You are on page 1of 13

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/281935762

Comparison of Seven Artificial Intelligence


Methods for Damage Detection of Structures

Conference Paper · September 2015


DOI: 10.4203/ccp.108.116

CITATIONS READS

0 405

3 authors, including:

Ramin Ghiasi Mohammad Noori


University of Sistan and Baluchestan California Polytechnic State University, San…
14 PUBLICATIONS 6 CITATIONS 132 PUBLICATIONS 1,457 CITATIONS

SEE PROFILE SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Ramin Ghiasi
letting you access and read them immediately. Retrieved on: 30 August 2016
Civil-Comp Press, 2015
Proceedings of the Fifteenth International Conference on
Paper 116 Civil, Structural and Environmental Engineering Computing,
J. Kruis, Y. Tsompanakis and B.H.V. Topping, (Editors),
Civil-Comp Press, Stirlingshire, Scotland

Comparison of Seven Artificial Intelligence Methods for


Damage Detection of Structures
R. Ghiasi1, M.R. Ghasemi1 and M. Noori2,3
1
Department of Civil Engineering, Faculty of Engineering
University of Sistan and Baluchestan, Zahedan, Iran
2
Mechanical Engineering Department, California Polytechnic State University
San Luis Obispo, United States of America
3
International Institute for Urban Systems Engineering
Southeast University, Nanjing, China
Abstract
Over the past two decades, significant research has been carried out in the area of
damage detection of structural systems and the field of structural health monitoring
(SHM) has become a major research field. The research in SHM has been mainly in
two distinct areas: a) development of sensing technologies and hardware for
identifying the location and the severity of damage, and b) development of
diagnostics and computational tools for the analysis and interpretation of the
structural response data in order to identify the location and the time of occurrence
of the damage. Despite extensive work carried out in the area of diagnostics, a
comprehensive and accurate methodology that can be used in conjunction with the
hardware and to identify the time, location and the extent of the damage, with a high
degree of reliability and especially taking into account the inherent uncertainties is
still lacking. In this paper, in order to address some of the current shortcomings in
this area, structural damage detection is performed incorporating several methods
including artificial intelligence (AI) including back-propagation neural networks,
least squares support vector machines (LS-SVMs), adaptive neural-fuzzy inference
system, radial basis function neural network, large margin nearest neighbour,
extreme learning machine (ELM), and Gaussian process. The comparative results
are presented. By considering the dynamic behaviour of a structure as input
variables, seven AI methods are constructed, trained and tested to detect the location
and severity of damage in civil structures. The variation of running time, mean
square error, number of training and testing data, and other indices for measuring the
accuracy in the prediction are defined and calculated in order to inspect advantages
as well as the shortcomings of each algorithm. The results indicate that the ELM and
LS-SVM methods demonstrate better performance in predicting the location and
severity of damage than other methods.

Keywords: damage detection, artificial intelligence, neural networks, support vector


machine, Gaussian process, extreme learning machine.

1
1 Introduction
Many structures may experience some local damages during functional age that
reduce their reliability and durability. These damages may even cause catastrophic,
economic and human life losses. How to detect the damage as soon as it appears in
the structure has attracted significant attention of engineers and researchers in recent
years. The process of utilizing a damage detection strategy for engineering structures
is defined as Structural Health Monitoring (SHM) [1]. SHM has great significance
in many engineering applications, such as enhancing the safety of the structures,
forecasting the failure of the structures, reducing the cost of the structural
maintenance and improving productive efficiency.
The structural damages are usually detected by the dynamic characteristics of the
structure, such as frequencies, mode shapes and frequency-domain transfer function
is an important method for structural damage detection.
Considering a change in the dynamic characteristics, as a measure for detecting
the damage, and the feasibility of applying artificial intelligence (AI) was considered
in order to detect the structural damages [2]. Ghiasi et al [3] presented a structural
damage detection method based on combining the wavelet packet decomposition
and least square support vector machine. Furthermore, they proposed new
combinational kernel functions for LS-SVM which combines Thin Plate Spline
Radial Basis Function (RBF) kernel with local characteristics and a modified
Littlewood-Paley Wavelet kernel function with global characteristics [4]. A study
was performed by Saeed et al. [5] to identify the cracks in curvilinear beams by
using ANN and adaptive neural fuzzy inference system (ANFIS). The results
demonstrated that the average prediction errors in the multiple ANN models is less
than those in the single ANN model and multiple ANFIS. Saadat et al. [6] proposed
Intelligent Parameter varying (IPV) for damage detection in a highly nonlinear
hysteretic system. It uses embedded radial basis function networks to estimate the
constitutive characteristics of inelastic and hysteretic restoring forces in a multi-
degree-of-freedom structure. Their results demonstrated the effectiveness of IPV in
identifying highly non-linear restoring forces, without a priori information, while
preserving a direct association with the structural dynamics.
Fathnejat et al. [7] proposed method to efficiently reduce the computational cost
of model updating during the optimization process of damage severity detection,
hence, the damage index of structural elements is evaluated using properly trained
cascade feed-forward neural network (CFNN). The results indicated that after
determining the damage location, the proposed solution method for damage severity
detection leads to significant reduction of computational time compared to finite
element method.
In this study, based on previous investigations of the present authors, structural
damage detection is performed incorporating several methods of Artificial
Intelligence (AI) including back-propagation neural networks (BPNN), Least Square
Support Vector Machines (LS-SVMs), Adaptive Neural-Fuzzy Inference System
(ANFIS), radial basis function neural network (RBFN), Large Margin Nearest
Neighbor (LMNN), Extreme Learning Machine (ELM), Gaussian process (GP) and
the comparative results are presented. By considering dynamic behaviour of a

2
structure as input variables, seven AI methods are constructed, trained and tested to
detect the location and severity of damage in civil structures. The variation of
running time, mean square error (MSE), number of training and testing data, and
other indices for measuring the accuracy in the prediction are defined and calculated
in order to inspect pros and cons of each algorithm.
The article is structured as follows. Section 2 presents the statement of damage
detection problem for a truss structure. The AI algorithms compared in this study are
concisely reviewed in Section 3. In Section 4, the relative merits of the algorithms
are assessed by solving Damage Detection (DD) problems. Some concluding
remarks are provided in Section 5.

2 Structural Damage Detection

Structural damage detection techniques are generally classified into two main
categories. They include the dynamic and static identification methods requiring the
dynamic and static test data, respectively. Furthermore, the dynamic identification
methods have shown their advantages in comparison with the static ones. Among
the dynamic data, the modal analysis information of a structure such as the natural
frequencies and mode shapes has been widely used for damage detection [8]. In this
study, we use the measured natural frequencies as a simple tool for identifying the
location and extent of structural damage. When the natural frequencies are employed
to identify the damage, a parameter vectors may be determined. A parameter vector
used consists of the ratios of the first n f natural frequency changes F due to
structural damage, i.e.
Fh  Fd
  F      
Fh

where Fh and Fd denote the natural frequency vectors of the healthy and damaged
structure, respectively. We use this vector as input vector of AI algorithm.

3 Artificial Intelligence Methods

Data-driven approaches use information from observed data to identify the


characteristics of damage progress and predict the future state without using any
particular physical model. Instead, mathematical models and/or weight parameters
are employed, which are determined based on the training data that are obtained
under various usage conditions. Since the data-driven approaches depend on the
trend of data, which often show a distinct characteristic near the end of life, they are
powerful in predicting near-future behaviours, especially toward the end of life.
The data-driven approaches are generally divided into two categories [9]: (1) the
artificial intelligence approaches that include neural network (NN) and fuzzy logic;
and (2) the statistical approaches that include the Gaussian process (GP) regression,
relevance/support vector machine, least squares regression, the gamma process, the
Wiener processes, hidden Markov model, etc. Among these algorithms, AI

3
algorithms mentioned in introduction are most commonly used for prognostics, and
thus, will be discussed in this paper.

3.1 Artificial Neural Networks


An ANN is a parallel information processing paradigm based on models of
biological neurons [10]. Neural network consists of a large number of highly
interconnected artificial neurons (also called cells, units, nodes, or processing
elements) operating in parallel. Neural network consists of an input layer, one or
more hidden layers, and an output layer. In the hidden and output layers, an
additional neuron with a constant output (usually 1) is often added to each neuron.
These nodes are known as bias neurons. Every neuron in a layer receives its input
from the output of the neuron in the previous layer or from the network’s input
layer. Each of the neurons has an associated transfer function which describes how
the sum of its inputs of the neurons in the hidden and output layers is converted into
an output.
A neural network acquires knowledge during learning process which is stored
within inter-neuron connection, strengths known as synaptic weights. Learning in
ANNs involves adjustments of the synaptic weights, in order to adjust the weights,
the neural networks make use of either supervised or unsupervised learning
paradigms. In the supervised learning stage, a neural network can learn from training
samples defining the various cases and can determine the class of new data based on
previous knowledge. Unsupervised learning algorithms patterns are recognized by
the neural networks based on the features present in them. The most common and
most popular neural network model among all the existing neural network
paradigms is the multilayer perceptron (MLP) trained by back propagation (BP)
algorithm. Back propagation is known as a supervised network because it requires a
desired output in order to learn. The multilayer perceptron neural networks are based
on the error correction learning. During the learning process, the interconnections
between neurons are associated with synaptic weights that are adjusted to reduce the
difference (error) between the desired output value and corresponding output. It has
been reported that neural networks are able to represent both linear and non-linear
relationships and in their ability to learn these relationships directly[11].
The structure of RBF neural network is similar to that of MLP [12]. It consists of
layer of neurons. The main distinction is that RBF has a hidden layer which contains
nodes called RBF units. Each RBF has two key parameters that describe the location
of the function’s center and its deviation or width. The hidden unit measures the
distance between an input data vector and the center of its RBF. The RBF has its
peak when the distance between its center and that of the input data vector is zero
and declines gradually as this distance increases. There is only a single hidden layer
in a RBF network there are only two sets of weights, one connecting the hidden
layer to the input layer and the other connecting the hidden layer to the output layer.
Those weights connecting to the input layer contain the parameters of the basis
functions. The weights connecting the hidden layer to the output layer are used to
form linear combinations of the activations of the basis functions (hidden units) to

4
generate the network outputs. Since the hidden units are nonlinear, the outputs of the
hidden layer may be combined linearly and so processing is rapid [12].

3.2 Adaptive Neuro-fuzzy Inference Systems


The adaptive neuro-fuzzy inference system (ANFIS) was introduced by Jang in
1993 [13]. The ANFIS integrates fuzzy logic with the neural network to improve the
system performance and also to rapidly and accurately approximate complex
functions. The ANFIS uses fuzzy rules for adaptation of a set of model parameters
and the neural network for training and updating these parameters.
Fuzzy logic and neural network technology were proved useful for the modelling
of relationships between the process parameters and process outputs where the
mathematical model of the process does not exist. However, both these methods
have their advantages and shortcomings. The ANNs provide high accuracy input–
output mapping for non-linear system modelling. However, one fundamental
weakness of the ANNs is that they are ‘black box’ models and they are incapable of
explaining a particular decision to the user in a human comprehensible form. On the
other hand, fuzzy logic has the ability of modelling human knowledge in a form of
‘if–then’ rules using easily interpretable and understandable linguistic term. For the
complex computation, it has the capability of transforming the linguistic and
heuristic terms into numerical values via fuzzy rules and membership functions. The
main disadvantage is that the fuzzy rules and the membership function cannot to be
selected and adjusted automatically. The ANFIS combines the advantages of fuzzy
logic and ANN techniques and the individual dis- advantages from fuzzy logic or
ANN alone can be resolved [14].

3.3 SVM Summarization


Vapnik, et al. developed the SVM in the initial stage of 1990s, and SVM is a new
machine learning method setting up in the statistical learning theory [15]. Its
primary characteristics are: On the basis of structured risk minimization, specially
against the limited sample situation, its goal is receiving optimization solution under
existing information but not merely the ones while the samples becoming infinitely
great; The algorithm is turned into secondary optimization finally, and will receiving
global optimization solution in theory; The algorithm transforms the practical
problem into the high-dimensional space by non-linear transform, constructs the
function of linear discriminant in the high-dimensional space to realize the non-
linear distinguishing in the former space, so it solved dimension problem and
guaranteed to the learning machine have stronger generalization ability. SVM have
overcome the ANN’s (Artificial Neural Network) shortcomings of the bad
generalization ability, easily falling into the partial optimization, being difficult to
confirm the latent number of network, and so on. It is the best method of estimating
and predicting learning about the little sample at present. For the outstanding
learning performance of SVM, it has been applied to structure health monitoring [3].
The LS-SVM is a kind of expansion of standard SVM. It turns the inequality
constraint into equality constraint of traditional SVM, regards the square errors and

5
losing function as the experience losing of training sets, and transforms the problem
of secondary optimization into solving linear equation group issue. So LS-SVM can
reduce calculating difficulty, accelerate solving pace and anti-interference ability
[16].

3.4 Large Margin Nearest Neighbor Classification


The k-nearest neighbors (kNN) rule is one of the oldest and simplest methods for
pattern classification. Nevertheless, it often yields competitive results, and in certain
domains, when cleverly combined with prior knowledge, it has significantly
advanced the state-of the-art [17]. The kNN rule classifies each unlabeled example
by the majority label among its k-nearest neighbors in the training set. Its
performance thus depends crucially on the distance metric used to identify nearest
neighbors.
Weinberger et al [17] proposed Large Margin Nearest Neighbor Classification
(LMNN) and shows how to learn a Mahanalobis distance metric for k-nearest
neighbor (kNN) classification by semidefinite programming. The metric is trained
with the goal that the k-nearest neighbors always belong to the same class while
examples from different classes are separated by a large margin. On seven data sets
of varying size and difficulty, they find that metrics trained in this way lead to
significant improvements in kNN classification. As in support vector machines
(SVMs), the learning problem reduces to a convex optimization based on the hinge
loss. Unlike learning in SVMs, however, they framework requires no modification
or extension for problems in multiway (as opposed to binary) classification.

3.5 Extreme Learning Machines


An ELM is a novel and fast learning method based on the structure of multi-layer
perceptrons, recently proposed in and applied thereafter to a large number of
classification and regression problems [18]. The ELM approach is a novel way of
training feedforward neural networks, with perceptron structure. The most
significant characteristic of the ELM training is that it is carry out just by randomly
setting the network weights, and then obtaining the inverse of the hidden-layer
output matrix. The advantages of this technique are its simplicity, which makes the
training algorithm extremely fast, and also its outstanding performance when
compared to avantgarde learning methods, usually better than other established
approaches such as classical multilayer perceptrons or support vector machines.
Moreover, the universal approximation capability of the ELM network, as well as its
classification capability, have been already proven [19].

3.6 Gaussian Processes


The prior of GP is that functions over outputs have a multivariate Gaussian (MVG)
distribution , and that target values are observed with the Gaussian noise, so that the
distributions of output functions also follow the MVG distribution. Since the MVG
distribution is defined with the mean and covariance function, if we design the mean

6
and covariance function properly, we can estimate the conditional mean and
variance of future values. Recently, GP has been commonly used in various fields,
such neuroscience, geoscience, as image processing [20].
The GP regresses a function over the observation data on the target data under the
assumption that possible functions have the MVG distribution with finite dimension.
The number of target data becomes the dimension of MVG distribution. The means
of a MVG distribution are assumed to be zero, and the variance of a MVG
distribution is the covariance function of observation data. Therefore, the time-series
of target data is a sample path from the MVG distribution. This assumption gives the
prior on the target data. It is also assumed that the outputs of a function are affected
by noise processes, which have the multivariate Gaussian distribution, Poisson
distribution, or Laplace distribution. These noise processes decide the likelihood
over the parameters in the regression analysis. Under these two assumptions, the
structure of the GP is constructed. In addition, the GP assumes that the dimension of
the output is one, but a multiple output method is also used, called the co-Kriging
method [21].

4 Damage Detection Procedure

Neural network (and other mentioned AI methods) must be put through a training
cycle, for it to predict the severity and location of damages of structure. This training
set should therefore encompass all damage sizes and locations. The input to the
neural network is the damage indices of different damage signature, and the desired
output (target) is the actual damage position and damage severity [22]. In order to
create a training set to train the network, a good number of finite element models
have to be created. This is because various permutations and combinations of
damage sizes and locations have to be modelled for the network to generalize
accurately. In this study all possible combination of damage sizes and locations are
considered and dynamic response of structure for each of them calculated.

5 Numerical Results of Damage Detection

In order to validate the ability of AI methods for identifying the multiple structural
damages, below test example is considered.
The 31-bar planar truss shown in Fig. 1 selected from Ref. [23] is modelled using
the conventional finite element method without internal nodes leading to 25 degrees
of freedom. In this example, the first 5 vibrating modes are utilized for damage
detection. The material density and elasticity modulus are 2770 kg/m and 70 GPa,
respectively. Damage in the structure is simulated as a relative reduction in the
elasticity modulus of individual element as

E  Ei
xi  , i  1,..., n  
E

7
where E is the original modulus of elasticity and Ei is the final modulus of
elasticity of i th element. Two different damage cases given in Table 1 are induced
in the structure and the proposed method is tested for each case.

Figure 1: The 31-bar planar truss.

Case 1 Case 2
Element Number Damage Ratio Element Number Damage Ratio
11 0.25 1 0.30
25 0.15 2 0.20

Table 1: Damage cases.

The damage identification results for cases 1–2 are shown in Fig. 2-3, respectively.
It is observed that the optimization process achieves to the site and extent of actual
damage truthfully.

Figure 2: Final identified damage variables of the 31-bar planar truss for case 1

8
Figure 3: Final identified damage variables of the 31-bar planar truss for case 2

Table 2 shows the results of comprehensive comparing between seven solution


methods in terms of computational speed and accuracy.

Number Of
AI method Number Of
Training MSE Time (S)
Testing Data
Data
LS-SVM 8137 3488 0.00051 300
RBFNN 9300 2325 0.00678 440
ANFIS 10000 1625 0.048 700
LMNN 10000 1625 0.1 1200
BPNN 9300 2325 0.00981 680
ELM 8137 3488 0.00031 350
GP 10000 1625 0.048 910

Table 2: Comparison the results between seven solution methods

Based on Table 2, LS-SVM have better results in comparison to ANN, meanwhile


uses lesser training data to acquire knowledge of data. This due to that LS-SVM was
very effective for sparse and high dimensional data. Furthermore, LS-SVM have
better generalization abilities than the ANN (Both BPNN and RBFNN). The major
drawback of using BPNN was computational cost for the potentially large size of the
hidden layer which could be equal to the size of the input vector.
Due to that DD was high dimensional data problem, creating Sugeno fuzzy model
is time consuming and complicated process for ANFIS, hence resulted ANFIS
model is poor in mapping input-output relation.
The most time consuming and inaccurate results belong to LMNN method,
because LMNN is classification method basically, and we use it in regression
problem. Furthermore, frequencies of damaged scenarios was very close to each

9
other, therefore creating appropriate clusters and boundaries between them is
difficult for algorithm.
In the data-driven approaches, NN is chosen because GP is the same as the global
model (a polynomial function) at the extrapolation cycles, while different
combinations of transfer functions of NN can predict better than a polynomial
function.
Fundamentally, ELM tends to have better scalability and achieve similar (for
regression and binary class cases) or much better (for multi-class cases)
generalization performance at much faster learning speed (up to thousands times)
than traditional SVM. Furthermore, ELM needs much less training time compared to
popular ANN method and LMNN. Hence, the prediction accuracy of basic ELM is
better than other IA method presented in this study and similar to LS-SVM in SHM
application.
Finally, In comparison to another AI method presented in this paper LS-SVM
have lesser computation time.

6 Concluding Remarks

This paper provides a comprehensive comparison of AI method for damage


detection of structure. As common prognostics algorithms, NN, GP, LS-SVM,
ANFIS, ELM and LMNN are reviewed and employed for case studies to discuss
their attributes, pros and cons and applicable conditions. Even if advanced
algorithms are available, basic algorithms are employed in this study and the results
are analysed by focusing on their intrinsic characteristics. This will be helpful for
beginners in prognostics to choose the best algorithm for their field of application.
Furthermore, one can improve each algorithm’s performance and can develop hybrid
approaches after understanding intrinsic characteristics of each algorithm, which are
expected as future works on the topic of prognostics algorithms.

References

[1] M. Zhang and R. Schmidt, “Sensitivity analysis of an auto-correlation-


function-based damage index and its application in structural damage
detection,” J. Sound Vib., vol. 333, no. 26, pp. 7352–7363, 2014.
[2] R. Yan, R. X. Gao, and X. Chen, “Wavelets for fault diagnosis of rotary
machines : A review with applications,” Signal Processing, pp. 1–15, 2013.
[3] R. Ghiasi, P. Torkzadeh, and M. Noori, “Structural damage detection using
artificial neural networks and least square support vector machine with
particle swarm harmony search algorithm,” Int. J. Sustain. Mater. Struct.
Syst., vol. 1, no. 4, pp. 303–320, 2014.
[4] R. Ghiasi, P. Torkzadeh, and M. Noori, “Structural Damage Detection Using
LS-SVM Based on a New Combinational Kernel Function,” Struct. Heal.
Monit., 2014.

10
[5] R. A. S. A. N. G. V Popov, “Crack identification in curvilinear beams by
using ANN and ANFIS based on natural frequencies and frequency response
functions,” pp. 1629–1645, 2012.
[6] S. Saadat, G. D. Buckner, T. Furukawa, and M. N. Noori, “An intelligent
parameter varying (IPV) approach for non-linear system identification of base
excited structures,” Int. J. Non. Linear. Mech., vol. 39, no. 6, pp. 993–1004,
2004.
[7] H. Fathnejat, P. Torkzadeh, E. Salajegheh, and R. Ghiasi, “Structural Damage
Detection by Model Updating Method Based on Cascade Feed-Forward
Neural Network as an Efficient Approximation Mechanism,” Int. J. Optim.
Civ. Eng., vol. 4, no. 4, pp. 451–472, 2014.
[8] O. Yazdanpanah, S. M. Seyedpoor, and H. A. Bengar, “A new damage
detection indicator for beams based on mode shape data,” Struct. Eng. Mech.,
vol. 53, no. 4, pp. 725–744, 2015.
[9] D. An, N. H. Kim, and J. Choi, “Practical options for selecting data-driven or
physics-based prognostics algorithms with reviews,” Reliab. Eng. Syst. Saf.,
vol. 133, pp. 223–236, 2015.
[10] B. B. Adhikary and H. Mutsuyoshi, “Artificial neural networks for the
prediction of shear capacity of steel plate strengthened RC beams,” Constr.
Build. Mater., vol. 18, pp. 409–417, 2004.
[11] D. Crivelli, M. Guagliano, and A. Monici, “Development of an artificial
neural network processing technique for the analysis of damage evolution in
pultruded composites with acoustic emission,” Compos. Part B Eng., vol. 56,
pp. 948–959, 2014.
[12] I. Yilmaz and O. Kaynar, “Multiple regression, ANN (RBF, MLP) and
ANFIS models for prediction of swell potential of clayey soils,” Expert Syst.
Appl., vol. 38, no. 5, pp. 5958–5966, 2011.
[13] J.-S. Jang, “ANFIS: adaptive-network-based fuzzy inference system,” Syst.
Man Cybern. IEEE Trans., vol. 23, no. 3, pp. 665–685, 1993.
[14] A. Khajeh and H. Modarress, “Prediction of solubility of gases in polystyrene
by Adaptive Neuro-Fuzzy Inference System and Radial Basis Function
Neural Network,” Expert Syst. Appl., vol. 37, no. 4, pp. 3070–3074, 2010.
[15] J. A. K. Suykens and J. Vandewalle, “Least squares support vector machine
classifiers,” Neural Process. Lett., vol. 9, no. 3, pp. 293–300, 1999.
[16] W. U. Sen and W. E. I. Zhuo-bin, “Application of Least Squares Support
Vector Machine in the Damage Identification of Plate Structure,” in
Intelligent System Design and Engineering Application (ISDEA), 2010
International Conference on, 2010, vol. 1, pp. 351–354.
[17] K. Q. Weinberger and L. K. Saul, “Distance metric learning for large margin
nearest neighbor classification,” J. Mach. Learn. Res., vol. 10, pp. 207–244,
2009.
[18] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, “Extreme learning machine: theory
and applications,” Neurocomputing, vol. 70, no. 1, pp. 489–501, 2006.
[19] P. K. Wong, K. I. Wong, C. M. Vong, and C. S. Cheung, “Modeling and
optimization of biodiesel engine performance using kernel-based extreme

11
learning machine and cuckoo search,” Renew. Energy, vol. 74, pp. 640–647,
2015.
[20] D. Lee, S. Member, and R. Baldick, “Short-Term Wind Power Ensemble
Prediction Based on Gaussian Processes and Neural Networks,” IEEE Trans.
SMART GRID, vol. 5, no. 1, pp. 501–510, 2014.
[21] C. E. Rasmussen and H. Nickisch, “Gaussian processes for machine learning
(GPML) toolbox,” J. Mach. Learn. Res., vol. 11, pp. 3011–3015, 2010.
[22] S. J. S. Hakim and H. A. Razak, “Modal parameters based structural damage
detection using artificial neural networks-a review,” SMART Struct. Syst., vol.
14, no. 2, pp. 159–189, 2014.
[23] S. M. Seyedpoor, “A two stage method for structural damage detection using
a modal strain energy based index and particle swarm optimization,” Int. J.
Non. Linear. Mech., vol. 47, no. 1, pp. 1–8, 2012.

12

You might also like