Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Contents lists available at ScienceDirect

Journal of Wind Engineering & Industrial Aerodynamics

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

Machine learning strategy for predicting flutter performance of streamlined

box girders
Haili Liao a, Hanyu Mei b, Gang Hu c, Bo Wu b, Qi Wang a, *
a
Research Center for Wind Engineering, Wind Engineering Key Laboratory of Sichuan Province, Department of Bridge Engineering, Southwest Jiaotong University, Chengdu,
610031, China
b
Research Center for Wind Engineering, Department of Bridge Engineering, Southwest Jiaotong University, Chengdu, 610031, China
c
School of Civil and Environmental Engineering, Harbin Institute of Technology, Shenzhen, 518055, China

A R T I C L E I N F O A B S T R A C T

Keywords: Engineers often heavily rely on wind tunnel tests or computational fluid dynamics (CFD) to evaluate the flutter
Streamlined box girder performance of bridges in their preliminary design, which is costly and time-consuming. To quickly obtain the
Flutter critical flutter wind speed of streamlined box girders in the preliminary design, a machine learning (ML) strategy
Machine learning
was proposed in this paper. A big dataset was built by testing critical flutter wind speeds of 30 sectional models of
Random forest
Gradient boosting regression
streamlined box girders with and without railings at 5 angles of attack through free vibration wind tunnel tests.
The flutter predicting models, taking geometric information and dynamic parameters as inputs, were built based
on four widely-used ML algorithms, i.e., support vector regression (SVR), neural network (NN), random forest
regression (RFR), and gradient boosting regression tree (GBRT). It is shown that the NN and GBRT models exhibit
the highest prediction accuracy for the girders regardless of railings, respectively. A comparative study revealed
that the ML models were superior over those simplified formulas for flutter estimation including the Van der Put
formula, Selberg formula, and Haifan Xiang formula. A case study was also given to demonstrate the practical
application of the proposed method. These ML models provide an efficient supplement to wind tunnel tests and
CFD simulations for flutter predictions of streamlined box girders in the preliminary design.

1. Introduction 2D-3DOF method (Yang et al., 2007), closed-form solution (Chen and
Kareem, 2006; Chen, 2007), and so on. However, this strategy still
Flutter, a self-excitation vibrating phenomenon, has been widely strongly relies on wind tunnel testing (Gu et al., 2000; Chowdhury and
studied since the collapse of the old Tacoma bridge. Flutter performance Sarkar, 2004; Chen and Hu, 2005; Sarkar et al., 2009; Lin et al., 2019; Wu
is a particularly critical design criterion in the wind-resistant design of et al., 2020a) or CFD simulations (Zhu et al., 2007; Huang et al., 2009;
bridges and is closely associated with the box geometry of bridge sections Brusiani et al., 2013; Xu and Zhang, 2017) because the flutter derivatives
(Larsen, 1993). Streamlined box girder has been widely applied to the in this classical force model need to be determined by using either wind
engineering design of long-span bridges due to its superior comprehen- tunnel testing or CFD simulations.
sive aerodynamic performance (Larsen, 1993; Ding et al., 2002; Matsu-

moto et al., 2010; Zhou et al., 2019; Mei et al., 2020) and its mature 1 h_ Bα_ h
Lse ¼ ρU 2 ð2BÞ KH1* þ KH2* þ K 2 H3* α þ K 2 H4* (1a)
manufacturing techniques. In general, wind tunnel tests or numerical 2 U U B
simulations (e.g., computational fluid dynamics, CFD) can be used to


directly estimate the critical flutter wind speeds of bridge sections. 1   h_ Bα_ h
However, they are costly and very time-consuming. Another method to Mse ¼ ρU 2 2B2 KA*1 þ KA*2 þ K 2 A*3 α þ K 2 A*4 (1b)
2 U U B
fast assess the bridge flutter performance is the flutter analysis method
based on Scanlan’s linear self-excited force model (Eq. (1)), such as the Hence, to rapidly predict the flutter performance of bridge girders,
step-by-step method (Matsumoto et al., 1997; Matsumoto, 1999), several simplified formulas have been proposed for estimating flutter

* Corresponding author.
E-mail addresses: [email protected] (H. Liao), [email protected] (H. Mei), [email protected] (G. Hu), [email protected] (B. Wu), wangchee_
[email protected] (Q. Wang).

https://doi.org/10.1016/j.jweia.2020.104493
Received 14 August 2020; Received in revised form 27 November 2020; Accepted 21 December 2020
Available online 7 January 2021
0167-6105/© 2021 Elsevier Ltd. All rights reserved.
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Fig. 1. Diagram of sectional models in G1 and G2, and the detailed size of railings.

derivatives (Matsumoto, 1996; Matsumoto et al., 1996; Scanlan et al., (2011) predicted the acceleration response time history of a bridge sec-
1997; Chen and Kareem, 2002; Tubino, 2005) and critical flutter wind tion under turbulence via ANN, and the prediction accuracy was not
speeds (Kloeppel, 1967; Selberg, 1963; Van der Put, 1976; Xiang and Ge, sufficient, which can be attributed to the weakness of ANN on the fitting
2002; Chen, 2007), which facilitates bridge designs. However, the for- accuracy of time series. Li et al. (2018) classified the vortex-induced vi-
mulas of estimating flutter derivatives derived from the quasi-steady bration modes using a decision tree model and also used SVM to predict
theory do not consider the unsteady effect of the self-excited force. vortex-induced vibration of a long-span suspension bridge based on field
Furthermore, the simplified formulas for calculating the critical flutter measurement data. It was shown that both models achieved high accu-
wind speed contain undetermined coefficients, which are associated with racy and can be used in structural health monitoring of long-span bridges.
the box geometry and can only be determined by empiricism or inter- Chen and Ge (2019) identified the aerostatic coefficients and flutter
polation based on the existing database. Although these methods can fast derivatives of streamlined and inverted trapezoidal box girders based on
estimate the flutter performance, the accuracy is doubtful, and the in- experimental results of long-span bridges using an ANN. It was found that
accuracy is amplified with increasing the bridge span. this ANN-based model performed well in predicting aerostatic co-
Machine learning is a data-driven method that has been extensively efficients, while the prediction results of flutter derivatives were unsat-
introduced to address the aforementioned issues successfully. And due to isfactory. Other applications of ML in wind engineering such as
its superior extrapolation ability, several advanced research has been short-term wind speed forecasting (Cadenas and Rivera, 2009; Zhou
made to solve problems in wind engineering in an intelligent way. Li and et al., 2011; Liu et al., 2014; Yu et al., 2018; Wilms et al., 2019), surface
Y (2000) used an artificial neural network (ANN) to identify the aero- wind pressure prediction of bluff bodies (Jin et al., 2018; Hu and Kwok,
static coefficients of typical bridge sections based on a dataset collected 2020; Tian et al., 2020; Hu et al., 2020) have also achieved great success,
from the wind tunnel tests. However, the results showed that the pre- confirming the feasibility and effectiveness of the data-driven methods in
diction errors were mostly above 20%, which is not suitable for engi- wind engineering (Bao and Li., 2020).
neering applications. Jung et al. (2004) adopted NN as a nonlinear This study aims to fully utilize the excellent nonlinear mapping and
mapping system to identify flutter derivatives of various cross-sections predictive abilities of machine learning (ML) to build a nonparametric
based on several bridge decks with different widths, depths, and model for estimating critical flutter wind speeds of streamlined box
heights of the center barrier, and the influence of the proposed network girders. The inputs of the model comprise the box geometry information
structure was also studied. Chen et al. (2008) assessed the flutter de- of bridge girders and dynamic parameters of the bridge; the output is the
rivatives of rectangular cross-sections with different width-height ratios critical flutter wind speed. Four widely-used and well-performed ML
by using ANN. The predicted flutter derivatives of unknown cases were models in wind engineering including SVR model (Lute et al., 2009; Li
then utilized to compute the critical flutter wind speeds. Results showed et al., 2018), NN model (Jung et al., 2004; Tian et al., 2020), RFR model
that this approach was capable of predicting flutter derivatives and (Hu et al., 2020), and GBRT model (Hu and Kwok, 2020), are tested. To
providing a feasible option for the determination of the basic shape of the build a dataset for this ML study, 30 streamlined box girders with or
bridge section in the preliminary design. Lute et al. (2009) employed without railings were fabricated and used to conduct the free vibration
support vector machines (SVM) to identify the flutter derivatives of wind tunnel tests to obtain their critical flutter wind speeds at 5 different
bridge sections and the corresponding critical flutter wind speeds. angles of attack. The optimum hyper-parameters of each ML model were
Compared to the assessment results of simplified empirical formulas and determined by using a combination of k-fold evaluation method and grid
the methods proposed by Xiang et al. (1998) based on three bridges, the search method. A case study is also given to illustrate the practical use of
proposed method exhibits sufficiently better accuracy. Wu and Kareem the proposed method.

Fig. 2. Design domain of sectional models for both groups.

2
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Fig. 3. Sectional model assembled in wind tunnel: (a) Group 1; (b) Group 2.

2. Data preparation Appendix A. The range of mass parameters and frequencies are consistent
with most of the practical bridge characteristics. And the damping ratios
2.1. Wind tunnel experiments were controlled below 0.35%. By normalizing these parameters, which
will be illustrated below, these dynamic parameters will be converted to
A big dataset is a prerequisite of training an applicable ML model for a normal distribution with a larger range. Thus, a total of 150
predicting the flutter performance of streamlined box girders. The (2  15  5) critical flutter wind speeds at 5 angles of attack were ob-
simplest way to build this dataset is to collect data from literature or wind tained, and consequently, two datasets for G1 and G2 (i.e., each group
tunnel test reports. However, considering the diversity of the section type has 75 samples) were built. All the critical flutter wind speeds are pre-
of girder in literature (e.g., streamlined box girder, steel truss girder, etc.) sented in Fig. 4.
and various testing conditions, it is challenging to systematically build a As shown in Fig. 4, the sectional models at positive angles of attack
practicable dataset for this study. To achieve this, 30 sectional models have a lower critical flutter speed than those at negative angles of attack
with different geometry of streamlined box girders were fabricated by for both groups. The critical flutter speed varies continuously with the
only considering the bottom plate/side panel angle, the ratio of bridge change of height (H) and bottom plate/side panel angle (θ). In addition,
width to height, and also the wind angle of attack, which are the key this trend is significantly influenced by the configuration of railings. In
point of flutter performance of streamlined box girder (Wang et al., general, the railing reduces the critical wind speed of flutter, especially
2011). 15 of them were not configured with any accessory structure under the positive angle of attack, with a few exceptions (e.g., G1H36A21
(labeled as G1); the rest (labeled as G2) has the same geometry as the vs. G2H36A21, at 3 angle of attack).
sectional models in G1, but were all configured with 3 different railings
chosen from real bridges. Considering that the size of railings of most
bridges is relatively close, the selected railings can meet most engineer- 2.2. Data preprocessing
ing requirements. Besides, every sectional model in G2 was configured
with the same railings, and hence the effect of railing size is eliminated in To train a ML model, data preprocessing, which includes feature se-
the ML model. All the models have the same length ðL ¼ 1100 mmÞ and lection and data rescaling, is essential. According to the coupled flutter
width ðB ¼ 400 mmÞ. The two group section models are sketched in mechanism revealed in previous studies (Chen and Kareem, 2006; Chen,
Fig. 1, together with the details of three types of railings. The reason for 2007), this divergent phenomenon is associated with both the geometry
setting “G2” is that sectional models in the service stage can provide a of the bridge section and the dynamic parameters in the coupled system.
certain reference value for the designer especially when it is essential to Considering the different shapes of streamlined box girder and the wind
assess the impact of railings on flutter. angle of attack, directly using vertex coordinates to uniformly charac-
A design domain was defined for the geometry of these sectional terize the box geometry properties will be beneficial to model building
models as shown in Fig. 2. It can be seen that there were 15 design points, and method extension. Thus, 7 vertex coordinates (mi ; ni ) of the section
corresponding to the sectional models in each group (i.e., G1 and G2). were selected to accurately characterize the box geometry of the
Three heights (H ¼ 36mm; 44mm; 57mm, labeled as H36, H44 and streamlined box girder, and then normalized to eliminate the nonuniform
H57, respectively) and five bottom plate/side panel angles (θ ¼ 24 ;21 ; input problem caused by the scaling ratio. For a bridge section with
18 ; 15 ; 12 , labeled as A24, A21, A18, A15, A12, respectively) were railings, the railings used in each section are identical, and hence the
sampled from the design domain. However, it should be noted that, the features of the railings were not be considered as input. For different
testing sections basically covers the scope of current bridge engineering’s wind angles of attack, it is essentially a change in the coordinates of the
need. Due to the cost of the wind tunnel test and the main purpose of this vertexes. Therefore, the effect of the angle of attack can be considered by
research is to promote the proposed method, only some factors that affect manipulating the coordinates of the vertexes (Fig. 5). The new vertex
0 0
the box geometry shape of the streamlined box girder are considered coordinates (mi ; ni ) of the section with the wind angle of attack can be
here. Other properties such as the slope of the bridge deck and the expressed as Eq. (2).
chamfer of the wind faring are ignored.


 0
!
The free vibration wind tunnel tests were performed in the XNJD-2 cosðαÞ sinðαÞ mi 1=B mi
⋅ ⋅ ¼ (2)
wind tunnel (Fig. 3). The sectional model was elastically mounted on 8 sinðαÞ cosðαÞ ni 1=H 0
ni
extension springs, which allowed the sectional model to rotate and move
vertically. The mechanical frequencies and masses of the test model were where α is the clockwise direction of the section, and airflow comes from
modified by adjusting the length of springs and adding mass blocks at the the left-hand side. When α is positive, the wind angle of attack is positive,
0 0
end of the model, respectively. Meanwhile, a chamfered rigid plate with a and vice versa. (mi ; ni ) are the coordinates of the section vertices after
thickness of 6 mm was added and fixed to the model ends to ensure the unified processing; i ¼ 1–7 represent the A–G vertices, respectively. It
model undergoing a two-dimensional flow. During testing, the wind should be noted that the practical bridge section is generally symmetri-
speed was increased gradually, and the critical flutter wind speed Ucr was cal. To reduce redundant input information and input feature dimensions
recorded until oscillation of the sectional model diverges. Each sectional as well as improve the generalization ability of the ML model, only 4
model was tested in a smooth flow with 5 angles of attack (α ¼ 0 ; 3 ; vertex coordinates of A, B, C, and D are reserved as input features in the
5 ). The dynamic parameters of the free vibration systems are listed in model training, which thus produces 8 input features.

3
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Fig. 4. Critical flutter wind speeds for building datasets: (a)(b)(c) G1; (d)(e)(f) G2.

The critical flutter wind speed is closely related to the structural dy- leading to a 120 times difference. Therefore, the training data must be
namic parameters, including vertical and torsional frequencies fv and ft , preprocessed for better model training in terms of efficiency and
vertical and torsional damping ratios ξv and ξt , unit length equivalent convergence. And the machine learning model can have a wider range of
mass M, and mass moment of inertia IM . These dynamic parameters were perception (for example, the ratio of torsional and vertical frequency can
also taken into account as the input features in the model training. These be easily transferred into the given range of training set). In this study,
parameters can be easily obtained from the finite element analysis in the the standard normalization method (z-score) was selected to process the
preliminary stage of bridge design. In the sectional model wind tunnel features of each dimension to conform to the standard normal distribu-
testing, different dynamic parameters were tested for all the sectional tion. For each feature xi , the standard normalized new value can be given
models so that the ML algorithm could systematically recognize the un- as Eq. (3).
derlying patterns and hence yield better robustness.
Generally, the magnitude and dimension of all the features are 0 xi  μðxÞ
xi ¼ (3)
different. It creates a scattered distribution of input features and signif- σ ðxÞ
icantly reduces the training efficiency. For example, the total mass of the
dynamic system is around 6, while the damping ratio is less than 5%, where μðxÞ and σ ðxÞ are the mean and standard deviation of the values of

4
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

are used to determine whether the training sample locates in or out of the
range of ε or not; m is the number of training samples. C is the regula-
rization parameter or penalty parameter; Ucr; i is the true value of the i-th
training sample (or the critical flutter wind speed).
0  0 
Kðx; x Þ ¼ exp  γkx  x k2 ; (5)
0
where Kðx; x Þ denotes the kernel function, γ is a hyper-parameter which
needs to be adjusted, and k ⋅k represents the squared Euclidean distance.
Thus, two hyper-parameters, i.e., γ and C, are needed to be optimized by
Fig. 5. Coordinates of streamlined box girder with a positive angle of attack.
the following optimization procedure.

each dimension respectively. 3.1.2. Neural network model

To sum up, the input features of G1’s sectional model selected herein Recently, the rapid development of deep learning (DL) has resulted in
include 4 vertex coordinates, i.e., 8 features, and 6 input features of the its extensive application in computer vision and natural language pro-
dynamic parameters, a total of 14 input features. Because the sections in cessing by the powerful fitting ability in revealing complex nonlinear
G2 are configured with the same railings, the input feature dimension dependency relations between input-output patterns. Its core idea is
was identical to that of G1. The output of both G1 and G2 was the critical based on the nonlinear transformation of the basic hidden neuron shown
flutter wind speed. Each of G1 and G2 datasets had 75 sets of data, and in Fig. 8. In this study, the most typically used fully-connected NN for
the ML models were built separately based on the two groups of datasets. regression analysis was selected to address problem. A 4-layer NN
structure (i.e., 1 input layer, 2 hidden layers, and 1 output layer) was
3. Machine learning model building and training built with 14 input features, i.e. box geometry (i.e., 8 vertex coordinates)
and dynamic properties (i.e., 6 dynamic parameters). Two hidden layers
Machine learning is a path to achieve artificial intelligence. It is the were added with numbers of basic hidden neurons activated by a ReLU
scientific study of algorithms and statistical models that provides systems function (Krizhevsky et al., 2012) (Eq. (6)) after the previous input layer.
the ability to automatically learn and improve from big data by explicit The critical flutter wind speeds Ucr were then output by fully-connecting
programming. Its main strategy is based on input-output patterns. In view the last hidden layer. The constructed NN architecture is shown in Fig. 8.
of the limited number of learning samples, ML methods are better at The gradient descent method was used to minimize the Mean Squared
pattern mining of small sample data than DL methods. Fig. 6 represents Error (MSE) loss function (Eq. (7)), and the backpropagation method was
the issue to be studied based on this pattern. It is clear that ML serves as a adopted to update the weight of each neuron with a specific learning rate
“black box” that can map relevant features (both geometric and dynamic by an optimizer.
information) to the prediction of interest by self-learning. In this study, 3 hyper-parameters, including the learning rate, the
number of hidden neurons, and the hyper-parameters were optimized by
3.1. Model building the following optimization methods. Other hyper-parameters such as the
number of the hidden layer, activation function, etc., were prescribed
3.1.1. Support vector regression model with a specific value and were not included in the optimization proced-
Support vector regression (SVR), a variant of the support vector ma- ure for weighting the trade-off between the fitting ability and resource
chine classification model, was developed for regression tasks. Its strat- consumption.
egy is to maximize the margin between the predicted value and the true 
x if x > 0
value based on convex quadratic programming (Fig. 7), which is σ ðxÞ ¼ (6)
0 if x  0
formulated in Eq. (4). The kernel trick is used to map the input feature
space to a higher-dimensional feature space (hyperplane) to solve the
N
2
nonlinear problem by kernelized linear discriminant analysis. The details 1 X
L ¼ yi  by i (7)
of the algorithm can be found in the literature (Cortes and Vapnik, 1995). N i¼1
Herein, the radial basis function (RBF) is selected as the kernel function
(Eq. (5)). where L represents the loss function, yi is the true value, and b
y i is the
predicted value.
1 Xm
 
min w2 þ C ξi þ ξ*i
2 i¼1 3.1.3. Random forest regression model
RFR is a ML algorithm that belongs to ensemble bagging learning
s:t: Ucr ðxi Þ  Ucr;i  ε þ ξi ; (4) (Liaw and Wiener, 2002). Its general idea is to construct a multitude of
Ucr;i  Ucr ðxi Þ  ε þ ξbi ; decision tree models (weak learners) (Quinlan, 1986, 2014) during
training and package them into a stronger model. This method uses a
ξi  0; ξbi  0; i ¼ 1; 2; …; m bootstrap sampling to divide different training subsets from the input
training dataset to train multiple different decision trees (i.e., weak
where w is the orientation of hyperplane; ε is the prescribed tolerance; learners) and to obtain the average prediction results of multiple internal
Ucr ðxÞ is the regression function; x is the input features with a dimension decision trees in the prediction stage. The final result is more robust and
of 14  ðnumber of training samplesÞ; ξi and ξ*i are slack variables that stable than a single weak learner, because the algorithm is insensitive to

Fig. 6. Input-output pattern for the target issue.

5
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Fig. 7. Schematic diagram of SVR.

the dependence of hyper-parameters, and it can learn a model easily with ML model. Especially, this testing set should be sampled based on the
satisfying performance. Fig. 9 represents the constructed RFR model same distribution of the whole dataset, that is to say, the testing set and
based on this procedure. In this study, the number of decision trees to be the training set should be mutually exclusive. For training a well-
used, the number of features used in every weak learner when seeking performed ML model that can be used both for later research and engi-
the best split, and the maximum depth of the decision trees, is set as the neering application, it is essential to choose the optimum parameter
hyper-parameters to be optimized in the model by using the following combination to achieve the best performance. In ML, the hyper-
optimization methods. parameter is a parameter whose value is pre-defined and then opti-
mized to control the learning process. Unlike the parameters that can be
3.1.4. Gradient boosting regression tree model learned from the training samples, e.g., weights and biases in NN model,
Compared with the RFR model, GBRT is also an ensemble technique the hyper-parameters are directly related to the model structure (e.g., the
but belongs to boosting methods. Like other boosting methods, GBRT number of hidden neurons in NN) and nonlinear mappings (e.g., the
combines weak learners (usually decision trees) into a single stronger selection of kernel function in SVR), etc. In this study, a combination of k-
leaner in an iterative forward stage-wise fashion by minimizing the MSE fold evaluation and grid search method was used for model training and
(Eq. (7)) loss function using gradient descent algorithm. To avoid over- hyper-parameter optimization.
fitting as well as increase the generalization ability, a specific rescale
ratio (or weight), which is called learning rate (Lr), is exerted on the 3.2.1. k-fold evaluation
predicting results of each weak learner, which is similar to the fully- Several methods have been developed in the ML field to evaluate the
connected layer in NN. A GBRT model is constructed and shown in model performance, including hold-out validation, cross-validation (CV),
Fig. 10. In this model, for the weak learners after the previous one, the and the bootstrap. In the deep learning field, the common practice is to
prediction residual (for MSE loss, the negative gradient can be thought as set aside a validation dataset by a specific dividing ratio, which is the
residual) in the previous step is served as the input training data. The hold-out method. However, this approach functions well only when data
critical flutter wind speed Ucr is the summation of weighted Ucr predicted is sufficient. For the bootstrap method, it changes the original distribu-
by each weak learner. tion of the dataset when the dataset is small, which will introduce esti-
In this model, the number of weak learners n, maximum depth of mation bias additionally and thus should not be used here. To maximize
decision trees, and the learning rate will be optimized to obtain the op- the use of data and satisfy the requirements concerning data distribution
timum GBRT model. and bias reduction, the k-fold CV method proposed by (Reich and Barai,
1999) was adopted here to achieve better performance. The general
procedure for the k-fold CV method is described as follows and shown in
3.2. Evaluation strategy and hyper-parameter optimization
Fig. 11:
Generally speaking, a separate dataset called testing set can be set
aside from the whole dataset to evaluate the generalization ability of a

Fig. 8. Architecture of fully-connected NN: (a) basic hidden neuron (b) neural network structure.

6
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Fig. 9. The constructed RFR model.

Fig. 10. The constructed GBRT model.
1. A testing dataset is firstly set aside for testing generalization ability
and the visualization representation in terms of prediction. (MAPE) shown as Eq. (8). R2 can provide a measure of how well the
2. The rest of the dataset is shuffled randomly and then split into k predicted value (between 0 and 1) are obtained by a regression model.
subsets. RMSE represents the square root of the difference between the predicted
3. Choose k-1 subsets as the training set and the rest one as the valida- value and the true value (non-negative), whose unit here is m=s. MAPE
tion set by k rounds. The training set is used for model training and can show the percentage error between the true value and the predicted
the validation set is used for error (e.g., Mean square error) or score value and can also be illustrated as the average relative percentage error.
(R-squared value) evaluation.
Pm
4. Return the average error or score as the final results of the trained ðyi  by i Þ2
R2 ¼ 1  Pi¼1
m 2 (8a)
models in k validation rounds.
i¼1 ðyi  yÞ

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2ffi
m

In this study, 20% of samples in each group were randomly extracted
from the whole dataset (G1 and G2) by the same random seed and were 1 X
RMSE ¼ yi  by i (8b)
set as the testing sets for evaluating the ML model generalization ability. m 1
That is to say, the two testing sets in each group have 15 (20% 75)
m
 
testing samples, leading to 60 (80%  75) training samples in each group 1 X 
yi  by i   100%
MAPE ¼  (8c)
for k-fold CV procedure. Besides, research (Friedman et al., 2001) shows m 1 yi 
that setting k ¼ 10 can strike a balance between the size of the training set
(i.e., k-1 subsets) and the overlap between different training sets in every where yi and b y i are the true and predicted value, respectively; m ¼ 15
validation round. Thus, in this study, the 10-fold CV method was intro- represents the number of testing samples.
duced for model training. Figs. 12 and 13 represent the R-squared score (R2 ) predicted by four
different ML models on the training set and test set of both groups,
3.2.2. Grid search method respectively. Firstly, it can be seen that the distribution of samples in the
Several ways have been developed to optimally find a tuple of hyper- training set and testing set are roughly close to each other from a qual-
parameters that yields an optimal model by minimizing the errors ob- itative point of view. For example, in G1, a smaller number of samples is
tained in CV procedure, such as grid search, random search, Bayesian distributed in low wind speed regions and a larger number of samples is
optimization, gradient-based optimization, etc. In this study, the tradi- distributed in high wind speed range. Secondly, it can be seen from
tional grid search method was utilized to optimize the hyper-parameters Fig. 12 that the NN, RFR, and GBRT models show great performance in
in each ML model. This method is simply an exhaustive searching the training set, while the SVR model achieves worse results on the
through a manually specified subset of the searching space of hyper- training set (R2 ¼ 0:800) but shows great generalization ability on
parameters. This method can easily be guided by the performance testing set (R2 ¼ 0:867). On the other hand, the performance of the NN,
metric measured by the above-mentioned k-fold CV procedure. Literature RFR, and GBRT models on the testing set is not as good as the perfor-
(Hu and Kwok, 2020) compared the influence of different mance on the training set. From a comprehensive perspective, the NN
hyper-parameters on the prediction performance of different models in model shows the greatest fitting and prediction ability on G1, which has a
detail based on this method. In this study, only the searching space for
R2 ¼ 0:985 on training set and R2 ¼ 0:880 on testing set (the highest
hyper-parameter tuning and the corresponding optimum values were
score among the four). As for the predicting results on G2 shown in
given for brevity (Table 1).
Fig. 13, the NN model still performs great who achieves a R2 ¼ 0:889
score on testing set, but a little bit lower than the RFR model and GBRT
4. Result representation and comparison
model (R2 ¼ 0:903 and R2 ¼ 0:910 on testing set, respectively). Thus,
the GBRT model should be treated as the best one in terms of fitting and
4.1. Prediction results representation
generalization ability on the G2 dataset.
Figs. 14 and 15 show the RMSE and MAPE results of different ML
Based on Python 3.7.1, the four ML algorithms above (SVR, NN, RFR,
models of both G1 and G2. Similarly, these two metrics illustrate the
GBRT) were implemented based on the training samples in both groups
predicting performance as well as the generalization ability on both
by using the optimized hyper-parameters shown in Table 1. Their per-
training set and testing set. It can be seen from Fig. 14 that the GBRT
formance on the testing sets was then evaluated. It should be noted that, a
model achieves the lowest RMSE and MAPE on training set, while the
certain amount of time will be spent in model parameter optimization
corresponding errors on testing set are not satisfying, indicating that the
and model training. Once the ML model is trained, the prediction pro-
GBRT model has an overfitting problem and thus has a lower general-
cedure will be finished in a few seconds, which should be very conve-
ization ability than SVR, NN, and RFR model. In contrast, the NN model
nient for engineering applications. Herein, three different error metrics
performs the best on G1 (RMSE : 0:369m=s; MAPE : 2:61%Þ than the
were used to synthetically quantify the generalization ability of different
other models, which has also been proved by the R2 results shown in
ML models, including R-squared score (R2 , coefficient of determination),
Fig. 12. As for G2 shown in Fig. 15, the RMSE and MAPE results of GBRT
root-mean-squared error (RMSE), and mean absolute percentage error

7
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Fig. 11. Diagram of k-fold CV procedure.

model on the testing set are 0:500 m=s and 5:86%, respectively. Thus, it Table 1 should be chosen as the optimum ML models for predicting the
achieves the better generalization ability than the other three ML models. critical flutter wind speeds of G1 and G2 datasets, respectively. Quali-
In addition, the NN model and RFR model also show great ability on G2 tatively speaking, the reason why the two different data sets have
in predicting the critical flutter wind speed, but still are not as good as the different optimal ML models is that the evolution pattern of the critical
performance of the GBRT model. flutter wind speed of G1 and G2 varying with the box geometry is
To sum up, the four ML methods all have good performance in pre- different, and different ML models have their pattern mining capabilities.
dicting the critical flutter wind speed of streamlined box girder, but still
exhibit different prediction ability on different issues. By combining the 4.2. A comparative study
prediction results of three different error metrics, the NN model and the
GBRT model configured with the optimized hyper-parameters given in To further demonstrate the superiority of ML approach in the pre-

Fig. 12. R-squared scores of Group1 (G1) predicted by different ML models: (a) SVR; (b) NN; (c) RFR; (d) GBRT.

8
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Fig. 13. R-squared scores of Group2 (G2) predicted by different ML models: (a) SVR; (b) NN; (c) RFR; (d) GBRT.

diction of critical flutter wind speeds of the streamlined box girder, three rffiffiffiffiffiffiffiffiffiffi
mr
commonly-used simplified formulas including the Van der Put formula Ucr ¼ 2:5μs μα Bft ; (9c)
(Van der Put, 1976) (Eq. (9a)), the Selberg formula (Selberg, 1963) (Eq. πρb3
(9b)), and the Haifan Xiang formula (Xiang and Ge, 2002) (Eq. (9c)),
where μs and μα are the reduction coefficients for the box geometry and
were used to predict the critical flutter wind speed of the bridge. These qffiffiffiffi
formulas have been included in the Chinese specifications ((Ministry of angle of attack effect, respectively; r ¼ Imm is the radius of inertia; ρ ¼
Communications of the People’s Republic of China, 2004)) for bridge 1.225 kg/m3 is the air density. For a streamlined box girder, μs ¼ 0.7. For
design reference. an angle of attack of 0 , μα ¼ 1:0, and for an angle of attack of 3 , μα ¼

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
0.8. Especially, for an angle of attack of 5 , the reduction coefficient is
ft mr B empirically defined as μα ¼ 0:7 based on the wind tunnel testing data
Ucr ¼ μs μα 1 þ  0:5 5:76 ð2π ft Þ ⋅ (9a)
fv πρB3 2 shown in Fig. 4. It is noteworthy that the above formulas do not consider
s

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi the bridge section configured with railings. That is to say, for a specific
 pffiffiffi ffi
f 2 8 2mr bridge section in the erection and under-serviced stages, they have the
Ucr ¼ 0:44μs μα ð2π ft Þ 1  v2 (9b) same critical flutter wind speeds computed by the above formulas, which
ft πρB

Fig. 14. RMSE and MAPE results of different ML models in G1: (a) RMSE (b) MAPE

9
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Fig. 15. RMSE and MAPE results of different ML models in G2: (a) RMSE (b) MAPE

is an obvious weakness of these empirical formulas. Based on the dy- method to the assessment of flutter performance during the preliminary
namic parameters and box geometry in the testing set, the critical flutter stage of bridge design. To visually show the application of the method
wind speeds of bridge sections in both Groups 1 and 2 can be computed proposed in this study and give instructions to the bridge designers, a
and their corresponding prediction errors can be obtained. flow chart summarizing the prediction process of critical flutter wind
It can be seen from Fig. 16 and Fig. 17 that all the three formulas speeds of bridges with streamlined box girder at different angles of attack
perform unsatisfactory confirmed by the three different error results on based on the above ML method is given in Fig. 18. This flow chart con-
both G1 and G2. The negative values of R2 represent that the predicted tains the acquisition of cross-section geometry information (i.e., vertex
results of the simplified formulas are even worse than the average values coordinates) and the acquisition of dynamic properties (i.e., m;Im ;fv ;ft ;ξv ;
y illustrated by Eq. (7a). In detail, the predicted flutter speeds of the three ξt ) based on finite element analysis. Furthermore, the transformed vertex
simplified formulas are conservative on G1, and the prediction results on coordinates and dynamic parameters are normalized in terms of Eq. (2)
G2 are discrete (i.e., Fig. 16). Compared with the three simplified for- and then input into the well-trained corresponding ML models (i.e., NN
mulas, the prediction results of the machine learning model (i.e., NN for or GBRT model), to predict the critical flutter wind speed.
G1, and GBRT for G2) are basically distributed close to the true value,
which is more accurate. Fig. 17 illustrates the RMSE and MAPE results of 5.2. An engineering example
different prediction models, which further represents that a higher pre-
dicting accuracy can be obtained by ML models than simplified formulas. Here takes an engineering project as an example for illustrating the
The evaluation results of the NN model and the GBRT model demonstrate practical application of our proposed approach based on ML strategy. The
the superior performance of accurately predicting the critical flutter wind selected project herein, the Fuma Yangtze River Bridge, which has a main
speeds of the bridge section. It is worth mentioning that these simplified span of 1050 m, is a highway suspension bridge over the Yangtze River in
formulas are more suitable in predicting the critical flutter wind speeds of Chongqing, China. A streamlined box girder with a width of 32.0 m and a
bridge sections in G2 rather than the sections in G1 (illustrated by height of 3.2 m is used in this bridge, and the standard cross-section
Fig. 17), which are not configured with railings on the bridge deck. (Fig. 19) is compared to Fig. 2 to facilitate the observation of the posi-
tion of the cross-section in the design domain (The geometry size is scaled
5. Practical application to match the width of both). Three different kinds of railings are
configured on the bridge deck at different positions to ensure traffic
5.1. Prediction procedure illustration safety.
According to the results of finite element analysis, the dynamic pa-
So far, the feasibility and superiority of ML strategy on the prediction rameters required for wind tunnel testing were figured out. The free vi-
issue of critical flutter wind speeds of the streamlined box girders have bration wind tunnel tests were conducted in the XNJD-1 wind tunnel
been demonstrated. However, another question is how to leverage this (shown in Fig. 20). The specific wind tunnel testing procedure can refer

Fig. 16. R-squared scores of testing set predicted by different simplified formulas (a) G1 (b) G2.

10
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Fig. 17. RMSE and MAPE results of different prediction models: (a) RMSE (b) MAPE

to Section 2.1. Sectional models configured without and with railings are parameters and cases are summarized in Table 2. Although the dynamic
labeled as model 1 and model 2, respectively. The critical flutter wind parameters of this example are not in the range of the dataset for machine
speeds of both sectional models at 0 ; 3 angles of attack were directly learning model training, these parameters will be normalized by Eq. (3)
measured by micromanometer in a smooth flow condition when the to the range of training data to fit the machine learning model. And this
motion of the sectional model became divergent. All the dynamic combination of dynamic parameters can further examine the

Fig. 18. Flow chart for practical application concerning the prediction of critical flutter wind speeds of streamline box girder.

11
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Fig. 19. Standard cross-section of the Fuma Yangtze River Bridge (all units are in m).

Fig. 20. Sectional model assembled in wind tunnel: (a) model 1: without railing (b) model 2: with railings.

extrapolation ability (or generalization ability) of the model proposed in although the RE is generally higher than that of model 1, the GBRT
this paper. model’s prediction errors of flutter critical wind speeds at three wind
To make a comparison of the prediction performance on the selected angles of attack are all less than 6%. This may result from the different
engineering example, the four well-trained ML models are used to predict shapes of railings between the dataset for model training and the prac-
critical flutter wind speeds by following the flow chart shown in Fig. 18. tical bridge. The prediction results of the engineering example further
By inputting the vertex coordinates and dynamic parameters, the critical verify the effectiveness of the proposed method. Thus, the ML strategy,
flutter wind speeds at three angles of attack were output. Here uses the especially NN and GBRT prediction model for G1 and G2, promises an
relative error (RE), which is another form of MAPE and can be formalized accurate and applicable tool for practical engineering applications con-
 
 b cerning the prediction of critical flutter wind speed of streamlined box
as RE ¼ yi yi y i   100%, to assess the prediction results. Fig. 21 presents
girder.
the comparison results and relative errors of the critical flutter wind
speeds measured by the wind tunnel tests and predicted by the ML 6. Discussions
models at different angles of attack. All the wind speeds are converted to
prototype wind speeds using the wind speed ratio of 3.12 and 3.38 for It is noteworthy that there are several limitations of the proposed
model 1 and model 2, respectively. method to predict the flutter performance of bridges:
It can be seen from Fig. 21 that the NN model and GBRT model can
also accurately predict the critical flutter wind speeds of the target long- (1) The effects of accessory structures. The effects of railings with
span bridge at three different angles of attack than the other three ML different ventilation rates, height, numbers, and locations of
models (illustrated by the tables at the bottom of Fig. 21). The maximum configuration were not considered owing to the limited testing
RE of NN for model 1 is 3.08% (i.e., at 3 angle of attack). For model 2, resources. Moreover, other accessory structures’ characteristics

Fig. 21. Comparison between the tested and predicted critical flutter wind speeds of different ML models: (a) model 1 (b) model 2.

12
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Table 1 Table 2
Grid search space for hyper-parameter optimization of different ML models. Dynamic parameters and cases for flutter tests in wind tunnel.
Algorithm Hyper-parameter Search space Optimum value Properties Practical bridge Sectional model

G1 G2 Model 1 Model 2 Model Model

1 2
SVR Regularization {1, 5, 10, 100, 200, 500, 1.0 750
parameter: C 750, 1000} Model scale ratio 1 1:50
RBF kernel parameter: Logspace(-2, 2, 10) 0.077 0.028 Length/m 100.475 2.095
γ Equivalent mass: M/kg 15651 24069 6.260 9.628
NN Learning rate {0.001, 0.0015, 0.002, 0.002 0.003 Equivalent mass moment of inertia: 1882377 2351960 0.301 0.376
0.0025, 0.003} Im /kg ⋅ m2 =m
Number of hidden {4, 8, 16, 32, 64, 128}in 32 32 Vertical modal frequency: fv /Hz 0.1636 0.1519 2.625 2.25
neurons each layer Torsional modal frequency: ft /Hz 0.4550 0.4099 7.293 6.03
Optimizer {SGD, Adam} Adam Adam Vertical modal damping ratio: ξv / 0.46% 0.43%
RFR Number of weak {5, 10, 15, …, 100} 55 30 Torsional modal damping ratio: ξt / 0.37% 0.35%
learners Wind speed ratio 1 3.12 3.38
Number of features {2, 3, 4, …, 10} 2 4 Angle of attack/ 0; 3
Maximum depth of {2, 4, 6, …, 20} 14 6
the tree Note: model 1 and model 2 corresponds to the sectional models configured
GBRT Number of weak {5, 10, 15, …, 100} 65 95 without and with railings on bridge deck, respectively.
learners
Maximum depth {2, 4, 6, …, 20} 4 4
7. Conclusions
Learning rate {0.005, 0.01, 0.05, 0.1, 0.1 0.1
0.2}
In this study, a ML strategy was proposed herein for quickly assessing
Note: “Logspace” returns numbers spaced evenly on a log scale by base 10. For
the flutter performance of streamlined box girders for the preliminary
Logspace(-2, 2, 10), it returns a sequence containing 10 samples ranging from
design of bridges. Considering that different ML algorithms are adapted
102 to 102 .
to different issues that lead to difficulty in choosing the most qualified
one, four frequently-used and well-performed ML models including SVR,
(e.g., the heights and locations of stabilizers and wind barriers) NN, RFR, and GBRT are introduced to cater to the demand of the pro-
configured on the bridge deck or the bottom plate were neglected. posed issue. For building the essential datasets for modeling training, the
Previous studies (Wang et al., 2011) have shown that they may critical flutter wind speeds of 30 streamlined box girders (15  2, two
cause unexpected results in terms of flutter. However, it was clear groups, configured with or without railings on bridge deck, respectively)
that these factors could be easily included by expanding the at 0 ; 3 ; 5 angles of attack were obtained by free vibration wind
training set. tunnel testing and were served as the datasets for training different ML
(2) In this study, only the bottom plate/side panel angle, the ratio of models. k-fold CV method and grid search method were introduced for
bridge width and height, and also the wind angle of attack, are model training and hyper-parameter pre-defining. By using the well-
considered as the input features of ML model, other potential trained ML models with the optimized hyper-parameters, these models’
changes in the chamfer of wind faring and the slope of bridge deck prediction performance and generalization ability on the testing sets
will also influence the flutter performance of streamlined box
were evaluated by three different error metrics including R2 , RMSE, and
girder. This is not achieved in this study due to the cost of wind
MAPE. Overall speaking, the NN model and GBRT model should be
tunnel testing. However, these features can also be easily taken
chosen as the optimum ML models for predicting the flutter speeds of
into consideration by the proposed method if these factors are
streamlined box girders respectively configured without and with rail-
expressed as the variation of vertex coordinates.
ings in terms of fitting ability and prediction ability. Moreover, a
(3) This research is carried out based on typical streamlined box
comparative study was conducted by carrying out predicting critical
girder. However, there are still a small number of other types of
flutter speeds based on three different simplified formulas. Results also
cross-sections (truss-girder, twin-box streamlined box girder, etc.)
showed that the proposed ML models achieve superior prediction accu-
being applied in bridge projects (Wu et al., 2020b). If the same
racy than that of simplified formulas.
method is extended to these sections, other input features should
For giving instructions to the practical application of the proposed
be taken into account, such as the ventilation rate of truss girder
approach, an engineering example was given to verify the effectiveness
and the space between the two girders of twin-box streamlined
and feasibility of this approach on predicting the flutter performance of
box girder.
streamlined box girders. The successful application and accurate pre-
(4) Considering that only total 140 samples are used for ML model
diction results of critical flutter wind speed demonstrate that the devel-
building, it is not easy to assess whether the amount of data is
oped NN and GBRT model can be directly applied to the prediction issue
sufficient for model building or not. And it is impractical to assess
with promising accuracy, making it a possible and feasible approach to
the potential relationships between the dependency of input fea-
skip the experimental procedure in the preliminary stage of wind-
tures and the output results. This is the biggest limitation of ML
resistant design of bridge flutter. Especially, the datasets used in this
methods than the traditional simplified formulas. With the
study are a significant reference for bridge engineers and researchers.
extension of this method and the expansion of data set in the
Moreover, it should be emphasized that the purpose of this study is
future, the Deep Learning method used by Hu et al. (2020) can be
not to propose a complete alternative to wind tunnel testing but to give
referenced to explore the dependency between the input features
an approach to preliminarily evaluate the flutter performance of
and the output.
streamlined box girders. Some limitations of the proposed approach
(5) Similarly, other outputs such as flutter derivatives of streamlined
should also be paid attention to applications. The final design of the
box girder can also be taken into account, which should also be an
bridge section configured with detailed accessory structures still has to
effective way to predict flutter performance by combining the
rely on the wind tunnel testing with higher accuracy, at least at the
predicted results of ML models with flutter analysis methods.
present stage.
Further study can be carried out based on this.

13
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

CRediT authorship contribution statement interests or personal relationships that could have appeared to influence
the work reported in this paper.
Haili Liao: Conceptualization, Methodology, Supervision. Hanyu
Mei: Data curation, Writing - original draft, Software, Formal analysis. Acknowledge
Gang Hu: Writing - review & editing, Validation. Bo Wu: Visualization,
Investigation. Qi Wang: Supervision, Validation, Investigation. This paper was financially supported by the National Natural Science
Foundation of China under grant numbers 51778547 and 51678508.
Declaration of competing interest

The authors declare that they have no known competing financial

Appendix A. Dynamic parameters of free vibration wind tunnel tests


Group Angle of Attack Model label fv ðHzÞ ft ðHzÞ ξv ð%Þ ξt ð%Þ kg Im ðkg ⋅mÞ
M
m

G1 0, 3 , 5 G1H36A24 1.28 2.07 0.32 0.13 5.88 0.201

G1H36A21 1.35 1.98 0.33 0.13 5.62 0.235
G1H36A18 1.30 1.97 0.33 0.13 5.62 0.220
G1H36A15 1.26 1.91 0.31 0.13 5.60 0.219
G1H36A12 1.27 1.92 0.34 0.14 5.58 0.220
G1H44A24 1.27 2.05 0.33 0.12 6.00 0.206
G1H44A21 1.29 2.15 0.33 0.12 5.80 0.187
G1H44A18 1.29 2.14 0.31 0.13 5.72 0.187
G1H44A15 1.29 2.15 0.32 0.11 5.70 0.185
G1H44A12 1.30 2.14 0.32 0.11 5.68 0.187
G1H57A24 1.24 2.07 0.31 0.13 6.22 0.202
G1H57A21 1.29 2.05 0.33 0.10 5.96 0.211
G1H57A18 1.28 2.05 0.34 0.09 5.94 0.210
G1H57A15 1.29 2.11 0.35 0.10 5.90 0.198
G1H57A12 1.30 2.06 0.35 0.10 5.82 0.208
G2 G2H36A24 1.28 2.01 0.32 0.13 5.90 0.215
G2H36A21 1.22 2.25 0.33 0.13 5.64 0.299
G2H36A18 1.23 2.2 0.33 0.13 5.64 0.316
G2H36A15 1.22 2.25 0.31 0.13 5.62 0.298
G2H36A12 1.22 2.22 0.34 0.14 5.6 0.305
G2H44A24 1.27 1.99 0.34 0.11 6.02 0.220
G2H44A21 1.29 2.12 0.33 0.10 5.82 0.192
G2H44A18 1.29 2.14 0.34 0.09 5.74 0.188
G2H44A15 1.29 2.14 0.35 0.10 5.72 0.187
G2H44A12 1.22 2.22 0.33 0.09 5.70 0.155
G2H57A24 1.25 2.01 0.31 0.13 6.26 0.218
G2H57A21 1.28 2.12 0.34 0.14 6.00 0.196
G2H57A18 1.28 2.11 0.33 0.10 5.98 0.198
G2H57A15 1.28 2.12 0.34 0.09 5.94 0.196
G2H57A12 1.31 2.09 0.35 0.10 5.86 0.206

Appendix B. Specific testing set

H36A18 (0 ) H36A12 (þ5 ) H44A18 (3 ) H57A12 (0 ) H57A12 (þ5 )

H36A24 (þ3 ) H44A21 (0 ) H44A15 (3 ) H57A24 (þ3 ) H57A24 (3 )
H36A18 (þ5 ) H44A15 (þ5 ) H57A15 (0 ) H57A21 (þ3 ) H57A15 (3 )
Note: the value in parentheses is the angle of attack. The testing sets were set aside by the same random seed, leading to the same testing set (or
cases) in both Group 1 and Group 2.

References Chen, X., Kareem, A., 2006. Revisiting multimode coupled bridge flutter: some new
insights. J. Eng. Mech. 132, 1115–1123.
Chen, X., Kareem, A., 2002. Advances in modeling of aerodynamic forces on bridge decks.
Bao, Y., Li, H., 2020. Machine learning paradigm for structural health monitoring. Struct.
J. Eng. Mech. 128, 1193–1205.
Health Monit. 1475921720972416.
Chen, Z., Hu, J., 2005. A comparative study between time-domain method and frequency-
Brusiani, F., De Miranda, S., Patruno, L., Ubertini, F., Vaona, P., 2013. On the evaluation
domain method for identification of bridge flutter derivatives. Eng. Mech. 22,
of bridge deck flutter derivatives using RANS turbulence models. J. Wind Eng. Ind.
127–133 (in Chinese).
Aerod. 119, 39–47.
Chowdhury, A.G., Sarkar, P.P., 2004. Identification of eighteen flutter derivatives of an
Cadenas, E., Rivera, W., 2009. Short term wind speed forecasting in La Venta, Oaxaca,
airfoil and a bridge deck. Wind Struct. 7, 187–202.
Mexico, using artificial neural networks. Renew. Energy 34, 274–278.
Cortes, C., Vapnik, V., 1995. Support-vector networks. Mach. Learn. 20, 273–297. https://
Chen, C.-H., Wu, J.-C., Chen, J.-H., 2008. Prediction of flutter derivatives by artificial
doi.org/10.1007/BF00994018.
neural networks. J. Wind Eng. Ind. Aerod. 96, 1925–1937.
Ding, Q., Chen, A., Xiang, H., 2002. Coupled flutter analysis of long-span bridges by
Chen, N., Ge, Y., 2019. Aerodynamic parameter identification of typical bridge sections
multimode and full-order approaches. J. Wind Eng. Ind. Aerod. 90, 1981–1993.
based on artificial neural network. China Civ. Eng. J. 52, 91–97 (in Chinese).
Friedman, J., Hastie, T., Tibshirani, R., 2001. The Elements of Statistical Learning.
Chen, X., 2007. Improved understanding of bimodal coupled bridge flutter based on
Springer series in statistics, New York.
closed-form solutions. J. Struct. Eng. 133, 22–31.
Gu, M., Zhang, R., Xiang, H., 2000. Identification of flutter derivatives of bridge decks.
J. Wind Eng. Ind. Aerod. 84, 151–162.

14
H. Liao et al. Journal of Wind Engineering & Industrial Aerodynamics 209 (2021) 104493

Hu, G., Kwok, K.C.S., 2020. Predicting wind pressures around circular cylinders using Quinlan, J.R., 1986. Induction of decision trees. Mach. Learn. 1, 81–106. https://doi.org/
machine learning techniques. J. Wind Eng. Ind. Aerod. 198, 104099. 10.1007/BF00116251.
Hu, G., Liu, L., Tao, D., Song, J., Tse, K.T., Kwok, K.C.S., 2020. Deep learning-based Reich, Y., Barai, S.V., 1999. Evaluating machine learning models for engineering
investigation of wind pressures on tall building under interference effects. J. Wind problems. Artif. Intell. Eng. 13, 257–272. https://doi.org/10.1016/S0954-1810(98)
Eng. Ind. Aerod. 201, 104138. https://doi.org/10.1016/j.jweia.2020.104138. 00021-1.
Huang, L., Liao, H., Wang, B., Li, Y., 2009. Numerical simulation for aerodynamic Sarkar, P.P., Caracoglia, L., Haan Jr., F.L., Sato, H., Murakoshi, J., 2009. Comparative and
derivatives of bridge deck. Simulat. Model. Pract. Theor. 17, 719–729. sensitivity study of flutter derivatives of selected bridge deck sections, Part 1: analysis
Jin, X., Cheng, P., Chen, W.-L., Li, H., 2018. Prediction model of velocity field around of inter-laboratory experimental data. Eng. Struct. 31, 158–169.
circular cylinder over various Reynolds numbers by fusion convolutional neural Scanlan, R.H., Jones, N.P., Singh, L., 1997. Inter-relations among flutter derivatives.
networks based on pressure on the cylinder. Phys. Fluids 30, 047105. J. Wind Eng. Ind. Aerod. 69, 829–837.
Jung, S., Ghaboussi, J., Kwon, S.-D., 2004. Estimation of aeroelastic parameters of bridge Selberg, A., 1963. Aerodynamic effects on suspension bridges. In: Proceedings of
decks using neural networks. J. Eng. Mech. 130, 1356–1364. https://doi.org/ International Symposium on Wind Effects on Buildings and Structures,. Proceedings
10.1061/(ASCE)0733-9399(2004)130:11(1356). of International Symposium on Wind Effects on Buildings and Structures,
Kloeppel, K., 1967. Modellversuche im Windkanal zur Bemessung von Brucken gegen die pp. 462–486.
Gefahr winderregter Schwingungen. Stahlbau 36 (12). Tian, J., Gurley, K.R., Diaz, M.T., Fernandez-Caban, P.L., Masters, F.J., Fang, R., 2020.
Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. ImageNet classification with deep Low-rise gable roof buildings pressure prediction using deep neural networks.
convolutional neural networks. In: Pereira, F., Burges, C.J.C., Bottou, L., J. Wind Eng. Ind. Aerod. 196, 104026.
Weinberger, K.Q. (Eds.), Advances in Neural Information Processing Systems, vol. 25. Tubino, F., 2005. Relationships among aerodynamic admittance functions, flutter
Curran Associates, Inc., pp. 1097–1105 derivatives and static coefficients for long-span bridges. J. Wind Eng. Ind. Aerod. 93,
Larsen, A., 1993. Aerodynamic aspects of the final design of the 1624 m suspension 929–950. https://doi.org/10.1016/j.jweia.2005.09.002.
bridge across the Great Belt. J. Wind Eng. Ind. Aerod. 48, 261–285. Van der Put, M., 1976. Rigidity of Structures against Aerodynamic Forces. IABSE, Zurich.
Li, Q., Y, X., 2000. Neural network method for aerodynamic characteristic analysis of Wang, Q., Liao, H., Li, M., Ma, C., 2011. Influence of aerodynamic configuration of a
bridge girder. Central South Highway Engineering 25, 56–59 (in Chinese). streamline box girder on bridge flutter and vortex-induced vibration. Journal of
Li, S., Laima, S., Li, H., 2018. Data-driven modeling of vortex-induced vibration of a long- Modern Transportation 19, 261–267.
span suspension bridge using decision tree learning and support vector regression. Wilms, H., Cupelli, M., Monti, A., Gross, T., 2019. Exploiting spatio-temporal
J. Wind Eng. Ind. Aerod. 172, 196–211. dependencies for RNN-based wind power forecasts. In: 2019 IEEE PES GTD Grand
Liaw, A., Wiener, M., 2002. Classification and regression by randomForest. R. News 2, International Conference and Exposition Asia (GTD Asia). Presented at the 2019 IEEE
18–22. PES GTD Grand International Conference and Exposition Asia (GTD Asia). IEEE,
Lin, S., Wang, Q., Nikitas, N., Liao, H., 2019. Effects of oscillation amplitude on motion- Bangkok, Thailand, pp. 921–926. https://doi.org/10.1109/GTDAsia.2019.8715887.
induced forces for 5:1 rectangular cylinders. J. Wind Eng. Ind. Aerod. 186, 68–83. Wu, B., Chen, X., Wang, Q., Liao, H., Dong, J., 2020a. Characterization of vibration
https://doi.org/10.1016/j.jweia.2019.01.002. amplitude of nonlinear bridge flutter from section model test to full bridge
Liu, D., Niu, D., Wang, H., Fan, L., 2014. Short-term wind speed forecasting using wavelet estimation. J. Wind Eng. Ind. Aerod. 197, 104048.
transform and support vector machines optimized by genetic algorithm. Renew. Wu, B., Wang, Q., Liao, H., Mei, H., 2020b. Hysteresis response of nonlinear flutter of a
Energy 62, 592–597. truss girder: experimental investigations and theoretical predictions. J. Comput.
Lute, V., Upadhyay, A., Singh, K.K., 2009. Support vector machine based aerodynamic Struct. 238, 106267.
analysis of cable stayed bridges. Adv. Eng. Software 40, 830–835. https://doi.org/ Wu, T., Kareem, A., 2011. Modeling hysteretic nonlinear behavior of bridge aerodynamics
10.1016/j.advengsoft.2009.01.008. via cellular automata nested neural network. J. Wind Eng. Ind. Aerod. 99, 378–388.
Matsumoto, M., 1999. Recent study on bluff body aerodynamics and its mechanism. In: Xiang, H., Ge, Y., 2002. Refinements on aerodynamic stability analysis of super long-span
10th ICWE, vol. 1, pp. 67–78. bridges. J. Wind Eng. Ind. Aerod. 90, 1493–1515. https://doi.org/10.1016/S0167-
Matsumoto, M., 1996. Aerodynamic damping of prisms. J. Wind Eng. Ind. Aerod. 59, 6105(02)00266-0. Fifth Asia-Pacific Conference on Wind Engineering.
159–175. https://doi.org/10.1016/0167-6105(96)00005-0. Meeting on Structural Xiang, H.F., Chen, A.R., Lin, Z.X., 1998. An introduction to the Chinese wind-resistant
Damping International Wind Engineering Forum and Additional Papers. design guideline for highway bridges. J. Wind Eng. Ind. Aerod. 74, 903–911.
Matsumoto, M., Daito, Y., Yoshizumi, F., Ichikawa, Y., Yabutani, T., 1997. Torsional Xu, F., Zhang, Z., 2017. Free vibration numerical simulation technique for extracting
flutter of bluff bodies. J. Wind Eng. Ind. Aerod. 69, 871–882. flutter derivatives of bridge decks. J. Wind Eng. Ind. Aerod. 170, 226–237.
Matsumoto, M., Kobayashi, Y., Shirato, H., 1996. The influence of aerodynamic Yang, Y., Ge, Y., Xiang, H., 2007. Investigation on flutter mechanism of long-span bridges
derivatives on flutter. J. Wind Eng. Ind. Aerod. 60, 227–239. https://doi.org/ with 2d-3DOF method. Wind Struct. 10, 421–435.
10.1016/0167-6105(96)00036-0. The Wind Engineering Society’s 2nd UK Yu, C., Li, Y., Xiang, H., Zhang, M., 2018. Data mining-assisted short-term wind speed
Conference. forecasting by wavelet packet decomposition and Elman neural network. J. Wind
Matsumoto, M., Matsumiya, H., Fujiwara, S., Ito, Y., 2010. New consideration on flutter Eng. Ind. Aerod. 175, 136–143. https://doi.org/10.1016/j.jweia.2018.01.020.
properties based on step-by-step analysis. J. Wind Eng. Ind. Aerod. 98, 429–437. Zhou, J., Shi, J., Li, G., 2011. Fine tuning support vector machines for short-term wind
Mei, H., Wang, Q., Liao, H., Fu, H., 2020. Improvement of flutter performance of a speed forecasting. Energy Convers. Manag. 52, 1990–1998.
streamlined box girder by using an upper central stabilizer. J. Bridge Eng. https:// Zhou, R., Ge, Y., Yang, Y., Liu, S., Du, Y., Zhang, L., 2019. A nonlinear numerical scheme
doi.org/10.1061/(ASCE)BE.1943-5592.0001584. to simulate multiple wind effects on twin-box girder suspension bridges. Eng. Struct.
Ministry of Communications of the People’s Republic of China, 2004. Wind-resistant 183, 1072–1090.
Design Specification for Highway Bridges. JTG/T D60-01-2004. China Zhu, Z., Gu, M., Chen, Z., 2007. Wind tunnel and CFD study on identification of flutter
Communications Press, Beijing (in Chinese). derivatives of a long-span self-anchored suspension bridge. Comput. Aided Civ.
Quinlan, J.R., 2014. C4.5: Programs for Machine Learning. Elsevier. Infrastruct. Eng. 22, 541–554.

15