You are on page 1of 8

Available online at www.sciencedirect.

com

Expert Systems
with Applications
Expert Systems with Applications 34 (2008) 3099–3106
www.elsevier.com/locate/eswa

An adaptive neuro-fuzzy inference system for bridge risk assessment


a,* b
Ying-Ming Wang , Taha M.S. Elhag
a
Institute of Soft Science, Fuzhou University, Fuzhou 350002, PR China
b
School of Mechanical, Aerospace and Civil Engineering, The University of Manchester, P.O. Box 88, Manchester M60 1QD, UK

Abstract

Bridge risks are often evaluated periodically so that the bridges with high risks can be maintained timely. This paper develops an
adaptive neuro-fuzzy system (ANFIS) using 506 bridge maintenance projects for bridge risk assessment, which can help Highways
Agency to determine the maintenance priority ranking of bridge structures more systematically, more efficiently and more economically
in comparison with the existing bridge risk assessment methodologies which require a large number of subjective judgments from bridge
experts to build the complicated nonlinear relationships between bridge risk score and risk ratings. The ANFIS proves to be very effective
in modelling bridge risks and performs better than artificial neural networks (ANN) and multiple regression analysis (MRA).
Ó 2007 Elsevier Ltd. All rights reserved.

Keywords: Adaptive neuro-fuzzy inference system; Bridge risk assessment; Artificial neural networks

1. Introduction However, the current aggregation process from risk rat-


ings to risk score employed by the Highways Agency
According to the British Highways Agency (2004), requires a large number of subjective judgments from
bridge risk refers to any event or hazard that could hinder bridge experts such as ‘‘what’’ a maintenance project can
the achievement of business goals or the delivery of stake- deliver, ‘‘how’’ and ‘‘when’’ they can be delivered, and so
holder expectations and is defined as the product of the on. Such an aggregation process proves to be highly subjec-
likelihood and consequences of an event occurring. Risk tive, costly and time-consuming. If the complicated nonlin-
events associated with bridge maintenance activities ear relationships between the risk score and risk ratings can
include deterioration, failure to meet the Agency’s obliga- be learned and memorized by some mathematical models,
tions for freedom of movement on the network and failure then the relationships can be used to predict the risk scores
of a component, element or structure. of bridge structures in the future. These models will obvi-
Bridge risks have to be assessed periodically so that ously be of great benefit to the Highways Agency. With
highly risky bridges can be maintained timely to make sure these developed models, the Highways Agency can manage
of the safety of the public. Usually, bridge risks can be and control bridge risks more easily than before and needs
assessed against different criteria such as safety, functional- no longer to spend time and costs in judging what benefits
ity, sustainability and environment and characterized by and disbenefits each bridge maintenance project can deliver
risk ratings such as high, medium, low or none. The risk rat- because the complicated nonlinear relationships between
ings on different criteria can then be aggregated into an risk score and risk ratings have been learned and memorized
overall risk score, based on which the maintenance priority by the models from the past experiences and examples.
of the bridge structure under evaluation can be determined. To learn the relationships, artificial neural networks
Big risk score means high maintenance priority. (ANN), multiple regression analysis (MRA) and the evi-
dential reasoning (ER) approach have all been tested using
*
Corresponding author. 506 bridge maintenance projects. Elhag and Wang (in
E-mail address: msymwang@hotmail.com (Y.-M. Wang). press) reveal that ANN has very strong capability of

0957-4174/$ - see front matter Ó 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2007.06.026
3100 Y.-M. Wang, T.M.S. Elhag / Expert Systems with Applications 34 (2008) 3099–3106

modelling bridge risks and the average fitting and testing For a first-order Sugeno fuzzy model (Sugeno, 1985), a
accuracies for bridge risk scores by ANN are both over typical rule set with four fuzzy if–then rules can be
96%. They also find that ANN outperforms MRA, but a expressed as
hybrid model, which combines ANN and MRA models,
produces better accuracies than any of the individually Rule 1: if x is A1 and y is B1, then f11 = p11x +
developed models. Wang and Elhag (2007) conduct a q11y + r11,
detailed comparison among ANN, MRA and a recently Rule 2: if x is A1 and y is B2, then f12 = p12x +
developed evidential reasoning (ER) approach. The com- q12y + r12,
parison covers not only the performances of the three alter- Rule 3: if x is A2 and y is B1, then f21 = p21x +
native methodologies in modelling bridge risks, but also q21y + r21,
their modelling mechanisms, merits and demerits. How- Rule 4: if x is A2 and y is B2, then f22 = p22x +
ever, adaptive neuro-fuzzy inference system (ANFIS), q22y + r22,
which has been widely used for different purposes such as
prediction (Bateni & Jeng, 2007; Bateni, Borghei, & Jeng, where A1, A2, B1 and B2 are the MFs for the inputs x and y,
2007; Chang & Chang, 2006; Firat & Güngör, 2007; Mous- respectively, pij, qij and rij (i, j = 1, 2) are consequent param-
avi, Ponnambalam, & Karray, 2007; Zaheeruddin & Gar- eters (Jang, 1993).
ima, 2006), knowledge discovery (Huang, Tsou, & Lee, As can be seen from Fig. 1, the architecture of a typical
2006), medical decision making and disease diagnosis ANFIS consists of five layers, which perform different
(Huang, Chen, & Huang, 2007; Polat & Günesß, 2006; Polat actions in the ANFIS and are detailed below.
et al., in press; Tokmakçi, Ünalan, Soyuer, & Öztürk, in
press), has not been tested yet. In this paper, we test Layer 1: All the nodes in this layer are adaptive nodes.
ANFIS and compare its performances with those of They generate membership grades of the inputs. The
ANN in modelling bridge risks. This will provide more outputs of this layer are given by
methodological comparisons for bridge risk assessment.
O1Ai ¼ lAi ðxÞ; i ¼ 1; 2;
The paper is organized as follows. Section 2 provides a
brief description of ANFIS. Section 3 develops an ANFIS O1Bj ¼ lBj ðyÞ; j ¼ 1; 2;
using 506 bridge maintenance projects and compares its
where x and y are crisp inputs, and Ai and Bj are fuzzy
performances with those of ANN. Conclusions are pre-
sets such as low, medium, high characterized by appro-
sented in Section 4.
priate MFs, which could be triangular, trapezoidal,
Gaussian functions or other shapes. In this study, the
2. Adaptive neuro-fuzzy inference system (ANFIS) generalized bell-shaped MFs defined below are utilized
1
ANFIS is a multilayer feed-forward network which uses lAi ðxÞ ¼  2bi ; i ¼ 1; 2;
xci
neural network learning algorithms and fuzzy reasoning to 1þ ai
map inputs into an output. It is a fuzzy inference system
1
(FIS) implemented in the framework of adaptive neural lBj ðyÞ ¼  2bj ; j ¼ 1; 2;
ycj
networks. Fig. 1 shows the architecture of a typical ANFIS 1þ aj
with two inputs, four rules and one output for the first-
order Sugeno fuzzy model, where each input is assumed where {ai, bi, ci} and {aj, bj, cj} are the parameters of the
to have two associated membership functions (MFs). MFs, governing the bell-shaped functions. Parameters

Layer 1 Layer 2 Layer 3 Layer 4 Layer 5


x y

W11 W11 W11 f11


A1 ∏ N
W12
x x y

W12 W12 f12


A2 ∏ N
x y ∑ z
W21 W21 f 21
B1 ∏ N
W21 x y
y
W22 W22 W22 f 22
B2 ∏ N

Fig. 1. ANFIS structure for a two-input Sugeno model with four rules.
Y.-M. Wang, T.M.S. Elhag / Expert Systems with Applications 34 (2008) 3099–3106 3101

in this layer are referred to as premise parameters or ground of the hybrid learning algorithm can be found in
antecedent parameters. Jang (1993).
Layer 2: The nodes in this layer are fixed nodes labelled
Õ, indicating that they perform as a simple multiplier. 3. Modelling bridge risks using ANFIS
The outputs of this layer are represented as
This section presents the development of an ANFIS for
O2ij ¼ W ij ¼ lAi ðxÞlBj ðyÞ; i; j ¼ 1; 2;
bridge risk assessment and tests its performance. Compar-
which represents the firing strength of each rule. The fir- isons with ANN are also made.
ing strength means the degree to which the antecedent
part of the rule is satisfied. 3.1. Data description
Layer 3: The nodes in this layer are also fixed nodes
labelled N, indicating that they play a normalization role The dataset used for developing an ANFIS was pro-
in the network. The outputs of this layer can be repre- vided by the British Highways Agency. The dataset con-
sented as tains 506 bridge maintenance projects and is randomly
split into two sample sets: training dataset with 390 projects
W ij and testing dataset with 116 projects. Both the training and
O3ij ¼ W ij ¼ ; i; j ¼ 1; 2;
W 11 þ W 12 þ W 21 þ W 22 testing datasets cover all levels and types of bridge risks.
which are called normalized firing strengths. Inputs to the ANFIS are the safety risk rating (SRR),
Layer 4: Each node in this layer is an adaptive node, functionality risk rating (FRR), sustainability risk rating
whose output is simply the product of the normalized (SURR), and environment risk rating (ERR) of the 506
firing strength and a first-order polynomial (for a first- bridge maintenance projects, which all range from 0 to 3
order Sugeno model). Thus, the outputs of this layer with 0 representing no risk, 1 low risk, 2 medium risk
are given by and 3 high risk. Output to the ANFIS is the risk scores
(RSs) of the 506 bridge projects, which range from 5 to
O4ij ¼ W ij fij ¼ W ij ðpij x þ qij y þ rij Þ; i; j ¼ 1; 2: 99, as shown in Fig. 2.
Parameters in this layer are referred to as consequent
3.2. Development of the ANFIS
parameters.
Layer 5:PThe single node in this layer is a fixed node
With the 390 training dataset, we choose two general-
labelled , which computes the overall output as the
ized bell-shaped MFs for each of the four inputs to build
summation of all incoming signals, i.e.
the ANFIS, which leads to 16 if–then rules containing
2 X
X 2 2 X
X 2
104 parameters to be learned. Note that it is inappropriate
z ¼ O51 ¼ W ij fij ¼ W ij ðpij x þ qij y þ rij Þ to choose three or more MFs for each input because the
i¼1 j¼1 i¼1 j¼1
parameters needing to be learned in that case will be
2 X
X 2
greater than the number of training samples. Fig. 3 shows
¼ ½ðW ij xÞpij þ ðW ij yÞqij þ ðW ij Þrij ;
the model structure of the ANFIS that is to be built for
i¼1 j¼1
bridge risk assessment in this study. The model structure
which is a linear combination of the consequent param- is implemented using the fuzzy logic toolbox of MATLAB
eters when the values of the premise parameters are fixed. software package.

It can be observed that the ANFIS architecture has two 3.3. Performance analysis
adaptive layers: Layers 1 and 4. Layer 1 has modifiable
parameters {ai, bi, ci} and {aj, bj, cj} related to the input Figs. 4 and 5 show the initial and final MFs before and
MFs. Layer 4 has modifiable parameters {pij, qij, rij} per- after 500 epochs of training (Epoch is set as 500 in this
taining to the first-order polynomial. The task of the learn- study), from which it can be seen that significant modifica-
ing algorithm for this ANFIS architecture is to tune all the tions have been done to the shapes of initial MFs through
modifiable parameters to make the ANFIS output match the learning process. The trained if–then rules are pre-
the training data. Learning or adjusting these modifiable sented in Fig. 6, which can be used for prediction. For
parameters is a two-step process, which is known as the example, if we change the values of the four inputs from
hybrid learning algorithm. In the forward pass of the 1.5 to 3, then we immediately get the new output value
hybrid learning algorithm, the premise parameters are hold of the ANFIS as 99. This is illustrated in Fig. 7.
fixed, node outputs go forward until Layer 4 and the con- The trained ANFIS is validated by the testing dataset.
sequent parameters are identified by the least squares Fig. 8 shows the testing errors for the testing dataset. For
method. In the backward pass, the consequent parameters convenience, the fitting errors for the training dataset are
are held fixed, the error signals propagate backward and also shown in Fig. 8, from which it can be observed that
the premise parameters are updated by the gradient descent except for three bridge structures (BSs), the fitting and test-
method. The detailed algorithm and mathematical back- ing errors for all the other 503 BSs are all nearly zero. The
3102 Y.-M. Wang, T.M.S. Elhag / Expert Systems with Applications 34 (2008) 3099–3106

100
80

Risk score
60
40
20
0
0 30 60 90 120 150 180 210 240 270 300 330 360 390 420 450 480 510
Bridge maintenance project

Fig. 2. Bridge risk scores of 506 bridge maintenance projects.

Fig. 3. Model structure of the ANFIS for bridge risk assessment.

Fig. 4. The membership functions before training. Fig. 5. The membership functions after training.

three exceptional BSs are BS178, BS407 and BS446, whose 7.27% and 7.27%, respectively. It should not be expected
risk scores are 56, 97 and 97, respectively, but the fitted or that the ANFIS produce very good results for all the train-
predicted values for them by the ANFIS are 44, 104 and ing and testing samples particularly in the case that there
104. The relative errors for these three BSs are 20.99%, might be conflicting data in the training or testing dataset.
Y.-M. Wang, T.M.S. Elhag / Expert Systems with Applications 34 (2008) 3099–3106 3103

Fig. 6. If–then rules after training.

Fig. 7. If–then rules for prediction by changing the values of inputs.

Looking into the training dataset, we find that both BS170 but the same risk score of 56. These two samples also con-
and BS178 have the same risk ratings: 1, 1, 1, 0 for SRR, flict with each other. It can be concluded that the risk score
FRR, SURR, ERR, respectively, but different risk scores: of BS178 is very likely to be an outlier. The performance of
44 and 56. These two samples are obviously in conflict with the developed ANFIS is in fact very good if the fitting error
each other. Moreover, it is also found that BS177 and of BS178 is not included. This can be seen clearly from
BS178 have different risk ratings: (2, 1, 1, 1) and (1, 1, 1, 0), Fig. 8.
3104 Y.-M. Wang, T.M.S. Elhag / Expert Systems with Applications 34 (2008) 3099–3106

10
5

Error
0
-5
-10
0 30 60 90 120 150 180 210 240 270 300 330 360 390 420 450 480 510
Bridge maintenance project

Fig. 8. Fitting and testing errors of the 506 bridge maintenance projects by ANFIS.

3.4. Comparisons with ANN Table 1


Performances of ANFIS and ANN in modelling bridge risk scores
To compare the performances of the ANFIS and ANN, Model Training dataset Testing dataset
the following evaluation criteria are adopted. RMSE MAPE (%) R RMSE MAPE (%) R
Root mean squared error (RMSE): ANFIS 0.63 0.32 0.9995 0.99 0.68 0.9990
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ANN 1.54 3.24 0.9981 1.48 3.54 0.9982
u
u1 X N
RMSE ¼ t
2
ðAt  F t Þ ;
N t¼1
shows the fitting and testing errors for the 506 bridge main-
where At and Ft are actual (desired) and fitted (or pre- tenance projects obtained by the ANN. It is very clear from
dicted) values, respectively, and N is the number of training Table 1 and Figs. 8 and 10 that the ANFIS has smaller
or testing samples. RMSE and MAPE as well as bigger R for both the training
Mean absolute percentage error (MAPE): and testing datasets than the ANN model. In other words,
N   the ANFIS achieves better performances than the ANN
1 X  At  F t  model. Therefore, ANFIS is a good choice for modelling
MAPE ¼  100:
N t¼1  At  bridge risks. Moreover, ANN is a black box in nature
and its relationships between inputs and outputs are not
Correlation coefficient (R): easy to be interpreted, while ANFIS is transparent and
PN
t¼1 ðAt  AÞðF t  F Þ
its if–then rules are very easy to understand and interpret.
R ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn ffi; But the drawback of ANFIS is its limitation to the number
2 Pn 2
t¼1 ðAt  AÞ  t¼1 ðF t  F Þ of outputs. It can only model a single output.
PN PN
where A ¼ N1 t¼1 At and F ¼ N1 t¼1 F t are the average val-
ues of At and Ft over the training or testing dataset. The 3.5. Choice of MFs
smaller RMSE and MAPE and larger R mean better
performance. The ANFIS has been built so far using the generalized
According to Elhag and Wang (in press), the best ANN bell-shaped MFs and proves to perform well, but how
structure for the 506 bridge maintenance projects is a three- about using other-shaped MFs? In this section, we test
layer back propagation network with 10 hidden neurons, as three more MFs, which are triangular, trapezoidal and
shown in Fig. 9. The performances of the ANFIS and Gauss MFs. They are respectively defined as follows:
8
ANN in modelling bridge risks are presented in Table 1,
< ðx  aÞ=ðb  aÞ; a 6 x 6 b
>
where the two models are trained using the same training lA ðxÞ ¼ ðc  xÞ=ðc  bÞ; b 6 x 6 c ;
dataset and validated by the same testing dataset. Fig. 10 >
:
0; otherwise
8 xa
> ; a 6 x 6 b
>
> ba
< 1; b6x6c
Safety lA ðxÞ ¼ dx ;
>
> ; c6x6d
>
: dc
Functionality 0; otherwise
1
lA ðxÞ ¼ 2
:
Sustainability Risk Score 1 þ ððx  cÞ=aÞ
For these three different MFs, Figs. 11–13 show the fitting
Environment
and testing errors of the 506 bridge maintenance projects
by the ANFIS trained with the same dataset but different
MFs. It is easy to find that the generalized bell-shaped
Fig. 9. ANN architecture for bridge risk assessment. MFs are the best choice for modelling bridge risks because
Y.-M. Wang, T.M.S. Elhag / Expert Systems with Applications 34 (2008) 3099–3106 3105

10
5

Error
0
-5
-10
0 30 60 90 120 150 180 210 240 270 300 330 360 390 420 450 480 510
Bridge maintenance project

Fig. 10. Fitting and testing errors of the 506 bridge maintenance projects by ANN.

10
5
Error

0
-5
-10
0 30 60 90 120 150 180 210 240 270 300 330 360 390 420 450 480 510
Bridge maintenance project

Fig. 11. Fitting and testing errors by the ANFIS with triangular membership functions.

10

0
Error

-10

-20
0 30 60 90 120 150 180 210 240 270 300 330 360 390 420 450 480 510
Bridge maintenance project

Fig. 12. Fitting and testing errors by the ANFIS with trapezoidal membership functions.

10
5
Error

0
-5
-10
-15
0 30 60 90 120 150 180 210 240 270 300 330 360 390 420 450 480 510
Bridge maintenance project

Fig. 13. Fitting and testing errors by the ANFIS with Gauss membership functions.

they lead to minimum fitting and testing errors when com- we have developed an adaptive neuro-fuzzy inference sys-
pared with Figs. 11–13. tem for bridge risk assessment. The developed ANFIS
learns the if–then rules between bridge risk scores and risk
4. Conclusions ratings from the past bridge maintenance projects and
memorizes them for generalization and prediction. It has
Modelling bridge risks is a challenging job facing High- been observed that ANFIS outperforms artificial neural
ways Agencies because good mathematical models can save networks, which have been found in Elhag and Wang (in
them a significant amount of cost and time. In this paper press), Wang and Elhag (2007) to perform better than
3106 Y.-M. Wang, T.M.S. Elhag / Expert Systems with Applications 34 (2008) 3099–3106

multiple regression models. Differing from ANN, ANFIS is Huang, M. L., Chen, H. Y., & Huang, J. J. (2007). Glaucoma detection
transparent rather than a black box. Its if–then rules are using adaptive neuro-fuzzy inference system. Expert Systems with
Applications, 32, 458–468.
easy to understand and interpret. In summary, ANFIS is Huang, M. J., Tsou, Y. L., & Lee, S. C. (2006). Integrating fuzzy data
a good choice and powerful tool for modelling bridge risks. mining and fuzzy artificial neural networks for discovering implicit
knowledge. Knowledge-Based Systems, 19, 396–403.
Jang, J. S. R. (1993). ANFIS: Adaptive-network-based fuzzy inference
Acknowledgement systems. IEEE Transactions on Systems Man and Cybernetics, 23,
665–685.
This research was supported by the UK Engineering and Mousavi, S. J., Ponnambalam, K., & Karray, F. (2007). Inferring
Physical Sciences Research Council (EPSRC) under the operating rules for reservoir operations using fuzzy regression and
ANFIS. Fuzzy Sets and Systems, 158, 1064–1082.
Project No. GR/S66770/01. Polat, K., & Günesß, S. An expert system approach based on principal
component analysis and adaptive neuro-fuzzy inference system to
diagnosis of diabetes disease. Digital Signal Processing.
References Polat, K., & Günesß, S. (2006). A hybrid medical decision making system
based on principles component analysis, k-NN based weighted pre-
Bateni, S. M., & Jeng, D. S. (2007). Estimation of pile group scour using processing and adaptive neuro-fuzzy inference system. Digital Signal
adaptive neuro-fuzzy approach. Ocean Engineering, 34, 1344–1354. Processing, 16, 913–921.
Bateni, S. M., Borghei, S. M., & Jeng, D.-S. (2007). Neural network and Sugeno, M. (1985). Industrial applications of fuzzy control. Amsterdam:
neuro-fuzzy assessments for scour depth around bridge piers. Engi- Elsevier.
neering Applications of Artificial Intelligence, 20, 401–414. Tokmakçi, M., Ünalan, D., Soyuer, F., & Öztürk, A. (in press). The
British Highways Agency (2004). Value Management of the Structures reevaluate statistical results of quality of life in patients with
Renewal Programme, Version 2.2. cerebrovascular disease using adaptive network-based fuzzy inference
Chang, F. J., & Chang, Y. T. (2006). Adaptive neuro-fuzzy inference system. Expert Systems with Applications, doi:10.1016/
system for prediction of water level in reservoir. Advances in Water j.eswa.2006.10.026.
Resources, 29, 1–10. Wang, Y. M., & Elhag, T. M. S. (2007). A comparison of neural network,
Elhag, T. M. S., & Wang, Y. M. (in press). Bridge risk assessment using evidential reasoning and multiple regression analysis in modelling
neural networks. Journal of Computing in Civil Engineering. bridge risks. Expert Systems with Applications, 32, 336–348.
Firat, M., & Güngör, M. (2007). River flow estimation using adaptive Zaheeruddin & Garima (2006). A neuro-fuzzy approach for prediction of
neuro fuzzy inference system. Mathematics and Computers in Simula- human work efficiency in noisy environment. Applied Soft Computing,
tion, 75, 87–96. 6, 283–294.

You might also like