Professional Documents
Culture Documents
To cite this article: Xinhua Xue & Ming Xiao (2016): Application of adaptive neuro-fuzzy
inference system for prediction of internal stability of soils, European Journal of Environmental
and Civil Engineering, DOI: 10.1080/19648189.2016.1271363
Article views: 3
Download by: [The UC San Diego Library] Date: 28 December 2016, At: 00:00
European Journal of Environmental and Civil Engineering, 2016
http://dx.doi.org/10.1080/19648189.2016.1271363
This paper presents an adaptive neuro-fuzzy inference system (ANFIS) for the
assessment of internal stability of soils under seepage. The training of fuzzy system
was performed by a hybrid method of back-propagation (BP) and least mean square
algorithm, and the subtractive clustering algorithm was utilised for optimising the
number of fuzzy rules. Experimental data on internal stability of soils in the literature
were used to validate and evaluate the performance of the proposed ANFIS model.
Further, predictions from three models (the BP model, the particle swarm optimisa-
tion-BP model and the ANFIS model) were compared with the experimental data.
The results show that the ANFIS model is a feasible, efcient and accurate tool for
predicting the internal stability of soils according to Wan and Fells criterion.
Keywords: internal stability; adaptive neuro-fuzzy inference system; particle swarm
optimisation; back propagation; subtractive clustering algorithm
1. Introduction
Seepage-induced internal erosion of soils is widely observed in dams, levees and other
earthen embankments. Internal erosion occurs when soil particles within an earthen
embankment or its foundation are carried downstream by seepage. It starts when the
erosive forces imposed by the hydraulic loads exceed the resistance of the materials in
the embankment to erosion. Internal erosion customarily includes suffusion and
piping phenomena. Suffusion is the entrainment of ne soil grains and their possible
subsequent deposition in the pores that are formed by coarse grains; suffusion may be
caused by discontinuity or segregation of soil particles and it can form large local
cavities. In piping, soil particles inside the soil matrix are entrained and washed out of
the matrix by concentrated seepage, forming a tubular pipe that progresses from
downstream to upstream; the pipe can develop into a large tunnel that may collapse.
Suffusion and piping may result in the increase of porosity and compressibility, settle-
ment, variation of permeability and the potential for a reduction in soil strength (Xiao &
Shwiyhat, 2012). Internal erosion can occur within an earthen embankment or its
foundation, along conduits and their outlet works, spillway walls or concrete gravity
structure supporting the embankment. In a survey of 11,192 dams, Foster, Fell, and
Spannagle (2000) concluded that approximately 46% of the dam failures among
the entire survey sample could be attributed to internal erosion. Internal erosion in
embankment dams is one of the main causes of failures and accidents of dams, and the
development of internal erosion can pose a serious threat to the stability of these water
retaining structures (Bonelli, 2013).
The threat of internal erosion to earthen dams and levees has long been recognised
and its mechanisms and behaviors have been studied in the past century. The rst engi-
neers who analysed this problem were probably Bligh (1910) and Lane (1935), who
dened the soils susceptibility to erosion using a percolation factor that took into
account of soil type, seepage path and hydraulic head difference. Recently, Foster et al.
(2000) evaluated the probability of failure of dams by internal erosion and dened a
four-stage process consisting of initiation of erosion, continuation of erosion, progres-
sion to form a pipe or occasionally cause surface instability and initiation of a breach.
Fell, Wan, Cyganiewicz, and Foster (2003) presented a new approach to estimate the
time of internal erosion development in embankment dams.
Many factors affect the internal stability of a soil, such as geometric conditions (i.e.
grain sizes and their distribution, pore sizes and their distribution, and pore constriction
sizes), hydraulic conditions (i.e. hydraulic gradient or ow velocity, and ow direction)
and mechanical conditions (i.e. compaction and soil strength) (Chang & Zhang, 2013).
These factors inherently exhibit the characteristics of randomness (i.e. they follow cer-
tain probabilistic distributions), fuzziness (i.e. they are difcult to determine) and spatial
variability. Furthermore, denitive consideration of all the pertinent parameters is not
presently possible because some factors are not yet clearly understood or their effects
on soils stability are difcult to quantify. With such limitations, internal erosion predic-
tion calls for new and innovative methods such as the back-propagation (BP) neural net-
works, neuro-fuzzy networks that are able to account for randomness, fuzziness and
variability. However, the standard BP neural network model exhibits slow convergence,
local minima and lack of robustness. An efcient training algorithm, such as the particle
swarm optimisation (PSO) developed by Kennedy and Eberhart (1995), may provide
better optimisation for BP neural network model. In addition to the BP neural networks,
neuro-fuzzy networks, which combine both advantages of the fuzzy systems (e.g.
humanlike if-then rules) and the learning and optimisation abilities of the articial neu-
ral networks (ANNs), may be one of the most competent articial intelligence subsys-
tems to predict the internal erosion potential. Moreover, the network model obtained
this way would not remain a black box, since this model would have fuzzy logic capa-
bilities to interpret in terms of linguistic variables. A linguistic variable is a variable
whose values are words in a natural language, and it plays a key role in its applications,
especially in the fuzzy expert system. The main idea behind the neuro-fuzzy systems is
that they learn adaptively from experience and extract various discriminants, each appro-
priate for its purpose. Neuro-fuzzy networks have been used successfully in geotechni-
cal engineering such as the evaluation of slope failure potential (Ni, Lu, & Juang,
1996), mapping of cone penetration test (CPT) values into soil dynamics properties
(Romo & Garca, 2003), settlement prediction of shallow foundations on granular soil
(Shahin, Maier, & Jaksa, 2003) and prediction of the pullout capacity of marquee
ground anchors (Shahin & Jaksa, 2005).
In this study, the adaptive neuro-fuzzy inference system (ANFIS) is developed for
the assessment of internal stability of soils. The training of fuzzy system was performed
by a hybrid method of back-propagation (BP) and the least mean square (LMS) algo-
rithm, and the subtractive clustering algorithm (SCA) was utilised for optimising the
number of fuzzy rules. In addition, we also compared the prediction performances of
ANFIS, PSO-based BP neural network and a standard Levenberg Marquardt (L-M)
European Journal of Environmental and Civil Engineering 3
2. Model development
2.1. Theory of adaptive neuro-fuzzy inference system (ANFIS)
2.1.1. Fuzzy logic
The modelling of many systems involves the consideration of some uncertain variables.
Besides the statistical uncertainties that handle variables through probability theory,
there also exists non-statistical uncertainty that handles variables in a rational framework
of fuzzy set theory (Chern, Lee, & Wang, 2008). According to Hisdal (1994), the great
achievements of fuzzy set theory are as follows: (1) the introduction of linguistic values
of variables, and (2) the introduction of the possibility of a partial grade of a member-
ship value of an object in a class instead of having a choice solely between the member-
ship values of 0 and 1 (in crisp sets).
In a classical set, the set has a crisp boundary, and the membership value of an
object in a class is only 0 or 1. The membership or non-membership of an element x in
the crisp set A is represented by the characteristic function of A, dened by
1 if x 2 A
lA x (1)
0 if x 62 A
Mathematically, the fuzzy set A is
A f x; lA xjx 2 U g (2)
where U refers to the universe of discourse dened for a specic problem and lA x is
the membership degree of the variable x that is dened as
lA x ! 0; 1 (3)
Figure 1 shows the basic framework of a fuzzy set that contains several functions:
fuzzication, fuzzy rule base, decision-making unit and defuzzication. In addition to
establishing a fuzzy rule base through linguistic fuzzy rules that are transformed from
expert knowledge and experience, a fuzzy rule base can also be established through
special algorithmic rules that obtain inputs and outputs with mathematical calculation. This
study adopted the SCA to analyse the hidden rules and to establish a fuzzy rule base.
parts. For simplicity, the ANFIS structure is set to have two inputs x and y as shown in
Figure 2 and explained as follows.
This fuzzy model has rules in the following forms:
where A1, A2, B1 and B2 are the membership functions for inputs x and y, respectively;
p1, q1, r1, p2, q2 and r2 are the parameters of the output function. The functioning of
each layer is described as follows (Jang, 1993):
Layer 1: Each node i in this layer generates membership grades of an input variable.
For instance, the node function of the ith node might be
1
O1;i lAi x 2 bi (4)
1 xci
ai
where x is the input to node i, and Ai is the linguistic label associated with this node,
and fai ; bi ; ci g is the parameter set that changes the shapes of the membership function.
Parameters in this layer are referred to as the premise parameters.
European Journal of Environmental and Civil Engineering 5
Layer 2: Each node in this layer calculates the ring strength of each rule via mul-
tiplication:
Layer 3: The ith node of this layer, labelled as N, computes the normalised ring
strength as
wi
i
O3;i w i 1; 2 (6)
w1 w2
Layer 4: Node i in this layer computes the contribution of the ith rule towards the
model output, with the following node functions:
i fi w
O4;i w i pi x qi y r i (7)
Layer 5: The single node in this layer computes the overall output of the ANFIS as
X P
i fi
w
O5;i i fi Pi
w (8)
i wi
where kk denotes the Euclidean distance, and ra is a positive constant representing a
neighbourhood radius. The potential of a data point to be a cluster centre is when it has
a high density value, which means that more data points are closer to it.
Let xc1 be the rst cluster centre and Dc1 is the potential value. The potential value
for each data point is dened as:
" #
kxi xc1 k2
Di Di Dc1 exp 2 (10)
r b
2
where rb is a positive constant that denes a neighbourhood that has measurable reduc-
tions in density. The constant rb is normally larger than ra to prevent closely spaced
cluster centres; generally rb is equal to 1.5 ra. After the density for each data point is
recalculated, the next cluster center xc2 is selected and all of the densities for data points
are recalculated again. This process is repeated until
individual best objective value obtained so far at time t. The global best particle is
denoted as g, which represents the best position found so far at time t in the whole
swarm. The new velocity of each particle is given as follows
vi;j t 1 jvi;j t c1 r1 pi;j xi;j t c2 r2 gj xi;j t ; j 1; 2; . . .; d (13)
where c1 and c2 are the acceleration coefcients, usually c1 = c2 = 2; r1 and r2 are two
independent random numbers uniformly distributed in the range [0,1]; is called the
inertia weight factor.
The position of each particle is then updated in each generation according to the
following equation:
xi;j t 1 xi;j t vi;j t 1; j 1; 2; . . .; d (14)
In PSO, Equation (13) is applied to calculate the new velocity according to its previous
velocity and to the distance of its current position from both its own best historical
position and its neighbours best position. Generally, the value of each component in Vi
can be constrained to the rangevmin ; vmax to control the excessive roaming of particles out-
side the search space. Then, the particle ies towards a new position according to Equation
(14). This process is repeated until the maximum number of generations is reached.
in which y is a normalised input parameter, x is the original input parameter, xmax and
xmin are the maximum and minimum parameters, respectively.
Step 2: Initialise the parameters of PSO, such as each particles velocity and
position, the acceleration coefcients c1 and c2, the inertia weight factor .
Step 3: Update each particles velocity and position according to Equations (13) and
(14), respectively.
Step 4: Evaluate each particles tness. In this study, the RMSE (as shown in
Equation (12)) of BP neural network is used as the tness function to guide the particle
population to search for the optimum solution.
Step 5: Update the personal best position pbest and the global best position gbest.
Step 6: Choose the maximum of the pbest as the current gbest of the particle.
Step 7: If the maximum of the iteration is achieved or the optimum solution is
acquired, then the algorithm is stopped, else return to Step 3.
The ow chart of the PSO-BP algorithm is shown in Figure 4.
3. Case study
In this study, the following six factors including the clay content (<0.005 mm), nes
content (<0.075 mm), sand fraction (0.0754.75 mm), gravel fraction (>4.75 mm), coef-
cient of uniformity Cu and coefcient of curvature Cc were taken into account as the
input parameters of the models of ANFIS, BP and PSO-BP. The database used in this
study was selected from Wan and Fell (2004). A few of the original records are omitted
because of incomplete data. The database includes 62 internal stability tests. The rst
50 of the total data were used to train the ANFIS model, whereas the remaining 12 of
the data were used to verify the accuracy and the effectiveness of the trained ANFIS
model for prediction of internal stability of soils. In this study, 1 represents stable and
0 denotes unstable, known as the nal classiers. In the determination of the nal
classier, the soils internal stability criterion was implicitly dened in the experimental
conditions and observations in the training data-sets. The detailed training and testing
datasets are summarised in Table 1.
Table 1 lists the inuencing factors in terms of grain size distribution. It should be
noted that other work has shown that the dry density, water content, liquid limit and satu-
ration ratio all play a role in the internal erosion potential or resistance (e.g. Regazzoni &
Marot, 2011). In this study, model predictions and comparisons were based on the internal
stability criteria by Wan and Fell (2004), which are based on particle gradations.
The ANFIS model structure consists of 6 inputs, 14 input membership functions
(MFs), 14 output membership functions (MFs) and 1 output, as shown in Figure 5. Dif-
ferent parameter types and their values used for training ANFIS are shown in Table 2.
The RMSE convergence curve of ANFIS was derived as shown in Figure 6. From the
curve, the nal convergence value is 1.1 102. In addition, the nal Gaussian-shaped
MFs of the input parameters after training are shown in Figures 712, respectively.
Notice that all the inputs and outputs in Figures 712 have exactly 14 membership
functions. The 14 membership functions represent the 14 clusters that were identied by sub-
cluster. By default, the rst membership function (the red curve as shown in Figures 712)
would be selected in the membership function editor. For example, the parameters of the rst
membership function named in1cluster1 (the red curve as shown in Figure 7) are
2:292 0:005948, where 2.292 represents the spread coefcient of the Gaussian curve
and 0.005948 represents the centre of the Gaussian curve. In1cluster1 captures the position
and inuence of the rst cluster for the input variable clay content. Similarly, the position and
inuence of other clusters for the input variables are captured by the other membership func-
tions. The parameters of the rst membership functions in Figures 812 are 4:688 0:826,
9:062 29:9, 8:876 69:18, 84:51 13:2, 6:966 1:343, respectively.
In addition, there are 14 rules generated based on ANFIS modelling and they are
listed as follows:
(1) If clay content is in1cluster1, and nes content is in2cluster1, and sand fraction
is in3cluster1, and gravel fraction is in4cluster1, and Cu is in5cluster1, and Cc
is in6cluster1, then out1 is out1cluster1 (1);
(2) If clay content is in1cluster2, and nes content is in2cluster2, and sand fraction
is in3cluster2, and gravel fraction is in4cluster2, and Cu is in5cluster2, and Cc
is in6cluster2, then out1 is out1cluster2 (1);
(3) If clay content is in1cluster3, and nes content is in2cluster3, and sand fraction
is in3cluster3, and gravel fraction is in4cluster3, and Cu is in5cluster3, and Cc
is in6cluster3, then out1 is out1cluster3 (1);
10
Table 1. Data-set used in model training and predictions in case study (1: stable; 0: unstable) (data from Wan and Fell (2004)).
Clay content Fines content Sand fraction (0.075 Gravel fraction Coeff. of Coeff. of Internal
Datasets No. (<0.005 mm) (%) (<0.075 mm) (%) 4.75 mm) (%) (>4.75 mm) (%) uniformity (Cu) curvature (Cc) stability
1 0.0 0.0 60.0 40.0 3.1 0.80 1
2 0.0 0.0 100.0 0.0 3.1 0.80 1
3 0.0 0.0 64.1 35.9 6.2 0.69 1
4 0.0 0.0 36.2 63.8 6.4 0.69 1
5 0.0 0.0 44.3 55.7 14.1 0.59 1
6 0.0 0.0 51.8 48.2 13.8 0.59 1
X. Xue and M. Xiao
Table 2. Different parameter types and their values used for training ANFIS.
ANFIS parameter type ANFIS (SCA)
Number of fuzzy rules 14
Number of nodes 205
Number of linear parameters 98
Number of nonlinear parameters 168
Total number of parameters 266
Number of training data sets 50
Number of verication data sets 12
(4) If clay content is in1cluster4, and nes content is in2cluster4, and sand fraction
is in3cluster4, and gravel fraction is in4cluster4, and Cu is in5cluster4, and Cc
is in6cluster4, then out1 is out1cluster4 (1);
European Journal of Environmental and Civil Engineering 13
(5) If clay content is in1cluster5, and nes content is in2cluster5, and sand fraction
is in3cluster5, and gravel fraction is in4cluster5, and Cu is in5cluster5, and Cc
is in6cluster5, then out1 is out1cluster5 (1);
(6) If clay content is in1cluster6, and nes content is in2cluster6, and sand fraction
is in3cluster6, and gravel fraction is in4cluster6, and Cu is in5cluster6, and Cc
is in6cluster6, then out1 is out1cluster6 (1);
(7) If clay content is in1cluster7, and nes content is in2cluster7, and sand fraction
is in3cluster7, and gravel fraction is in4cluster7, and Cu is in5cluster7, and Cc
is in6cluster7, then out1 is out1cluster7 (1);
(8) If clay content is in1cluster8, and nes content is in2cluster8, and sand fraction
is in3cluster8, and gravel fraction is in4cluster8, and Cu is in5cluster8, and Cc
is in6cluster8, then out1 is out1cluster8 (1);
(9) If clay content is in1cluster9, and nes content is in2cluster9, and sand fraction
is in3cluster9, and gravel fraction is in4cluster9, and Cu is in5cluster9, and Cc
is in6cluster9, then out1 is out1cluster9 (1);
European Journal of Environmental and Civil Engineering 15
(10) If clay content is in1cluster10, and nes content is in2cluster10, and sand frac-
tion is in3cluster10, and gravel fraction is in4cluster10, and Cu is in5cluster10,
and Cc is in6cluster10, then out1 is out1cluster10 (1);
(11) If clay content is in1cluster11, and nes content is in2cluster11, and sand frac-
tion is in3cluster11, and gravel fraction is in4cluster11, and Cu is in5cluster11,
and Cc is in6cluster11,then out1 is out1cluster11 (1);
(12) If clay content is in1cluster12, and nes content is in2cluster12, and sand frac-
tion is in3cluster12, and gravel fraction is in4cluster12, and Cu is in5cluster12,
and Cc is in6cluster12, then out1 is out1cluster12 (1);
(13) If clay content is in1cluster13, and nes content is in2cluster13, and sand frac-
tion is in3cluster13, and gravel fraction is in4cluster13, and Cu is in5cluster13,
and Cc is in6cluster13, then out1 is out1cluster13 (1);
16 X. Xue and M. Xiao
(14) If clay content is in1cluster14, and nes content is in2cluster14, and sand frac-
tion is in3cluster14, and gravel fraction is in4cluster14, and Cu is in5cluster14,
and Cc is in6cluster14, then out1 is out1cluster14 (1).
The rst rule can be simply explained as follows. If the inputs to the ANFIS, clay
content, nes content, sand fraction, gravel fraction, Cu, and Cc strongly belong to their
respective cluster membership functions (cluster1), then the output (out1) or the status
of stability must strongly belong to its cluster1 membership function. The (1) at the
end of each rule indicates that the rule has a weight or an importance of 1. Weights
can take any value between 0 and 1. Rules with less weights count less in the nal
output.
The developed fuzzy model can provide a precise evaluation of soils internal erosion
potential based on Wan and Fells criterion (2004) once we enter proper input data.
Figure 13 shows a model application in MATLAB environment. When input parameters
are clay content = 9.46%, nes content = 24.1%, sand fraction = 42.7%, gravel
fraction = 33.2%, Cu = 213 and Cc = 53, the status of stability (Out 1) would be 0.085
(Figure 13), which means the status of stability is unstable. Since the model has the
ability of interpolating input parameters, that is, if we take any value between the mini-
mum and the maximum of the data-set, the proposed ANFIS model is capable of pre-
dicting the internal stability of soils under seepage according to Wan and Fells
criterion.
To evaluate the performance of the proposed ANFIS method, a comparison between
the experimental results and the predictions by the PSO-BP model, the BP model and
the ANFIS model are made and shown in Figure 14 and Tables 34, respectively. It
should be noted that it is very important to select the number of hidden layers and the
number of neurons in various layers before using the BP neural network. The number
of neurons in input and output layers is usually dictated by the nature of the problem.
In this study, there are six parameters including the clay content (<0.005 mm), nes
content (<0.075 mm), sand fraction (0.0754.75 mm), gravel fraction (>4.75 mm), coef-
cient of uniformity Cu, and coefcient of curvature Cc were taken into account as the
input parameters; therefore, the number of neurons in input layers is 6. As mentioned,
the main objective of this paper is to predict the potential of internal stability of soils
under seepage, so the number of neurons in output layer is 1 and one hidden layer BP
neural network is adopted herein. For the number of neurons in hidden layer, the main
strategy is to use as few hidden layer neurons as possible because each unit adds to the
loads on the CPU during simulations. If the network fails to converge to a solution, it
means that more hidden neurons are required. If it does converge we might try for fewer
hidden neurons. Based on this idea, the number of hidden neurons was determined by
trial and found suitable network with ve neurons in hidden layer. Thus, the structure
of BP neural network is designed as 6-5-1. Many kinds of transfer functions have been
proposed in the literature and one of the most popular hidden layer transfer functions is
the tangent sigmoid function; therefore, the tangent sigmoid transfer function is
employed in the hidden layer herein. Because the pureline transfer function is sufcient
for BP neural network to approximate almost any complex function; therefore, it is
employed in the output layer in this study. In addition, the important parameters of PSO
used in this study are given in Table 5. In this case, it is assumed that the acceleration
coefcients c1 = c2 = 2 and inertia weight j0:5.
As shown in Table 4, the performance of ANFIS model is superior to those of BP
and PSO-BP models. For example, the total accuracy rates of ANFIS, BP and PSO-BP
models are 100, 96.77 and 100%, respectively. Whereas the total RMSE values of
ANFIS, BP and PSO-BP models are 0.038, 0.486 and 0.105, respectively. Clearly, the
smaller the RMSE values, the better the prediction accuracy and vice versa. However,
different from the PSO-BP model, the ANFIS model has the fuzzy logic capabilities to
interpret in terms of linguistic variables, while this capability lacks in the PSO-BP
model. These results indicate that the ANFIS is a valid tool to predict the internal stabil-
ity of soils under seepage according to Wan and Fells criterion (2004).
18 X. Xue and M. Xiao
4. Conclusions
In this study, the adaptive neuro-fuzzy inference system (ANFIS) is developed for the
assessment of internal erosion potential of soils. The training of fuzzy system was per-
formed by a hybrid method of BP-LMS algorithm and the SCA was utilised for optimis-
ing the number of fuzzy rules. In addition, we compared the prediction performances of
ANFIS, PSO-BP neural network model and a standard LevenbergMarquardt algo-
rithm-based BP neural network model. The performances of these techniques have been
evaluated in terms of the training performance and classication accuracies. The results
conrmed that the proposed ANFIS model is capable of predicting the internal stability
of soils under seepage according to Wan and Fells criterion (Wan & Fell, 2004). The
following conclusions are drawn from this study.
(1) The developed fuzzy model, ANFIS, can provide a precise evaluation of internal
stability if proper input data are provided. The model has the ability of interpolating
input parameters and the predictions can be made in various conditions. ANFIS may be
one of the most competent articial intelligence subsystems to evaluate the internal ero-
sion potential and stability of soils. In addition, the developed ANFIS model could be
used to predict other criteria proposed by the literature if proper input data are provided.
It could be an interesting tool to compare several criteria.
(2) The standard BP neural network model exhibits slow convergence, local minima
and lack of robustness. As an efcient training algorithm, PSO can be a good optimisa-
tion tool for the BP neural network model.
Disclosure statement
No potential conict of interest was reported by the authors.
References
Banerjee, T., & Das, S. (2012). Multi-sensor data fusion using support vector machine for motor
fault detection. Information Sciences, 217, 96107. doi:10.1016/j.ins.2012.06.016
Bligh, W. G. (1910). Dams, barrages and weirs on porous foundations. Engineering News, 64,
708710.
Bonelli, S. Ed. (2013). Erosion in Geomechanics Applied to Dams and Levees. ISTE Ltd:
London, UK; John Wiley & Sons: Hoboken, NJ, USA.
Chang, D. S., & Zhang, L. M. (2013). Extended internal stability criteria for soils under seepage.
Soils and Foundations, 53, 569583. doi:10.1016/j.sandf.2013.06.008
Chern, S. G., Lee, C. Y., & Wang, C. C. (2008). CPT-based liquefaction assessment by using
fuzzy-neural network. Journal of Marine Science and Technology, 16, 139148. doi:10.6119/
JMST
Chiu, S. L. (1994). Fuzzy model identication based on cluster estimation. Journal of Intelligent
& Fuzzy Systems, 2, 267278. doi:10.3233/IFS-1994-2306
European Journal of Environmental and Civil Engineering 19
Eftekhari Zadeh, E., Feghhi, S. A. H., Roshani, G. H., & Rezaei, A. (2016). Application of articial
neural network in precise prediction of cement elements percentages based on the neutron activa-
tion analysis. The European Physical Journal Plus, 131, 167. doi:10.1140/epjp/i2016-16167-6
Fell, R., Wan, C. F., Cyganiewicz, J., & Foster, M. (2003). Time for development of internal
erosion and piping in embankment dams. Journal of Geotechnical and Geoenvironmental
Engineering, 129, 307314. doi:10.1061/(ASCE)1090-0241(2003)129:4(307)
Fell, R., & Fry, J. J. (2013). State of the art on the likelihood of internal erosion of dams and
levees by means of testing. In S. Bonelli (Ed.), Erosion in geomechanics applied to dams and
levees (pp. 199). London, UK: ISTE; Hoboken, NJ, USA: John Wiley & Sons.
Foster, M., Fell, R., & Spannagle, M. (2000). The statistics of embankment dam failures and
accidents. Canadian Geotechnical Journal, 37, 10001024. doi:10.1139/t00-030
Garner, S. J., & Fannin, R. J. (2010). Understanding internal erosion: A decade of research fol-
lowing a sinkhole event. International Journal on Hydropower & Dams, 17, 9398.
Ghosh, S., Das, S., Kundu, D., Suresh, K., Panigrahi, B. K., & Cui, Z. H. (2012). An inertia-
adaptive particle swarm system with particle mobility factor for improved global optimization.
Neural Computing and Applications, 21, 237250. doi:10.1007/s00521-010-0356-x
Hisdal, E. (1994). Interpretative versus prescriptive fuzzy set theory. IEEE Transactions on Fuzzy
Systems, 2, 2226. doi:10.1109/91.273118
Jang, J. S. R. (1993). ANFIS: Adaptive-network-based fuzzy inference system. IEEE Transactions
on Systems, Man and Cybernetics, 23, 665685.
Kennedy, J., & Eberhart, R. C. (1995). Particle swarm optimization. In Proceedings of the 1995
IEEE International Conference on Neural Networks 4, Perth, Australia. (pp. 19421948). Pis-
cataway, NJ: IEEE Service Center.
Lane, E. W. (1935). Security from under-seepage masonry dams on earth foundations. Transac-
tions of the American Society of Civil Engineers, 100, 12351272.
Li, M., & Fannin, R. J. (2008). Comparison of two criteria for internal stability of granular soil.
Canadian Geotechnical Journal, 45, 13031309. doi:10.1139/T08-046
Marot, D., Rochim, A., Nguyen, H. H., Bendahmane, F., & Sibille, L. (2014, December 24). Sys-
tematic methodology for characterization of suffusion sensibility. In Scour and Erosion: Pro-
ceedings of the 7th International Conference on Scour and Erosion. Perth, Australia: CRC Press.
Nasir, M., Das, S., Maity, D., Sengupta, S., Halder, U., & Suganthan, P. N. (2012). A dynamic
neighborhood learning based particle swarm optimizer for global numerical optimization.
Information Sciences, 209, 1636. doi:10.1016/j.ins.2012.04.028
Ni, S. H., Lu, P. C., & Juang, C. H. (1996). A fuzzy neural network approach to evaluation of
slope failure potential. Computer-Aided Civil and Infrastructure Engineering, 11, 5966.
doi:10.1111/j.1467-8667.1996.tb00309.x
Regazzoni, P. L., & Marot, D. (2011). Investigation of interface erosion rate by Jet Erosion Test
and statistical analysis. European Journal of Environmental and Civil Engineering, 15,
11671185. doi:10.1080/19648189.2011.9714847
Roshani, G. H., Feghhi, S. A. H., & Setayeshi, S. (2015). Dual-modality and dual-energy gamma
ray densitometry of petroleum products using an articial neural network. Radiation Measure-
ments, 82, 154162. doi:10.1016/j.radmeas.2015.07.006
Romo, M. P., & Garca, S. R. (2003). Neurofuzzy mapping of CPT values into soil dynamic prop-
erties. Soil Dynamics and Earthquake Engineering, 23, 473482. doi:10.1016/S0267-7261(03)
00066-6
Shahin, M. A. & Jaksa, M. B. (2005). Neural network prediction of pullout capacity of marquee
ground anchors. Computers and Geotechnics, 32, 153163. doi:10.1016/j.compgeo.2005.02.003
Shahin, M. A., Maier, H. R., & Jaksa, M. B. (2003). Settlement prediction of shallow foundations
on granular soils using B-spline neurofuzzy models. Computers and Geotechnics, 30, 637
647. doi:10.1016/j.compgeo.2003.09.004
Takagi, T. & Sugeno, M. (1985). Fuzzy identication of systems and its applications to modeling
and control. IEEE Transactions on Systems, Man and Cybernetics, SMC-15, 116132.
Wan, C. F., & Fell, R. (2004). Experimental investigation of internal instability of soils in
embankment dams and their foundations (University Report). Sydney: The University of New
Southwales.
Xiao, M. & Shwiyhat, N. (2012). Experimental investigation of the effects of suffusion on physi-
cal and geo-mechanical characteristics of sandy soils. ASTM Geotechnical Testing Journal,
35, 890900. doi:10.1520/GTJ104594