1. Introduction
Due to their high flexibility and low structural damping, long-span cable-supported bridges are susceptible to a variety of wind-induced vibrations [
1,
2,
3,
4]. Wind buffeting and vortex-induced vibration are two typical wind-induced vibrations. Wind buffeting is caused by natural fluctuation of wind flow, and it is usually a broad-band, small amplitude vibration. On the other hand, vortex-induced vibration (VIV) is a kind of resonant vibration excited by periodic vortex shedding and it is single mode, large-amplitude vibration. VIV has been observed occasionally on several long-span bridges, such as Volgograd continuous bridge with steel box girder in Russia [
5], the Great Belt East suspension bridge in Denmark [
6,
7], the Trans-Tokyo Bay Crossing Bridge in Japan [
8], the Second Severn Bridge in UK [
9], the Yi Sun-sin suspension bridge in South Korea [
10], the Xihoumen Bridge in China [
11], among others. As VIV usually happens at low wind velocity of 6–12 m/s, it can result in discomfort to users and fatigue problem to structures.
VIV of bluff body is a complex aeroelastic phenomenon that is caused by nonlinear fluid and structure interaction. When the airflow passes through the bluff body like a bridge deck, periodic vortexes sheds at either side of the body. The vortex shedding frequency, fs, of the flow is generally proportional to the wind velocity as described by the Strouhal number, St = fsD/U. At a certain wind velocity, that is, when the vortex shedding frequency is close to one of the natural frequencies of the structure, fn, resonant self-excited vibration occurs if the structural damping is sufficiently low. Once the bridge decks are set into resonant vibrations, the vortex-shedding frequency of the flow becomes locked onto the vibration frequency of bridge decks, such that vortex-induced vibrations sustain over a range of wind velocity. Due to the structure–fluid interaction, the VIV is highly nonlinear as excitation forces due to vortex shedding are difficult to predict. At present, most of the researches on VIV are based on wind tunnel test methods, focusing on influencing factors of VIV, characteristics and modeling of excitation forces, and searching for effective measures to suppress vibration, among others. However due to inadequacy in modelling of Reynolds number, flow turbulence, and three-dimensional effect, wind tunnel tests sometimes fail to predict the full-scale aeroelastic responses of VIV of prototype bridges.
Some field measurements have been made to investigate the characteristics of VIV, and its correlation with wind conditions. Frandsen [
6] monitored the Great Belt East Bridge and employed the fast Fourier transform spectral analyses to identify the VIV from field measured data of the Great Belt East Bridge. The VIV with amplitudes up to 30 cm occurred at wind velocity about 4–12 m/s, while the wind direction is nearly perpendicular to bridge axis. The wind pressures and accelerations were measured simultaneously in that article and it was observed that pressures and accelerations were highly correlated at lock-in, but once out of this regime they become uncorrelated. The increase of turbulence intensity can reduce the correlation. Similar observations have been reported by Larsen et al. [
7]. Fujino [
8] investigated the VIV measured in Trans-Tokyo Bay Crossing Bridge and found that the first mode can only be observed when the wind direction is within ±20° of the bridge transverse axis. The maximum vertical amplitude can reach 50cm during the VIV, so tuned mass dampers (TMD) and vertical plates were implemented for the first two modes and higher modes, respectively. The result of wind tunnel tests agreed with the phenomenon of field observations. Li et al. [
11] carried out the research about the Xihoumen Bridge. During the monitoring period, thirty-seven VIV events were observed and the stochastic subspace identification method was used to obtain the modal parameters from the measured data. It was found that the inhomogeneity of wind field along the bridge axis can affect VIV as an important factor. Cantero et al. [
12] investigated the monitoring data of Hardanger Bridge, which is currently the longest suspension bridge in Norway. It found that the vortex shedding frequency of hangers is correlated to the 60 s mean wind velocity, and the possibility of identifying the VIV of hangers by analyzing the acceleration signal of deck is demonstrated. Kim et al. [
13] investigated the relationship between the amplitude of the VIV and the identified damping ratios based on the monitoring data of Jindo bridge. The multiple tuned mass dampers (MTMD) was then installed in the main span of the bridge to mitigate the vibration, and it was found that the effect of MTMD is gradually obvious when the wind velocity approaches the onset wind velocity of the VIV.
Long-term structural health monitoring has increasingly become an important technique for health monitoring and condition assessment of bridges [
14,
15,
16,
17,
18,
19,
20]. It records a large amount of data of bridges when subjected to vortex-induced vibrations; therefore, providing a powerful tool for verifying wind tunnel tests and further understanding of VIV. However, as VIV occurs only at special wind conditions, the majority of recorded acceleration responses are random vibrations caused by wind buffeting, passing vehicles, and other environmental effects. Only a few of them are vortex-induced vibrations due to periodic vortex shedding at specified wind conditions. In the previous studies [
6,
7,
8,
9,
10], the VIV was selected out mainly by visual inspection of measured data. How to identify the VIV signal from the massive monitoring data, without manual intervention, becomes an important problem to be addressed. Wind buffeting is a multi-modal random vibration, while VIV is a single-mode, nearly harmonic vibration. Therefore, there are significant differences between the VIV and random vibration. Automatic identification aims at distinguishing between those two modes, and can be cast into the machine learning framework, such as classical classification problem or outlier detection. In the context of structural damage detection, a great amount of work has been conducted by using supervised and unsupervised learning algorithms for discriminating the damage state from the intact structure [
16,
18,
21,
22]. Fugate et al. developed a damage detection algorithm that is based on statistical analysis of residual errors of fitted auto-regression models [
16]. Worden et al. suggested the damage detection to be addressed in the framework of outlier detection, especially in the case of low-level damage detection [
21]. Rogers et al. developed a Bayesian non-parametric clustering approach for damage detection [
23]. Unlike damage detection, VIV happens only at some specific wind conditions, and; therefore, studies addressing the automatic detection of VIV are rather limited. Li et al. [
24] proposed a technique based on cluster analysis to identify VIV events from long-term monitoring data, where the power spectral density of acceleration response is taken as feature of cluster analysis, and that method is applied to the Xihoumen Bridge. Later they developed a decision tree method to classify the mode branch of VIV [
25]. Their techniques are capable of identifying VIV events but also can discriminate different modes of vibration. However, in above studies the power spectral density of the acceleration response was taken as input, which may require a certain degree of manual intervention. Therefore, signal-based automatic identification of VIV events is an important and meaningful task.
The aim of this study is to determine the character of the response measured from a long-term monitoring system, that is, vortex-induced vibration or random vibration (wind buffeting or traffic-induced vibration). Random decrement technique (RDT) is an important tool for analyzing the random vibration signal [
26,
27]. It is well known that the RDT-processed random signal can be regarded as kind of free vibration data. As will be shown later, the VIV signal processed with RDT is a kind of sinusoidal response with nearly constant amplitude. In recognition of the difference, this paper proposes an automatic identification method for VIV based on random decrement technique. The coefficient of variation (COV) of the peak values of the processed signals is defined as the extracted feature, and the hypothesis of Gaussian distribution is employed to establish the threshold according to the Pauta criterion. The organization of the paper is as follows: In
Section 2, the RDT-based automatic identification method of VIV is presented. In
Section 3, numerical validation of the proposed method is carried out by using simulated response of VIV and random vibration. In
Section 4, the proposed method is applied to three-month monitoring data of Xihoumen Bridge in China for identifying the VIV, and the dynamic characteristics of VIV are also described.
2. Presentation of Method
2.1. Overview of Proposed Method
Figure 1 showed the flowchart of the proposed method for automatic identification of VIV from massive long-term monitoring data. In general, two steps are involved in this method. The first step is the establishment of a threshold of output, COV in this study, using random vibration data. In the second step, when the COV of new data fall below the threshold, the data will be identified as vortex-induced vibrations; otherwise it will be classified as random vibration.
In the first step, the raw acceleration data is processed with filtering to remove the undesired frequency components; a banded-pass filtering of 0.07–0.7 Hz is employed. The filtered data is processed by random decrement technique to obtain the random decrement signature from which the threshold is established.
2.2. Character of VIV and Random Vibration
A bridge at operation stage is usually subjected to ambient dynamic loadings, such as wind and traffic. When the dynamic loadings are random, the induced response will also be random, such as those caused by wind buffeting and passing vehicles. Wind buffeting is a kind of forced vibration caused by natural fluctuation of incoming wind. In some circumstances, the dynamic loading can be caused by periodic excitation such as vortex-shedding forces, and the induced response becomes periodic or even sinusoidal. Wind buffeting is usually not a serious concern due to its small amplitude at common wind velocity, while vortex-induced vibration can cause user discomfort and fatigue problems due to its large vibration amplitude up to 50 cm. The theoretical background of wind buffeting and vortex-induced vibration may be found in Simiu and Scanlan [
1].
The majority of measured structural responses of a bridge at normal operation stage are random responses induced by wind buffeting and passing vehicles. Due to the broad-banded nature of excitation, random response is characterized by multimodal vibration, indicating a number of predominant frequency components in the power spectrum of random vibration. On the other hand, vortex-induced vibration is a kind of resonant vibration that is caused by the periodic shedding and its resulting sinusoidal forces. Therefore, VIV is characterized as single mode vibration for a specific wind velocity. Several distinct differences between random vibration and VIV are highlighted:
- (1)
Random vibration is caused by the frequent random excitation, such as wind buffeting and passing vehicles. It is always occurring, regardless of the magnitude of wind velocity, while VIV occurs only at a specific wind velocity.
- (2)
Random vibration is characterized by multiple mode responses, while VIV only has a single predominant frequency in the power spectrum.
Figure 2 showed the two typical segments from the health monitoring system of the Xihoumen Bridge, which corresponds to random vibration and VIV. It is obvious that the VIV signal is close to the harmonic signal, and is usually considerably larger than wind buffeting.
2.3. Random Decrement Technique
In this section, the background of RDT was briefly introduced. The random decrement technique is a time-domain method that uses the ensemble average of ambient data to approximate the free vibration responses when structures are exposed to linear, Gaussian white noise excitation. It was initially proposed by Cole in 1973 [
28,
29]. Later in 1982, Vandiver provided a strict mathematical proof for a single-degree-of-freedom system under the assumption of input to be a zero-mean, stationary, and Gaussian random process [
27]. The technique can also be extended to multi-degrees-of-freedom systems naturally, and is identified as the multimode random decrement technique (MRDT) [
30,
31].
For a linear dynamical model, the forced vibration response,
y(
t), at a measurement point of a structure under any linear, Gaussian white noise excitation,
f(
t), can be expressed as:
where
D(
t) is free vibration response with an initial displacement of 1 and an initial velocity of 0;
V(
t) is free vibration response with an initial displacement of 0 and an initial velocity of 1;
y(0) and
(0) are initial displacement and velocity of the vibration system, respectively;
h(
t) is unit impulse response function of a viscously damped single-degree-of-freedom system;
f(
t) is dynamic loading; and
ζ,
ω, and
ωd are modal damping ratio, natural frequency, and damped frequency, respectively.
As shown in Equation (1), the response y(t) in each time instant t can be decomposed as a linear superposition of three parts: The first two parts in the right hand side of Equation (1) represent the free vibration response due to initial displacement and initial velocity, and the third part is the forced vibration response due to dynamic loading, f(t). When f(t) follows a Gaussian white noise process, the responses, y(t), will also obey a Gaussian white noise process.
Figure 3 shows the overview of random decrement technique. By selecting an appropriate constant
A to intercept a measured random vibration response signal,
y(
t), a series of sub-response process,
y(
t −
ti), can be obtained as follow:
Since the
f(
t) is a zero-mean stationary Gaussian process, the starting point of time does not affect its random characteristics. By moving a series of time starting points
ti of
y(
t −
ti) to the origin of coordinate, a series of sample function,
xi(
t) (
I = 1, 2, …,
N), of random processes can be obtained:
Because
f(
t) and
are zero-mean stationary Gaussian processes, the expectation of a sub- excitation process,
E[
f(
t)], and the expectation of the sub-response process,
E[
], are zero. The mathematical expectation of
xi(
t) (
I = 1, 2, …,
N) is obtained as follows:
In practice, due to the limited measurement time and data, the average of
x(
t) is used to replace the mathematical expectation:
Therefore, a damped free vibration response with an initial displacement of A and an initial velocity of 0 is obtained.
2.4. Analysis of VIV Signal by RDT
The buffeting of the bridge is a kind of random vibration caused by turbulent flow, and the pulsating wind component in the atmosphere is the main contributing factor. Since pulsating wind is a stationary random excitation with a mean of zero, the response of the structure is also a stationary random response. It is therefore obvious that the buffeting response processed with RDT reduces to a kind of damped free-decaying response. In contrast, vortex-induced vibration is nearly a sinusoidal response with one dominant frequency, and VIV processed with RDT will become a harmonic response with nearly constant amplitude, as shown below.
The vortex-induced vibration of the bridge is a kind of wind-induced limiting vibration with both forced and self-excited properties. VIV events can be observed at a certain wind speed, and the measured vibration signal is close to the harmonic signal. Therefore, the harmonic response at a certain measurement point can be described as follows,
where
Y is amplitude of the vibration;
ω is circle frequency of the vibration, which is close to one of structural modal frequencies;
θ is phase angle, which is a phase lag between force and structural response;
η(
t) is Gaussian white noise signal, which is a zero-mean stationary Gaussian process.
Likewise, the RDT is applied to vortex-induced vibration given by Equation (8). By selecting a constant
A to intercept the signal, a series of segments can be obtained. The subsample function is given as follow:
Two subsamples of segments are taken as example for illustration, as follows
Similarly, ensemble average of
yi(
t) is given as follows:
It can be seen from the above equation that the VIV data processed with RDT becomes a harmonic signal.
2.5. Etablishment of Threshold-Discrimiating VIV and Random Vibration
As seen in Equations (7) and (11), there is a significant difference between the results of random vibration and VIV obtained with RDT. Specifically, the random vibration reduces to a kind of damped free vibration, while the VIV become a harmonic vibration.
Figure 4 showed the obtained result of the acceleration data previously shown in
Figure 2. It is verified that the result obtained from the random vibration is a damped free vibration, with the initial amplitude of 5 × 10
−3 m/s
2. As the length of the segment is limited, the number of the subsample that intercepted from the segment is limited, which makes the signal different from an ideal free vibration. By contrast, the amplitude of the VIV signal after the random decrement process changes little with time, remaining at about 5 × 10
−2 m/s
2.
In order to distinguish the two kinds of responses given in
Figure 4, the COV of peak values of the processed signals is used. First, the peak value of processed data is selected, and the COV of peak value is calculated for random vibration and for VIV. For random vibration signals, the COV of the random decrement signal will be significantly larger than that of VIV signal.
The calculation formula for COV is:
In order to determine the characteristic value of COV for random vibration, a large number of COV of the random vibration signal should be used. The mean value
uc and the standard deviation
σc are obtained from large data set of COV. In this paper, the Pauta criterion was considered when calculating the threshold for determining the VIV, the confidence interval was chosen as 99.99%:
When the COV is greater than the above threshold dc, the signal is determined as random vibration, and when the COV is less than the threshold, it is determined as a VIV. The main steps of the proposed method are summarized as follows:
- (1)
Selecting the cutoff value A and the length of the output signal in the segment;
- (2)
Performing the random decrement method to obtain the processed signal;
- (3)
Extracting the peak value of the processed signal to obtain the COV of each signal segment;
- (4)
Selecting a certain number of COV of the random vibration signal to obtain a threshold value;
- (5)
Calculate the COV using steps (1)–(4) for a new data set. When the COV for the new data set is less than the threshold, it is identified as VIV.