Vous êtes sur la page 1sur 11

Computers and Geotechnics 30 (2003) 637647

www.elsevier.com/locate/compgeo

Settlement prediction of shallow foundations on granular soils


using B-spline neurofuzzy models
M.A. Shahin, H.R. Maier, M.B. Jaksa*
School of Civil and Environmental Engineering, The University of Adelaide, SA, 5005, Australia

Received 28 February 2003; received in revised form 6 August 2003; accepted 1 September 2003

Abstract
The design of shallow foundations on granular soils is generally controlled by settlement rather than bearing capacity. As a
consequence, settlement prediction is a major concern and is an essential criterion in the design process of shallow foundations. At
present, consistent accurate prediction of settlement of shallow foundations on granular soils has yet to be achieved using many
numerical modelling techniques. Recently, multi-layer perceptrons (MLPs) trained with the back-propagation algorithm have been
applied successfully to settlement prediction of shallow foundations on granular soils. However, a shortcoming of MLPs is that the
knowledge that is acquired during training is distributed across their connection weights in a complex manner that is often dicult
to interpret. Consequently, the rules governing the relationships between the network input/output variables are dicult to quan-
tify. One way to overcome this problem is to use neurofuzzy networks in which the acquired knowledge can be translated into a set
of fuzzy rules that describe the relationships between the network inputs and the corresponding outputs in a transparent fashion. In
the present paper, the ability of neurofuzzy networks to predict settlement of shallow foundations on granular soils and to assist
with providing a better understanding regarding the relationships between settlement and the factors aecting settlement is asses-
sed. The sensitivity of neurofuzzy models to a number of stopping criteria is investigated and the models obtained are compared in
terms of prediction accuracy, model parsimony and model transparency. The impact of incorporating existing engineering knowl-
edge on neurofuzzy model performance and interpretation is also investigated. The type of neurofuzzy networks used in this
research are B-spline networks that are trained with the adaptive spline modelling of observation data (ASMOD) algorithm. The
results indicate that B-spline neurofuzzy networks are capable of predicting well the settlement of shallow foundations on granular
soils and are able to provide a transparent understanding of the relationships between settlement and the factors aecting it. It is
found from this research that neurofuzzy models that use the Bayesian Information Criterion (BIC) are able to strike a balance
between model accuracy, parsimony and transparency. The results also indicate that modifying neurofuzzy networks by incorpor-
ating existing engineering knowledge can improve model performance and enhance the interpretation of the constructed model.
# 2003 Elsevier Ltd. All rights reserved.
Keywords: Settlement prediction; Shallow foundations; Granular soils; Neurofuzzy networks

1. Introduction propagation multi-layer perceptrons (MLPs) [2,3] have


been applied successfully by Shahin et al. [4] to settle-
Settlement prediction is an essential criterion in the ment prediction of shallow foundations on granular
design of shallow foundations on granular soils since soils and have been found to outperform the most
settlement, rather than bearing capacity, generally con- commonly used traditional methods. Despite the high
trols the design process. Over the years, consistent and capability of data mapping of MLPs, the knowledge
accurate prediction of settlement of shallow foundations acquired in the trained network is stored in the connec-
on granular soils has yet to be achieved by the use of a tion weights in a complex manner that is often dicult
variety of methods ranging from purely empirical to to interpret. This makes it dicult to extract the
complex non-linear nite elements [1]. Recently, back- knowledge that has been learnt by the trained network
regarding the relationships between the model inputs
* Corresponding author. Tel.: +61-8-8303-4317; fax: +61-8-8303-
and the corresponding outputs and, thus, MLPs trained
4359. with the back-propagation algorithm are often criticised
E-mail address: mjaksa@civeng.adelaide.edu.au (M.B. Jaksa). for being black boxes. One way to overcome this
0266-352X/$ - see front matter # 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compgeo.2003.09.004
638 M.A. Shahin et al. / Computers and Geotechnics 30 (2003) 637647

problem is by the use of neurofuzzy networks. Neuro- and local linear model trees (LOLIMOT) [14], among
fuzzy networks combine the explicit linguistic knowl- others. In this research, B-spline neurofuzzy networks
edge representation of fuzzy systems with the learning that are trained with the ASMOD algorithm will be
power of MLPs [5,6]. One advantage of neurofuzzy used. B-spline neurofuzzy networks use the fuzzy logic
networks is that they can be trained by processing data system [15] to store the knowledge acquired between a
samples to perform input/output data mappings, similar set of input variables (x1, x2, . . ., xn) and the corre-
to the way MLPs do, with the additional benet of sponding output variable (y) in a set of linguistic fuzzy
being able to provide a set of fuzzy rules that can rules that can be easily interpreted, such as:
describe the model input/output relationships in a
transparent fashion [7,8]. One additional advantage of
IFx1 is high AND x2 is lowTHENy is high;
neurofuzzy networks is that available engineering
knowledge can be incorporated into the trained network c 0:9
to optimise model performance and to enhance the where (c=0.9) is the rule condence which indicates
interpretation of a constructed model. In geotechnical the degree to which the above rule has contributed to
engineering, Ni et al. [9] have already applied a neuro- the output. As part of any fuzzy logic system, two
fuzzy network approach to the evaluation of slope fail- main components need to be determined (i.e. fuzzy sets
ure potential. However, a review of the literature and fuzzy rules). In order to determine the fuzzy sets,
indicates that neurofuzzy networks are new tools in the linguistic terms (e.g. small, medium and large) can be
eld of geotechnical engineering. More recently, Shahin interpreted mathematically in the form of membership
et al. [10] carried out a preliminary investigation regard- functions, and model variables are fuzzied to be partial
ing the feasibility of using B-spline neurofuzzy networks members of these membership functions in the interval
for settlement prediction of shallow foundations on grade (0,1). This means that, for a fuzzy set A, an input
granular soils. variable x is fuzzied to be a partial member of the
The work presented in this paper is an extension of fuzzy set A by transforming it into a degree of mem-
the research conducted by Shahin et al. [10]. The objec- bership of function uA(x) of interval (0,1). There are
tives of the paper are: (i) to assess the feasibility of many forms of membership functions including B-spline
adopting B-spline neurofuzzy networks for predicting and Gaussian functions [7]. Fig. 1 gives an example of
settlement of shallow foundations on granular soils; (ii) B-spline basis functions of dierent order, which are the
to investigate the impact of some of the parameters type of basis functions that will be used in the present
controlling the development of B-spline neurofuzzy research. For each input variable, the fuzzy sets overlap
networks; and (iii) to investigate the impact of incor- and cover the necessary range of variation for that
porating engineering knowledge on model performance variable in the process called fuzzication. It should be
and model interpretation. noted that the model output of a fuzzy set is also fuzzy
and, in order to obtain a real-valued output, defuzzic-
ation is needed. The mean of maxima and centre of grav-
2. B-spline neurofuzzy networks ity are the most popular defuzzication algorithms [7].
A typical structure of a B-spline neurofuzzy network
Neurofuzzy networks are modelling techniques that contains three layers: an input layer; a single hidden
combine the explicit linguistic knowledge representation layer and an output layer [7]. The input layer normalises
of fuzzy systems with the learning power of multi-layer the input space in a p-dimensional lattice (Fig. 2). Each
perceptrons (MLPs) [5,6]. There are many neurofuzzy cell of the lattice represents a similar region of the input
construction algorithms, including [11]: the adaptive space. The hidden layer consists of basis functions (i.e.
spline modelling of observation data (ASMOD) [12]; B-spline functions) which are dened on the lattice
adaptive neurofuzzy inference systems (ANFIS) [13] formed by normalising the input space. The size, shape

Fig. 1. B-Spline fuzzy membership functions of dierent order.


M.A. Shahin et al. / Computers and Geotechnics 30 (2003) 637647 639

One major disadvantage of B-spline networks is that


the number of basis functions (i.e. fuzzy sets or mem-
bership functions) is exponentially dependent on the
dimension of the input space [6]. Consequently, the num-
ber of rules is also exponentially dependent on the
dimension of the input space, resulting in impractical
model representation. This problem has been termed the
curse of dimensionality [6]. To illustrate this problem,
consider a fuzzy logic system that has ve input vari-
ables and that each input variable is presented over ve-
valued membership functions. This fuzzy system will
contain as many as (55=3125) rules. One useful
Fig. 2. Typical structure of a neurofuzzy network [6]. approach for overcoming this problem is to use the
analysis of variance (ANOVA) representation [6].
ANOVA decomposes an n-dimensional function to a
and overlap of the basis functions determine the struc- linear combination of a number of separate functions,
ture and complexity of the network. The output layer as follows [6]:
sums the weighted outputs from the basis functions to X
n n1 X
X n  
produce the network output using Eq. (1). fx f0 fi xi fi;j xi ; xj . . .
X
p i1 i1 ji1
y ai w i 1 f1;2;...;n x 4
i1
where f0 represents a constant (the function bias);
where: y=model output; ai=output from the pth basis and the other terms represent the univariate, bivariate and
function; and wi=connection weight associated with ai. high-order subfunctions. In many situations, the major-
This output is compared with the actual measured ity of high-order terms are zero or negligible, resulting
output and a correction error (the mean squared error, in a limited number of subfunctions (often called sub-
MSE, is usually used) is calculated. Using this error and networks) of much lower dimensions that approximate
implementing a learning rule, the neurofuzzy network the network input/output mapping. It should be noted
adjusts its weights and determines its fuzzy parameters that each subnetwork in the ANOVA description
(i.e. fuzzy sets and fuzzy rules). The Least Mean represents a neurofuzzy system of its own and the over-
Squared (LMS) and the Normalised Least Mean Squared all model output is produced by summing outputs of all
(NLMS) learning algorithms are generally used to subnetworks. An example of ANOVA decomposition
update the weights [7]. At time t and as part of these for the problem of ve input variables and ve mem-
algorithms, Eqs. (2) and (3), respectively, are used to bership functions for each of these is shown in Fig. 3.
adjust the weights for the LMS and NLMS algorithms The 5D function is decomposed into one 1D and two
[6]: 2D subnetworks, resulting in 5, 25 and 25 fuzzy rules for
wi t wi t  1 y^ t  ytai t 2 the rst, second and third subnetwork, respectively.
Consequently, the network with ANOVA decomposi-
tion will produce an overall number of rules equal to 55,
! instead of 3125 for the non-decomposed network.
y^ t  yt
wi t wi t  1    ai t 3 The adaptive spline modelling of observation data
ai t2 (ASMOD) proposed by Kavli [12] is an automatic
2

where: =learning rate; and y^=desired output.


It should be noted that, when the output error is zero,
the weights are not updated, whereas if it is not zero, the
weights are adjusted so as to reduce the output error. If
the basis functions have non-zero outputs in only a
small part of the input space, then only the numbers of
weights that are contributing to the network output are
updated during training. Consequently, similar network
inputs result in similar sets of non-zero basis functions
and therefore, the knowledge is stored locally in the
network without interfering with the knowledge that is
stored in other regions of the network. Fig. 3. ANOVA decomposition of a neurofuzzy rule base [6].
640 M.A. Shahin et al. / Computers and Geotechnics 30 (2003) 637647

algorithm for obtaining the optimal structure of B- the same number of model parameters, the selection is
spline neurofuzzy networks. ASMOD has been found to based only on the prediction errors. It can also be seen
perform well in a wide variety of modelling problems from the above equations that the BIC penalises com-
[7]. The algorithm starts with a simple model (e.g. only plex models more than the AIC and FPE. The eect of
one variable with two membership functions) and itera- using the aforementioned selection criteria on the per-
tively renes the model structure during training so as formance of neurofuzzy networks will be investigated
to gradually increase model capability until some stop- later in this paper.
ping criterion is met. Possible renements include add-
ing or deleting input variables, increasing the number
and dimension of an individual subnetwork by linking it 3. Development of neurofuzzy models
to an existing input, forming multi-variate subnetworks
using ANOVA and changing the number and spacing of It is generally accepted that ve parameters have the
the basis functions (i.e. the optimum partitioning of the most signicant impact on the settlement of shallow
input space). Changing the order of B-spline functions foundations on granular soils [4,17]. These include the
for an individual input variable is also a possible footing width (B), footing net applied pressure (q), soil
renement; however, the order of B-spline functions has compressibility (or density) which can be represented by
to be chosen manually in advance. It should be noted the average blow count (N) obtained from the standard
that higher order B-spline basis functions result in penetration test (SPT) over the depth of inuence of the
smoother model outputs; however, it is likely to lead to foundation, footing geometry
 L=B and footing
data overtting [7]. Consequently, lower order basis embedment ratio Df =B . The data used to calibrate and
functions are more desirable if they are able to model validate B-spline neurofuzzy networks in this research
the desired relationship with a satised accuracy [16]. In were collated from the literature and comprise a total of
the present study, second order (triangular) B-spline 189 individual cases that include eld measurements
functions will be used for all neurofuzzy models devel- of shallow foundations, as well as the corresponding
oped, as it was found by Maier et al. [16] that second information regarding footings and soils. Of the 189
order B-spline functions generally produce more trans- data cases, 125 cases were reported by Burland and
parent models. For every renement, the impact of net- Burbidge [17], 22 cases by Burbidge [18], ve cases by
work pruning is evaluated and the network that has the Bazaraa [19] and 30 cases by Wahls [20]. Another four
simplest structure with the best performance is chosen. cases were given by Briaud and Gibbens [21], one case
As part of ASMOD, stopping/selection criteria have to by Picornell and Del Monte [22] and two cases by
strike a balance between model performance and model Maugeri et al. [23]. The statistical properties of the data
complexity. Examples of such criteria are given by used are summarised in Table 1 and their histograms
Brown and Harris [7], which include: are shown in Fig. 4.
It should be noted that, as recommended by Burland
Bayesian Information Criterion
5 and Burbidge [17], the values of N used in this research
BIC nlnMSE plnn have not been corrected for overburden pressure nor
submergence. However, for very ne and silty sand
below the water table, the submergence correction pro-
Akaikes Information Criterion posed by Terzaghi and Peck [24], when N > 15 is used,
6
AIC nlnMSE p;  > 0 as follows:
Ncorrected 15 0:5N  15 8

Final Prediction Error FPE For gravel or sandy gravel, the correction proposed
  by Burland and Burbidge [17] is used, as follows:
np
nlnMSE nln 7 Ncorrected 1:25N 9
np
It should also be noted that the depth of inuence
where p is the number of parameters in the model, MSE over which N is measured is that proposed by Burland
is the mean square error and n is the number of data and Burbidge [17], as follows. When N is decreasing
pairs used to train the network. For each of the above with depth, the depth of inuence is taken to be equal to
criteria, the model that has the smallest selection criter- the lesser of 2B or the depth from the bottom of the
ion is preferred. It can be seen from the above criteria footing to bedrock. On the other hand, when N is con-
that each of them has two terms. The rst term concerns stant or increasing with depth, the depth of inuence is
the prediction errors computed on the data set used, taken to be equal to B0.75. Circular footings are also
whereas the second term penalises larger models with considered to be equivalent to square footings
the same prediction errors. In case of two models having L=B 1, as it was found by Burbidge [18] that there is
M.A. Shahin et al. / Computers and Geotechnics 30 (2003) 637647 641

Table 1
Statistical properties of the available data

Model variable Minimum value Maximum value Mean Standard deviation

Footing width, B (m) 0.8 60.0 8.8 10.2


Net applied pressure, q (kPa) 18.3 697.0 187.1 123.3
Average SPT blow count, N 4.0 60.0 24.6 13.5
Footing geometry, L=B 1.0 10.5 2.2 1.8
Footing embedment, Df =B 0.0 3.4 0.53 0.58
Measured settlement, Sm (mm) 0.6 121.0 20.4 26.6

Fig. 4. Histograms of the data used for the neurofuzzy models.

no signicant dierence between the settlement of cir- pressure (q), average SPT blow count (N), footing
  geo-
cular or square footings having the same width (B) on metry (L/B) and footing embedment ratio Df =B ] are
the same soil. presented to the neurofuzzy model as potential model
The B-spline neurofuzzy network trained with the input variables. The measured value of settlement (Sm)
adaptive spline modelling of observation data is the single output variable. It should be noted that, as
(ASMOD) algorithm, described in Section 2, is imple- mentioned earlier, the ASMOD algorithm auto-
mented using the personal computer-based commercial matically optimises model architecture and selects the
software Neuframe Version 4.0 [25]. The factors aect- input variables that have the most signicant impact on
ing settlement [i.e. footing width (B), footing net applied model output. It should also be noted that, as mentioned
642 M.A. Shahin et al. / Computers and Geotechnics 30 (2003) 637647

Table 2
Summary of the neurofuzzy models developed

Model No. Stopping No. of signicant Signicant No. of connection No. of fuzzy
criteria inputs inputs weights rules extracted

1 BIC 3 B, q, N 8 10
2 AIC 3 B, q, N 7 8
3 FPE 3 B, q, N 8 16

Table 3
Performance of the neurofuzzy models developed

Model No. Performance measures

Correlation coecient, r RMSE (mm) MAE (mm)

Training Validation Training Validation Training Validation

1 0.889 0.881 12.33 12.36 8.08 9.36


2 0.879 0.863 12.82 13.37 8.29 10.10
3 0.910 0.875 11.16 13.00 6.82 9.49

previously, the ASMOD algorithm uses stopping cri- that the models are trained using dierent stopping
teria [e.g. Bayesian Information Criterion (BIC), criteria (i.e. BIC, AIC and FPE). It can be seen from
Akaikes Information Criterion (AIC) and Final Pre- Table 2 that all models select only three input variables
diction Error (FPE)] that require the data to be divided (i.e. B, q and N) as the most signicant inputs, whereas
into two sets; a training set to build the model and an the footing
  geometry L=B and footing embedment
independent validation set to test the predictive ability ratio Df =B are not selected in any model. This is in
of the model in real-world situations. In total, 80% of agreement with the results of the sensitivity analysis
the data (i.e. 152 case records) are used for training and carried out by Shahin et al. [4] on a multi-layer percep-
20% (i.e. 37) are used for validation. In this research, trons (MLPs) model. The sensitivity analysis by Shahin
the available data are randomly divided into training et al. [4] indicated that B, q and N have the most sig-
and validation sets in such a way that they are repre- nicant impact on settlement, Df =B has a moderate
sentative of the same statistical population [26]. In impact on settlement and L=B has the smallest impact
order to achieve this in the present study, several ran- on settlement. The neurofuzzy models obtained in the
dom combinations of the training and validation sets present work are compared in terms of prediction accu-
are tried until two statistically consistent data sets are racy, model parsimony and model transparency and the
obtained. The statistical parameters considered include optimum model is described in more detail in Section 5.
the mean, standard deviation, maximum, minimum and The following conclusions can be drawn:
range. A detailed discussion regarding data division for
ANN models is beyond the scope of this paper and can  In terms of prediction accuracy, all models are
be found in Shahin et al. [27]. Once training has been comparable, although Model 1 performs slightly
successfully accomplished, the performance of the better than the other models with respect to the
model is tested on the validation data set using three validation set (Table 3).
dierent measures, including the coecient of corre-  In terms of model parsimony, all models are
lation (r); the root mean square error (RMSE) and the comparable, even though Model 2 is found to be
mean absolute error (MAE). slightly more parsimonious than the other mod-
els, as it has the lowest number of connection
weights (Table 2).
4. Results and discussion  In terms of model transparency, Models 1 and 2
are comparable, even though Model 2 is found to
A summary of the structure of the neurofuzzy models be more transparent, as it describes the rela-
developed in this work, and the number of fuzzy rules tionship between model inputs and outputs using
produced for each model, is given in Table 2. The per- a smaller number of fuzzy rules (Table 2). On the
formance of the models obtained is given in Table 3. It other hand, Model 3 is found to be the worst
should be noted that the models obtained are developed model in terms of transparency as it has the
using the same procedures described in Section 3, except highest number of fuzzy rules (Table 2), which is
M.A. Shahin et al. / Computers and Geotechnics 30 (2003) 637647 643

almost twice that obtained for Models 1 and 2. 5. Description of the optimum neurofuzzy model
This is probably because the FPE stopping
criterion does not penalise larger models as much A schematic view of the optimum neurofuzzy model
as the BIC and AIC. (i.e. Model 1) is given in Fig. 5. It can be seen that, as
mentioned previously, the model uses only 3 input vari-
It can be seen from Tables 2 and 3 that Model 1 is ables (i.e. B, q and N) as the most signicant inputs. It
able to strike a balance between model accuracy, parsi- can also be seen that the model has one 1D and one 2D
mony and transparency. In terms of model accuracy, subnetwork. In each of the subnetworks obtained, tri-
the performance of Model 1 is slightly better than the angular membership functions of order 2 are used, as
other models with respect to the validation set (Table 3). mentioned previously, for all input variables, as shown
In terms of model parsimony and transparency, Model in Fig. 6. It can be seen from this gure that the mem-
1 has as the same number of input variables as Models 2 bership functions of B, q and Sm are presented over a
and 3, combined with a number of connection weights two-valued linguistic universe (i.e. small and large for B,
and fuzzy rules that are approximately equal to the light and heavy for q, and low and high for Sm). On the
average of those obtained for Models 2 and 3 (Table 2). other hand, the membership functions of the soil den-
The results in Table 3 also show that the predictive sity, which is represented herein by the average SPT
ability of Model 1 on the validation set is generally blow count, N, is presented over a four-valued linguistic
consistent with that of the training set, indicating that universe (i.e. loose, medium, dense and very dense). As
the model is able to generalise within the range of the a result, the rst subnetwork contains 8 rules while the
data used for training. Overall, Model 1 can be con- second subnetwork contains 2 rules, resulting in a
sidered to be optimal. model with 10 fuzzy rules, as listed in Table 4. It should
be noted that the number between brackets in Table 4
represents the rule condence, which, as mentioned
previously, indicates the condence level that is given to
a certain fuzzy rule.
The fuzzy rules in Table 4 are considered to be a
valuable source of information from which knowledge
about the relationships between settlement and the fac-
tors aecting settlement can be extracted. It should be
noted that the range of applicability of the fuzzy rules in
Table 4 is a function of the quality of the data used
in the model calibration phase. Consequently, it is unli-
kely that these fuzzy rules provide a general representa-
Fig. 5. Schematic representation of the neurofuzzy model. tion of the relationship between settlement and the

Fig. 6. Membership functions of input variables used by the neurofuzzy model.


644 M.A. Shahin et al. / Computers and Geotechnics 30 (2003) 637647

Table 4
Fuzzy rules extracted by the neurofuzzy model

Subnetwork No. Rule No. Rule

1 1 IF Footing width, B is Small AND Soil is Loose


THEN Settlement, Sm is Low (0.84) OR Settlement, Sm is High (0.16)
2 IF Footing width, B is Large AND Soil is Loose
THEN Settlement, Sm is High (1.00)
3 IF Footing width, B is Small AND Soil is Medium density
THEN Settlement, Sm is Low (0.99) OR Settlement, Sm is High (0.01)
4 IF Footing width, B is Large AND Soil is Medium density
THEN Settlement, Sm is Low (0.44) OR Settlement, Sm is High (0.56)
5 IF Footing width, B is Small AND Soil is Dense
THEN Settlement, Sm is Low (0.96) OR Settlement, Sm is High (0.04)
6 IF Footing width, B is Large AND Soil is Dense
THEN Settlement, Sm is Low (0.86) OR Settlement, Sm is High (0.14)
7 IF Footing width, B is Small AND Soil is Very dense
THEN Settlement, Sm is Low (1.00)
8 IF Footing width, B is Large AND Soil is Very dense
THEN Settlement, Sm is Low (0.87) OR Settlement, Sm is High (0.13)
2 9 IF Footing net applied pressure, q is Light
THEN Settlement, Sm is Low (0.96) OR Settlement, Sm is High (0.04)
10 IF Footing net applied pressure, q is Heavy
THEN Settlement, Sm is Low (0.87) OR Settlement, Sm is High (0.13)

footings contained in the database used were designed


factors aecting it. It can be seen from Table 4 that so that the applied load does not result in high settle-
some of the fuzzy rules obtained are in agreement with ment or bearing capacity failure. Another reason is that
what one would expect based on the underlying physical there were insucient training data to cover the full
meaning of the settlement problem. On the other hand, range of possible high settlement conditions. A review
some of the fuzzy rules obtained seem to contradict of the data used indicates that almost 90% of the data
existing physical knowledge. records have settlement that may be considered to be
The knowledge that can be derived from Table 4 that low, as categorised by the settlement membership func-
is in agreement with existing physical knowledge is as tions of the neurofuzzy model. In order to calibrate
follows: models to cater for extreme settlement situations, data
on footings that produce excessive settlements need to
 Small footings are most likely to result in low be included in the model.
settlement regardless of the density of the soil One advantage of neurofuzzy networks is that existing
they are founded on (Rules 1, 3, 5 and 7); engineering knowledge can be incorporated into the
 Large footings are most likely to be susceptible to trained network to optimise model performance and to
high settlement when they are founded on loose enhance the interpretation of a constructed model. In
soils (Rule 2), and they are most likely to result in this work, this is done by optimising the membership
low settlement when they are founded on dense or functions of B and N so as to include available geo-
very dense soils (Rules 6 and 8); and technical engineering knowledge. The membership
 Large footings are equally likely to be susceptible functions of B are optimised to be presented over a
to either low or high settlement when they are three-valued linguistic universe (i.e. small, medium
founded on soils of medium density (Rule 4). and large) so that small footings are limited to footings
of maximum width of 5 m (see Fig. 7). The denition of
The above results indicate that neurofuzzy networks footing width using the three-valued linguistic universe
have the ability to extract rules from data that make shown in Fig. 7 is considered to be more rational than
physical sense, which may be used to gain under- that of the two-valued linguistic universe given by
standing in situations where data are available, but Model 1 based on the available data. This is because it
physical relationships are not well understood. does not make sense, from a geotechnical engineering
It can also be seen from Table 4 that Rules 9 and 10 point of view, that small footings can include widths up
seem to contradict with what one would expect based to 60.0 m, as dened by Model 1 and shown in Fig. 6.
on the underlying physical meaning of the settlement On the other hand, the membership functions of N are
problem. Rules 9 and 10 indicate that settlement is most optimised to be presented over a four-valued linguistic
likely to be low regardless of whether the applied load is universe (i.e. loose, medium, dense and very dense) so
light or heavy. The most likely reason for this is that the that the classication of soil density given by Terzaghi
M.A. Shahin et al. / Computers and Geotechnics 30 (2003) 637647 645

Fig. 7. Optimised membership functions of the neurofuzzy model.

and Peck [24] can be incorporated, as shown in Fig. 7 carrying out this procedure, Model 1 is retrained. The
and given in Table 5. It should be noted that the most performance of the optimised model is given in Table 6
probable values of N that are incorporated in the mem- together with the performance of Model 1. The fuzzy
bership functions for medium and dense soils are taken rules of the optimised neurofuzzy model are listed in
to be equal to the average of the range given by Terza- Table 7. It should be noted that this model again has
ghi and Peck [24]. On the other hand, the most probable three inputs (i.e. B, q and N) and the number of con-
values of N that are incorporated in the membership nection weights is equal to 9.
functions for loose and very dense soils are taken to be It can be seen from Table 6 that the performance of
equal to the minimum and maximum values of N, the optimised neurofuzzy model is slightly better than
respectively, that are found in the database used. By that of the model developed previously (i.e. Model 1),
with respect to the training and validation sets. The
predictive ability of this model is found to be similar to
Table 5
that of the MLP model developed by Shahin et al. [4],
Optimisation of membership functions of N to incorporate the classi-
cation of soil density of Terzaghi and Peck [24] which was found to outperform the most commonly
used traditional methods examined [4]. These results
Terzaghi and Peck [24] Most probable value of N indicate that neurofuzzy networks have the ability to
incorporated in the
membership functions of
predict well the settlement of shallow foundations on
the neurofuzzy model granular soils. It can be seen from Table 7 that the
Soil density Possible model has 14 fuzzy rules. As with Model 1, some of
values of N the fuzzy rules obtained from the optimised model are
Loose <10 4 in agreement with existing physical knowledge of the
Medium 1030 20 settlement problem, while others seem to contradict this
Dense 3050 40 existing knowledge. The knowledge that can be extrac-
Very dense >50 60 ted from the optimised model and make physical sense
is as follows (Table 7):

Table 6
Performance of model 1 and optimised neurofuzzy model

Model No. Performance measure

Correlation coecient., r RMSE (mm) MAE (mm)

Training Validation Training Validation Training Validation

1 0.889 0.881 12.33 12.36 8.08 9.36


Optimised 0.893 0.892 12.11 11.74 7.87 8.93
646 M.A. Shahin et al. / Computers and Geotechnics 30 (2003) 637647

Table 7
Fuzzy rules extracted by the optimised neurofuzzy model

Subnetwork No. Rule No. Rule

1 1 IF Footing width, B is Small AND Soil is Loose


THEN Settlement, Sm is Low (0.86) OR Settlement, Sm is High (0.14)
2 IF Footing width, B is Medium AND Soil is Loose
THEN Settlement, Sm is Low (0.64) OR Settlement, Sm is High (0.36)
3 IF Footing width, B is Large AND Soil is Loose
THEN Settlement, Sm is High (1.00)
4 IF Footing width, B is Small AND Soil is Medium density
THEN Settlement, Sm is Low (0.91) OR Settlement, Sm is High (0.09)
5 IF Footing width, B is Medium AND Soil is Medium density
THEN Settlement, Sm is Low (0.90) OR Settlement, Sm is High (0.10)
6 IF Footing width, B is Large AND Soil is Medium density
THEN Settlement, Sm is Low (0.46) OR Settlement, Sm is High (0.54)
7 IF Footing width, B is Small AND Soil is Dense
THEN Settlement, Sm is Low (0.94) OR Settlement, Sm is High (0.06)
8 IF Footing width, B is Medium AND Soil is Dense
THEN Settlement, Sm is Low (0.89) OR Settlement, Sm is High (0.11)
9 IF Footing width, B is Large AND Soil is Dense
THEN Settlement, Sm is Low (0.84) OR Settlement, Sm is High (0.16)
10 IF Footing width, B is Small AND Soil is Very Dense
THEN Settlement, Sm is Low (1.00)
11 IF Footing width, B is Medium AND Soil is Very Dense
THEN Settlement, Sm is Low (0.90) OR Settlement, Sm is High (0.10)
12 IF Footing width, B is Large AND Soil is Very Dense
THEN Settlement, Sm is Low (0.81) OR Settlement, Sm is High (0.19)
2 13 IF Footing net applied pressure, q is Light
THEN Settlement, Sm is Low (0.89) OR Settlement, Sm is High (0.11)
14 IF Footing net applied pressure, q is Heavy
THEN Settlement, Sm is Low (0.78) OR Settlement, Sm is High (0.22)

 Small footings are most likely to result in low constructed models. Also, Rules 13 and 14 again indi-
settlement regardless of the density of the soil cate that settlement is most likely to be low regardless of
they are founded on (Rules 1, 4, 7 and 10); whether the applied load is light or heavy, which con-
 Medium size footings are approximately equally tradicts what one would expect. As mentioned pre-
likely to be susceptible to either low or high set- viously, the most likely reason for this is that the
tlement when they are founded on loose soils footings contained in the data set were designed so that
(Rule 2), and they are most likely to result in low the applied load does not result in high settlement and
settlement when they are founded on medium, that there were insucient training data to cover the full
dense or very dense soils (Rules 5, 8 and 11); range of possible high settlement conditions.
 Large footings are most likely to be susceptible to
high settlement when they are founded on loose
soils (Rule 3), and they are most likely to result in 6. Summary and conclusions
low settlement when they are founded on dense or
very dense soils (Rules 9 and 12); and B-spline neurofuzzy networks trained with the
 Large footings are equally likely to be susceptible ASMOD algorithm were used to assess the feasibility of
to either low or high settlement when they are neurofuzzy models for predicting the settlement of
founded on soils of medium density (Rule 6). shallow foundations on granular soils and to investigate
the impact of incorporating engineering knowledge on
It should be noted that, in general, the knowledge neurofuzzy model performance and interpretation. The
acquired from the above fuzzy rules is in agreement with sensitivity of the B-spline neurofuzzy models to a num-
the knowledge obtained previously from model 1 ber of stopping criteria [i.e. Bayesian Information Cri-
(Table 4). However, the fuzzy rules in Table 7 describe terion (BIC), Akaikes Information Criterion (AIC) and
the relationship between settlement and the factors Final Prediction Error (FPE)], was also investigated.
aecting settlement in a more rational fashion. The The models were compared in terms of prediction accu-
above results suggest that it is benecial to add available racy, model parsimony and model transparency. Five
expertise to neurofuzzy models, as it can improve model potential input variables (i.e. B, q, N, L=B and Df =B)
performance and enhance the interpretation of the were presented to the neurofuzzy models and settlement
M.A. Shahin et al. / Computers and Geotechnics 30 (2003) 637647 647

was the single output. This paper has yielded the fol- [4] Shahin MA, Maier HR, Jaksa MB. Predicting settlement of
lowing results and conclusions: shallow foundations using neural networks. J Geotech Geoenv
Eng 2002;128(9):78593.
[5] Altrock CV. Fuzzy logic and neurofuzzy applications explained.
1. Neurofuzzy models have the ability to predict Englewood Clis (NJ): Prentice-Hall; 1995.
well the settlement of shallow foundations on [6] Brown M, Harris CJ. A perspective and critique of adaptive
granular soils and are able to extract rules from neurofuzzy systems used for modelling and control applications.
the data that make physical sense, which may be Int J Neural Systems 1995;6(2):197220.
used to gain a better understanding in situations [7] Brown M, Harris CJ. Neurofuzzy adaptive modelling and con-
trol. Englewood Clis (NJ): Prentice-Hall; 1994.
where data are available but physical relation- [8] Sayed T, Razavi A. Comparison of neural and conventional
ships are not well understood. In addition, neu- approaches to mode choice analysis. J Computing Civil Eng
rofuzzy networks can be modied by 2000;14(1):2330.
incorporating available engineering knowledge to [9] Ni SH, Lu PC, Juang CH. A fuzzy neural network approach to
improve model performance and enhance the evaluation of slope failure potential. J Microcomputers Civil Eng
1996;11:5966.
interpretation of the constructed model. [10] Shahin MA, Jaksa MB, Maier HR. Neurofuzzy networks applied
2. The footing width (B), footing net applied pres- to settlement of shallow foundations on granular soils. In: 9th int.
sure (q) and average SPT blow count (N) were conf. on applications of statistics and probability in civil eng.,
found to be the most signicant factors aecting ICASP9, San Francisco. Rotterdam: Millpress; 2003. p. 137983.
settlement. This is in agreement with the results [11] Harris C, Brown M, Gunn S. Neurofuzzy systems modelling: a
transparent approach. In: Karny M, Warwick K, Kurkova V,
of the sensitivity analysis carried out by Shahin et editors. Dealing with complexity: a neural networks approach.
al. [4] on a multi-layer perceptron model. Springer-Verlag; 1998. p. 11025.
3. All neurofuzzy models were found to be com- [12] Kavli T. ASMODan algorithm for adaptive spline modelling of
parable in terms of prediction accuracy, even observation data. Int J Control 1993;58(4):94767.
though the model that uses the BIC stopping [13] Jang J. ANFIS: adaptive-network-based fuzzy inference system.
IEEE Trans Sys, Man and Cybernetics 1993;23:66585.
criterion was found to perform marginally better [14] Nelles O, Sinsel S, Isermann R. Local basis function networks for
than the other models on an independent valid- identication of a turbocharger. IEE UKACC Control96, UK
ation set. 1996. p. 712.
4. All neurofuzzy models were found to be com- [15] Zadeh LA. Fuzzy sets. Information and Control 1965;8:33853.
[16] Maier HR, Sayed T, Lence BJ. Forecasting cyanobacterium
parable in terms of model parsimony, even
Anabaena spp. in River Murray, South Australia, using B-spline
though the model that uses the AIC stopping neurofuzzy models. Ecological Modelling 2001;146:8596.
criterion was found to be slightly more parsi- [17] Burland JB, Burbidge MC. Settlement of foundations on sand
monious than the other models, as it has the and gravel. Proc Institution of Civil Engineers, London 1985:
lowest number of connection weights. 132581.
5. The neurofuzzy models that use the BIC and AIC [18] Burbidge MC. A case study review of settlements on granular
soil. MSc thesis, Imperial College of Science and Technology,
stopping criteria were found to be more trans- University of London, London; 1982.
parent than the model that uses the FPE stopping [19] Bazaraa ARSS. Use of the standard penetration test for estimat-
criterion, as they have fewer fuzzy rules. This was ing settlements of shallow foundations on sand. PhD thesis,
attributed to the fact that the BIC and AIC University of Illinois, Urbana, Illinois; 1967.
penalise complex models more. [20] Wahls HE. Settlement analysis for shallow foundations on sand.
In: Proc. 3rd int. geotech. eng. conf., Cairo, Egypt; 1997. p. 728.
6. Neurofuzzy model that uses the BIC stopping [21] Briaud JL, Gibbens R. Behaviour of ve large spread footings in
criterion was found to be able to strike a balance sand. J Geotech Geoenv Eng 1999;125(9):78796.
between model accuracy, parsimony and trans- [22] Picornell M, Del Monte E. Prediction of settlement of cohesive
parency and consequently, was considered to be granular soils. In: Proc. measured performance of shallow found,
Nashville, Tennessee; 1988. p. 5572.
optimal.
[23] Maugeri M, Castelli F, Massimino MR, Verona G. Observed and
computed settlements of two shallow foundations on sand. J
Geotech Geoenv Eng 1998;124(7):595605.
References
[24] Terzaghi K, Peck RB. Soil mechanics in engineering practice.
[1] Poulos HG. Common procedures for foundation settlement ana- New York: John Wiley; 1948.
lysisare they adequate? In: Proc. 8th Australia New Zealand [25] Neusciences. Neuframe Version 4.0. Southampton, Hampshire:
Conf. on Geomechanics, Hobart, 1999. p. 325. Neusciences Corp., 2000.
[2] Fausett LV. Fundamentals neural networks: architecture, algo- [26] Masters T. Practical neural network recipes in C++. San Diego
rithms, and applications. Englewood Clis (NJ): Prentice-Hall; (CA): Academic Press; 1993.
1994. [27] Shahin MA, Maier HR, Jaksa MB. Data division for developing
[3] Zurada JM. Introduction to articial neural systems. St. Paul: neural networks applied to geotechnical engineering. J Comput-
West Publishing Company; 1992. ing Civil Eng ASCE, [in press].

Vous aimerez peut-être aussi