10 shows that as K increases, the SD over represent the discrete deflection data with noises.
depth in each subplot becomes increasingly small, which implies that Then, using the discrete measurements with noises in Fig. 11a and
11b as input to the proposed method, the pile lateral responses are

9
T. Zhao and Y. Wang Computers and Geotechnics 122 (2020) 103537

Deflection (mm) Deflection (mm) in Fig. 13, which utilizes the discrete and noisy measurements in
-10 0 10 20 30 -10 0 10 20 30
0 0 Fig. 11b. All these agreements indicate that: (1) the proposed method is
(a) (b) robust and accurate in reconstruction and differentiation, and (2) the
pile lateral responses interpreted are reasonable.
2 2

8. Conclusion
4 4
This paper proposes a Bayesian compressive sampling (BCS) based
approach to probabilistically calculate high order derivatives on dis-
6 6 crete data with unequal measurement intervals. The proposed method
Depth, z(m)

Depth (m)
not only provides high-order derivatives with a high resolution, but also
quantifies uncertainty associated with the derivatives obtained. It
8 8 showed that the uncertainty of the differentiation increases as the order
of derivatives increases, implying decreasing reliability in high order
derivatives on discrete data. This is consistent with the observation that
10 10 a small error in the measurements may lead to huge fluctuations in high
order derivatives obtained. The quantified uncertainty may be used as
an indicator to judge the quality and reliability of the derivatives cal-
Deflection data from Deflection data from
12 12 culated from the discrete measurements.
LPILE with Gaussian noise LPILE with uniform noise
Sparse deflection data y Sparse deflection data y Both real-life pressuremeter data and numerical example of pile
lateral responses were adopted to illustrate and validate the proposed
14 14 method. The results showed that the proposed method is capable of
Fig. 11. Deflection measurement data with (a) Gaussian white noise and (b) providing an accurate reconstruction of the quantity of interest, as well
uniform white noise. as reasonable high order derivatives with a high resolution from sparse
measurements with unequal intervals, even when data contain mea-
surement error. In addition, a comparison was made between the pro-
obtained for each of the two cases. The results corresponding to the
posed method and several existing methods (e.g., polynomials, spline
noisy measurements in Fig. 11a and 11b are summarized in Figs. 12 and
and cosine-exponential method). The results showed that the proposed
13, respectively, using the same symbols as those in Fig. 5. It is evident
method is more robust and accurate than the existing methods and it is
in Figs. 12 and 13 that both the reconstructed deflection data and the
applicable to problems with both elastic and plastic soil responses.
pile lateral responses interpreted by the proposed method are in good
agreement with those from LIPILE. Consider, for example, Fig. 12,
which corresponds to the noisy measurements in Fig. 11a. Fig. 12a
CRediT authorship contribution statement
shows that the dashed and two dotted lines are very consistent with the
bold solid line. This indicates that the BCS approach is robust and
Tengyuan Zhao: Software, Validation, Formal analysis,
confident in data reconstruction (e.g., [43]). More importantly, the pile
Investigation, Data curation, Writing - original draft, Writing - review &
lateral responses interpreted by the proposed method, which involves
editing. Yu Wang: Conceptualization, Methodology, Investigation,
differentiation of reconstructed data, are very consistent with those
Resources, Writing - original draft, Writing - review & editing,
from LPILE, as shown in Fig. 12b–d. Similar observations are observed
Visualization, Supervision, Project administration, Funding acquisition.

Deflection (mm) Bending moment (kN m) Shear force (kN) Soil resistance (kN/m)
-10 0 10 20 30 -100 0 100 200 300 400 -200 -100 0 100 200 -400 -200 0 200 400
0
(a) (b) (c) (d)
2

4
Depth, z(m)

Deflection data
10 from LPILE Bending moment Soil resistance
Shear force from
Sparse deflection from LPILE LPILE from LPILE
data y Best estimate Best estimate shear Best estimate soil
12
Best estimate bending moment force resistance
deflection data Best estimate 2SD Best estimate 2SD Best estimate 2SD
14 Best estimate 2SD

Fig. 12. Pile lateral response interpreted from discrete inclinometer data with Gaussian white noise.

10
T. Zhao and Y. Wang Computers and Geotechnics 122 (2020) 103537

Deflection (mm) Bending moment (kN m) Shear force (kN) Soil resistance (kN/m)
-10 0 10 20 30 -100 0 100 200 300 400 -200 -100 0 100 200 -400 -200 0 200 400
0
(a) (b) (c) (d)
2

4
Depth, z(m)

Deflection data
10 from LPILE Bending moment Soil resistance
Shear force from
Sparse deflection from LPILE LPILE from LPILE
data y Best estimate Best estimate shear Best estimate soil
12
Best estimate bending moment force resistance
deflection data Best estimate 2SD Best estimate 2SD Best estimate 2SD
14 Best estimate 2SD

Fig. 13. Pile lateral response interpreted from discrete inclinometer data with uniform white noise.

Declaration of Competing Interest parameter effects. J Soil Mech Found Div 1972;98(SM6):603–24.
[18] Hetenyi M. Beams on elastic foundation: theory with applications in the fields of
civil and mechanical engineering. Ann Arbor, Michigan: University of Michigan;
The authors declare that they have no known competing financial 1946.
interests or personal relationships that could have appeared to influ- [19] Hong Y, He B, Wang LZ, Wang Z, Ng CWW, Mašín D. Cyclic lateral response and
failure mechanisms of semi-rigid pile in soft clay: centrifuge tests and numerical
ence the work reported in this paper. modelling. Can Geotech J 2017;54(6):806–24.
[20] Huang Y, Beck JL, Wu S, Li H. Bayesian compressive sensing for approximately
Acknowledgments sparse signals and application to structural health monitoring signals for data loss
recovery. Probabilist Eng Mech 2016;46:62–79.
[21] Hyman JM, Larrouturou B. The numerical differentiation of discrete functions using
The work described in this paper was supported by grants from the polynomial interpolation methods. Appl Math Comput 1982;10:487–506.
Research Grants Council of the Hong Kong Special Administrative [22] Ilyas T, Leung CF, Chow YK, Budi SS. Centrifuge model study of laterally loaded pile
groups in clay. J Geotech Geoenviron Eng 2004;130(3):274–83.
Region, China (project no. 9042516 (CityU 11213117) and project no.
[23] Jardine R. Nonlinear stiffness parameters from undrained pressure meter tests. Can
8779012 (T22-603/15N)). The financial support is gratefully ac- Geotech J 1992;29(3):436–47.
knowledged. [24] Ji S, Xue Y, Carin L. Bayesian compressive sensing. IEEE Trans Signal Process
2008;56(6):2346–56.
[25] Ji S, Dunson D, Carin L. Multitask compressive sensing. IEEE Trans Signal Process
References 2009;57(1):92–106.
[26] Johnson RA, Wichern DW. Applied multivariate statistical analysis. New Jersey, NJ:
[1] Babacan SD, Molina R, Katsaggelos AK. Bayesian compressive sensing using laplace Prentice hall; 2002.
priors. IEEE Trans Image Process 2010;19(1):53–63. [27] Jonas Lundgren (2020). SPLINEFIT (https://www.mathworks.com/matlabcentral/
[2] Baecher GB. Bayesian thinking in geotechnics. Geo-Risk 2017: Geotechnical Risk fileexchange/71225-splinefit), MATLAB Central File Exchange. Accessed at
from Theory to Practice. ASCE, Colorado, USA; 2017. p. 1–18. February 4, 2020.
[3] Brandenberg SJ, Wilson DW, Rashid MM. Weighted residual numerical differ- [28] Juang CH, Zhang J. Bayesian methods for geotechnical applications—a practical
entiation algorithm applied to experimental bending moment data. J Geotech guide. In: Geotechnical Safety and Reliability, GSP 286, Reston, VA, USA; 2017. p.
Geoenviron Eng ASCE 2010;136(6):854–63. 215–246.
[4] Briaud JL. Geotechnical engineering: unsaturated and saturated soils. New Jersey: [29] Kong L, Zhang L. Rate-controlled lateral-load pile tests using a robotic manipulator
John Wiley & Sons. Hoboken; 2013. p. 283–6. in centrifuge. Geotech Test J ASTM 2006;30(3):192–201.
[5] Brown DA, Hidden SA, Zhang S. Determination of p-y curves using inclinometer [30] Li D-Q, Zhang F-P, Cao Z-J, Zhou W, Phoon K-K, Zhou C-B. Efficient reliability
data. Geotech Test J ASTM 1994;17(2):150–8. updating of slope stability by reweighting failure samples generated by Monte Carlo
[6] Candès EJ, Wakin MB. An introduction to compressive sampling. IEEE Signal Proc simulation. Comput Geotech 2015;69:588–600.
Mag 2008;25(2):21–30. [31] Li X-S, Yang J, Liu H. Differentiation of noisy experimental data for interpretation of
[7] Candès EJ, Romberg JK, Tao T. Stable signal recovery from incomplete and in- nonlinear stress-strain behavior. J Eng Mech ASCE 1998;124(7):705–12.
accurate measurements. Commun Pure Appl Math 2006;59(8):1207–23. [32] Lin S-S, Liao J-C, Chen JT, Chen L. Lateral performance of piles evaluated via in-
[8] Cao Z-J, Zheng S, Li D, Phoon KK. Bayesian identification of soil stratigraphy based clinometer data. Comput Geotech 2005;32(6):411–21.
on soil behavior type index. Can Geotech J 2018. [33] McClelland B, Focht JA. Soil modulus for laterally loaded piles. Trans ASCE
[9] Choo YW, Kim D. Experimental development of the p-y relationship for large-dia- 1958;123(1):1049–63.
meter offshore monopiles in sands: centrifuge tests. J Geotech Geoenviron Eng ASCE [34] Murio D. Automatic numerical differentiation by discrete mollification. Comput
2016;142(1):04015058. Math Appl 1987;13(4):381–6.
[10] Clarke BG. Pressuremeters in geotechnical design. Glasgow, U.K.: Blackie and Sons; [35] Prust R, Davies J, Hu S. Pressuremeter investigation for mass rapid transit in
1995. Bangkok, Thailand. J Trans Res B 2005;1928:205–17.
[11] Comerford L, Kougioumtzoglou IA, Beer M. Compressive sensing based stochastic [36] Robertson PK, Ferreira RS. Seismic and pressuremeter testing to determine soil
process power spectrum estimation subject to missing data. Probabilist Eng Mech modulus. In: Proc., Predictive Soil Mechanics: Proceedings of the Wroth Memorial
2016;44:66–76. Symposium, London, UK; 1993. p. 562–580.
[12] Comerford L, Jensen H, Mayorga F, Beer M, Kougioumtzoglou I. Compressive [37] Salomon D. Data compression: the complete reference. New York, USA: Springer
sensing with an adaptive wavelet basis for structural system response and reliability Science & Business Media; 2007.
analysis under missing data. Comput Struct 2017;182:26–40. [38] Sinnreich J, Ayithi A. Derivation of py curves from lateral pile load test instrument
[13] Donoho DL. Compressed sensing. IEEE Tran Inf Theory 2006;52(4):1289–306. data. Geotech. Test. J. 2014;37(6):1075–86.
[14] Dou F-F, Fu C-L, Ma Y-J. A wavelet-Galerkin method for high order numerical [39] Teodorescu P, Stanescu N, Pandrea N. Numerical analysis with applications in
differentiation. Appl Math Comput 2010;215(10):3702–12. mechanics and engineering. Hoboken, NJ: John Wiley; 2013.
[15] Ensoft 2015, LPILE v8.0-A Program for the Analysis and Design of Piles and Drilled [40] Wang CH, Harken B, Osorio-Murillo CA, Zhu HH, Rubin Y. Bayesian approach for
Shafts under Lateral Loads, 2015. probabilistic site characterization assimilating borehole experiments and Cone
[16] Haiderali AE, Madabhushi G. Evaluation of curve fitting techniques in deriving p–y Penetration Tests. Eng. Geol. 2016;207:1–13.
curves for laterally loaded piles. Geotech Geol Eng 2016;34(5):1453–73. [41] Wang Y, O’Rourke TD. Interpretation of secant shear modulus degradation char-
[17] Hardin BO, Drnevich VP. Shear modulus and damping in soils: measurement and acteristics from pressuremeter tests. J Geotech Geoenviron Eng ASCE

11
T. Zhao and Y. Wang Computers and Geotechnics 122 (2020) 103537

2007;133(12):1556–66. [48] Zhao T, Wang Y. Interpretation of pile lateral response from deflection measure-
[42] Wang Y, Zhao T. Interpretation of soil property profile from limited measurement ment data: a compressive sampling-based method. Soils Found 2018;58(4):957–71.
data: a compressive sampling perspective. Can Geotech J 2016;53(9):1547–59. [49] Zhao T, Hu Y, Wang Y. Statistical interpretation of spatially varying 2D geo-data
[43] Wang Y, Zhao T. Statistical interpretation of soil property profiles from sparse data from sparse measurements using Bayesian compressive sampling. Eng Geol
using Bayesian compressive sampling. Géotechnique 2017;67(6):523–36. 2018;246:162–75.
[44] Wang Y, Zhao T, Cao Z. Site-specific probability distribution of geotechnical [50] Zhao T, Montoya-Noguera S, Phoon KK, Wang Y. Interpolating spatially varying soil
properties. Comput Geotech 2015;70:159–68. property values from sparse data for facilitating characteristic value selection. Can
[45] Wei T, Hon Y, Wang Y. Reconstruction of numerical derivatives from scattered Geotech J 2018;55(2):171–81.
noisy data. Inverse Probl 2005;21(2):657–72. [51] Reese LC, Van Impe WF. Single piles and pile groups under lateral loading. 2nd ed. Boca
[46] Wilson DW, Boulanger RW, Kutter BL. Observed seismic lateral resistance of li- Raton, Florida: CRC press; 2010.
quefying sand. J Geotech Geoenviron Eng ASCE 2000;126(10):898–906. [52] Ooi PSK, Ramsey TL. Curvature and bending moments from inclinometer data. Int J
[47] Yang K, Liang R. Methods for deriving py curves from instrumented lateral load Geomech 2003;3(1):64–74.
tests. Geotech Test J 2006;30(1):31–8.

12

You might also like