TCN network-based large-span bridge wind-induced vibration response prediction system and TCN network-based large-span bridge wind-induced vibration response prediction method

Abstract

The invention relates to a TCN network-based large-span bridge wind-induced vibration response prediction system and method. The TCN network-based large-span bridge wind-induced vibration response prediction system comprises a front-end data acquisition and transmission subsystem, a later-stage data processing and calculation subsystem and a bridge wind-induced vibration response prediction subsystem. Firstly, a front-end data acquisition and transmission subsystem continuously monitors wind environment data and bridge vibration response data at a bridge site in real time and transmits the data to a database; then, the later data processing and calculating subsystem preprocesses the measured data, calculates measured wind characteristic data and bridge vibration response characteristic data, constructs a data set by taking the wind characteristic data as an input characteristic and taking the vibration response characteristic as an output characteristic, and performs characteristic data correlation analysis; and finally, the bridge wind-induced vibration response prediction subsystem is used for training a TCN deep learning network according to the data set after correlation analysis and improvement, and carrying out bridge wind-induced vibration response prediction based on the TCN network. According to the method, the wind vibration response of the bridge can be effectively predicted, and information reference is provided for bridge construction operation decision making.

Description

Technical Field

[0001] The present invention belongs to the technical field of bridge wind engineering, and specifically relates to a system and method for predicting wind-induced vibration response of a long-span bridge based on a TCN network. Background Art

[0002] In recent years, with the rapid economic development, my country's transportation infrastructure construction is booming. As an important infrastructure, long-span bridges are widely used, such as the Sutong Bridge, the Hutong Yangtze River Bridge, the Hong Kong-Zhuhai-Macao Bridge, etc. However, long-span bridges are highly flexible structures that are easily affected by wind loads. The location of the bridge site is mostly a high-risk area for wind resistance, so the bridge is very likely to be damaged by wind. Accurately predicting the wind-induced vibration response of long-span bridges can help construction personnel take reasonable countermeasures in a timely manner to avoid possible economic losses.

[0003] At present, there are two main methods for predicting bridge vibration response: statistical model prediction and data-driven prediction. The statistical model prediction method is a method based on mathematical principles and data analysis technology to predict the trend or result of future events, and is widely used in the field of bridge wind vibration prediction. However, the cross-section forms of bridges in different engineering practices are diverse, and the complex changes in wind field flows under different construction environments make it difficult to accurately quantify parameters such as bridge aerodynamic admittance functions and flutter derivatives. Moreover, the mathematical model is complex in form and the prediction model is not versatile, making it difficult to meet the prediction needs of variable-section cantilever bridges. In recent years, with the rapid development of machine learning technology, data-driven methods have provided new ideas for bridge wind vibration prediction based on structural dynamic response analysis.

[0004] In response to the above problems, the present invention proposes a large-span bridge wind-induced vibration response prediction system and method based on the TCN network, which integrates the parallel characteristics of convolutional neural networks and the time series modeling capabilities of recursive neural networks, and can better explore the abstract characteristics of bridge vibration, thereby achieving more accurate and efficient bridge wind-induced vibration response prediction. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] The purpose of the present invention is to propose a system and method for predicting the wind-induced vibration response of a large-span bridge based on a TCN network, to use deep learning network technology to realize a prediction model for the bridge vibration response, to perform a correlation analysis on the input wind characteristics and the output bridge vibration response characteristics, to determine the main associated parameters of the bridge wind-induced vibration response, to remove the input wind characteristics with poor correlation, to improve the training and prediction efficiency of the model, to use the improved data set to train the TCN network model, to realize the prediction of the bridge wind-induced vibration response, and to provide information reference for bridge construction and operation decisions.

[0007] The technical solution proposed by the present invention is as follows:

[0008] A wind-induced vibration response prediction system for a large-span bridge based on a TCN network includes a front-end data acquisition and transmission subsystem, a post-data processing and calculation subsystem, and a bridge wind-induced vibration response prediction subsystem; wherein:

[0009] The front-end data acquisition and transmission subsystem is a bridge monitoring system composed of multiple three-dimensional anemometers and acceleration sensors, which can continuously monitor the wind environment data and bridge vibration response data at the bridge site in real time and transmit them to the database;

[0010] The post-data processing and calculation subsystem can pre-process the wind environment data at the bridge site and the bridge vibration response data in the database, calculate the corresponding wind characteristic data and bridge vibration response characteristics, and perform wind vibration response analysis on the measured bridge;

[0011] The bridge wind-induced vibration response prediction subsystem is constructed based on the TCN network prediction model, with the wind characteristic data obtained in the later data processing and calculation subsystem as input, and the bridge vibration response data as output to train the TCN algorithm, and derive the wind-induced vibration response prediction data to provide information reference for bridge construction and operation decision-making;

[0012] The bridge wind-induced vibration response prediction subsystem can evaluate the prediction effect, compare it with the actual measured results of the bridge wind-induced vibration response during the prediction period, and provide feedback to update the TCN network prediction model parameters.

[0013] Preferably, the post-data processing and calculation subsystem processes the wind environment data at the original bridge site with a basic time interval of 10 minutes to obtain measured wind characteristics, including: average wind speed, wind direction angle, wind attack angle, turbulence intensity, turbulence integral scale, and pulsating wind power spectrum density.

[0014] The calculation formula of the wind characteristics is as follows:

[0015]

[0016]

[0017] u=xcos(θ)+ysin(θ)-U (4)

[0018] v=-xsin(θ)+ycos(θ) (5)

[0019]

[0020] I i =σ i / U (i=u,v,w) (7)

[0021]

[0022] In the formula, x, y, and z are the wind speeds along the bridge, across the bridge, and vertically, respectively; are the average wind speeds along the bridge, across the bridge and vertically in the basic time interval; U, θ, α are the average wind speed, wind direction angle and wind attack angle in the basic time interval; u, v, w are the along wind direction, across wind direction and vertical pulsating wind respectively; I i is the turbulence degree in each wind direction; L i is the turbulence integral scale of each wind direction; S i (f) is the power spectrum density of fluctuating wind in each wind direction.

[0023] Preferably, the post-processing data and calculation subsystem, when processing the original bridge vibration response data, uses a second-order Butterworth-based low-pass filter to remove unnecessary high-frequency noise, subtracts the signal mean to avoid drift, and converts the vibration time domain signal into a frequency domain signal through a fast Fourier transform (FFT) to obtain the bridge modal frequency and damping ratio.

[0024] The damping ratio ζ can be calculated by the logarithmic decay rate of the free decay vibration, assuming that the two adjacent peaks are x1 and x2:

[0025]

[0026] Damping ratio calculation formula:

[0027]

[0028] Preferably, the post-data processing and calculation subsystem can use the measured wind characteristic data as input features and the measured bridge vibration response characteristics as output features, construct a TCN machine learning data set, analyze the correlation between wind characteristics and vibration response characteristics, and determine the main correlation parameters of the bridge wind vibration response, so as to minimize the redundant input variables of subsequent neural network modeling and improve the training and prediction efficiency of the model.

[0029] Preferably, the bridge wind-vibration response prediction subsystem uses the wind characteristics with high correlation after analysis by the later data processing and calculation subsystem as input vectors, and the longitudinal, transverse, vertical and torsional accelerations of the bridge as output vectors, and predicts bridge vibrations under different environments based on the TCN deep learning network.

[0030] Preferably, the bridge wind-vibration response prediction subsystem uses three evaluation indicators, namely mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE), to achieve a comprehensive evaluation of the prediction ability of the TCN model, compares the results with the measured results of the bridge wind-vibration response during the prediction period, and provides feedback to update the TCN network prediction model parameters.

[0031] The calculation formulas for each evaluation index are as follows:

[0032]

[0033] Among them, y t represents the measured value of acceleration at time t; Represents the predicted value of acceleration at time t; N is the total number of samples.

[0034] The present invention proposes a method for predicting wind-induced vibration response of a long-span bridge based on a TCN network, which is characterized by comprising the following steps:

[0035] S1, the front-end data acquisition and transmission subsystem collects wind environment data and bridge vibration response data at the bridge site and transmits them to the database;

[0036] S2, the post-processing data and calculation subsystem analyze and process the wind environment measured data and bridge wind vibration response measured data

[0037] S2-1, call the original wind environment measured data in the database and process it with 10 minutes as the basic time interval to obtain the measured wind characteristics, including: average wind speed, wind direction angle, wind attack angle, turbulence intensity, turbulence integral scale, and fluctuating wind power spectrum density;

[0038] S2-2, call the measured data of wind-induced vibration response of bridges in the database, use a low-pass filter based on second-order Butterworth to remove unnecessary high-frequency noise, subtract the signal mean to avoid drift, and convert the vibration time domain signal into a frequency domain signal through fast Fourier transform (FFT) to obtain the bridge modal frequency and damping ratio;

[0039] S2-3, using the measured wind characteristic data as input features and the measured bridge vibration response characteristics as output features, a TCN machine learning data set was constructed to analyze the correlation between wind characteristics and vibration response characteristics, and determine the main correlation parameters of the bridge wind vibration response, so as to minimize the redundant input variables of the subsequent neural network modeling and improve the training and prediction efficiency of the model;

[0040] S3, the bridge wind-induced vibration response prediction subsystem realizes the bridge wind-induced vibration response prediction based on the TCN network and compares it with the measured data to update the improved algorithm;

[0041] S3-1, based on the TCN deep learning network, the wind characteristics with high correlation after the later data processing and calculation subsystem analysis are used as input vectors, and the longitudinal, transverse, vertical and torsional accelerations of the bridge are used as output vectors to predict the bridge vibration under different environments;

[0042] S3-2, using the three evaluation indicators of mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) to realize the comprehensive evaluation of the prediction ability of the TCN model, and comparing it with the measured results of the wind-induced vibration response of the bridge during the prediction period, and providing feedback to update the parameters of the TCN network prediction model.

[0043] In the preferred step S2, the calculation formula of the wind characteristics is as follows:

[0044]

[0045]

[0046] u=xcos(θ)+ysin(θ)-U (4)

[0047] v=-xsin(θ)+ycos(θ) (5)

[0048]

[0049] I i =σ i / U (i=u,v,w) (7)

[0050]

[0051] In the formula, x, y, and z are the wind speeds along the bridge, across the bridge, and vertically, respectively; are the average wind speeds along the bridge, across the bridge and vertically in the basic time interval; U, θ, α are the average wind speed, wind direction angle and wind attack angle in the basic time interval; u, v, w are the along wind direction, across wind direction and vertical pulsating wind respectively; I i is the turbulence degree in each wind direction; L i is the turbulence integral scale of each wind direction; S i (f) is the power spectrum density of fluctuating wind in each wind direction.

[0052] Preferably, the damping ratio ζ in the preferred S2 step can be calculated by the logarithmic decay rate of the free decay vibration, assuming that two adjacent peaks are x1 and x2:

[0053]

[0054] Damping ratio calculation formula:

[0055]

[0056] Preferably, in step S2, features whose absolute values ​​of the calculated wind features and bridge vibration response features are less than 0.3 will not be used as input features of the machine learning data set.

[0057] Preferably, the calculation formulas for the evaluation indicators in step S3 are as follows:

[0058]

[0059] Among them, y t represents the measured value of acceleration at time t; Represents the predicted value of acceleration at time t; N is the total number of samples.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] A large-span bridge wind-induced vibration response prediction method model based on TCN network proposed in the present invention shows good prediction effect in wind-induced vibration prediction in different directions of the bridge with the help of the special structure of TCN deep learning algorithm. It can realize the wind-induced vibration response prediction in areas with large wind speed changes in the short term, and provide information reference for bridge construction and operation decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0063] Figure 1 Flow chart of the wind-induced vibration response prediction method for long-span bridges based on TCN network

[0064] Figure 2(a) shows the structure of the TCN deep learning model

[0065] Figure 2(b) is a schematic diagram of the one-dimensional causal convolution principle

[0066] Figure 2(c) is the principle diagram of dilated convolution

[0067] Figure 2(d) shows the residual connection structure

[0068] Figure 3 Layout diagram for a rigid frame bridge

[0069] Figure 4 The measured wind speed time history curve

[0070] Figure 5 This is the measured bridge vibration response time history curve

[0071] Figure 6 This is the correlation analysis diagram between wind characteristics and bridge vibration response characteristics

[0072] Figure 7(a) shows the prediction effect of different models on the acceleration along the bridge.

[0073] Figure 7(b) shows the prediction effect of different models on the transverse acceleration of the bridge

[0074] Figure 7(c) shows the prediction effect of different models on vertical acceleration

[0075] Figure 7(d) shows the prediction effect of different models on torsional acceleration; DETAILED DESCRIPTION

[0076] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0077] Example 1

[0078] The present invention provides a system and method for predicting wind-induced vibration response of a long-span bridge based on a TCN network. Figure 1 As shown, a wind-induced vibration response prediction system for a large-span bridge based on a TCN network includes three parts: a front-end data acquisition and transmission subsystem, a later data processing and calculation subsystem, and a bridge wind-induced vibration response prediction subsystem. First, the front-end data acquisition and transmission subsystem continuously monitors the wind environment data and bridge vibration response data at the bridge site in real time and transmits them to the database; the later data processing and calculation subsystem calls the measured data in the database for preprocessing, and calculates the measured wind characteristic data and bridge vibration response characteristic data to analyze the measured wind-induced vibration response of the bridge; finally, the bridge wind-induced vibration response prediction subsystem uses the wind characteristic data as input features and the vibration response features as output features to construct a data set, trains the TCN deep learning network to predict the wind-induced vibration response of the bridge based on the TCN network, and derives the prediction results, while evaluating the prediction effect, and performing feedback comparison with the measured data of the bridge wind-induced vibration response to improve the algorithm; the present invention can effectively predict the wind-induced vibration response of the bridge and provide information reference for bridge construction and operation decisions.

[0079] A method for predicting wind-induced vibration response of a long-span bridge based on a TCN network comprises the following steps:

[0080] Step 1: Build a front-end data acquisition and transmission subsystem, use multiple three-dimensional anemometers and accelerometers to arrange a large-span bridge wind vibration response monitoring system, collect wind environment data and bridge vibration response data at the bridge site, and transmit them to the database;

[0081] Step 2: The post-data processing and calculation subsystem calls the original wind environment measured data in the database, and uses the pycharm platform to process the measured wind characteristics with a basic time interval of 10 minutes. The average wind speed, wind direction angle, and wind attack angle are calculated according to formulas (1) to (3), and the fluctuating wind in each direction is calculated according to formulas (4) to (6). The turbulence degree, turbulence integral scale, and fluctuating wind power spectrum density in each direction are calculated according to formulas (7) to (9);

[0082] Step 3: The post-processing and calculation subsystem calls the measured data of bridge wind-induced vibration response in the database and is programmed using the pycharm platform. The second-order Butterworth low-pass filter is used to remove unnecessary high-frequency noise, and the signal mean is subtracted to avoid drift. The vibration time domain signal is converted into a frequency domain signal through fast Fourier transform (FFT) to obtain the bridge modal frequency and damping ratio. The bridge damping ratio is calculated according to formulas (10) and (11).

[0083] Step 4: The post-processing and calculation subsystem is programmed on the pycharm platform. The measured wind characteristic data is used as the input feature, the measured bridge vibration response characteristics are used as the output feature, and the data volume of 100 times the basic time interval is used to construct the TCN machine learning data set. The correlation between the wind characteristics and the vibration response characteristics is analyzed, and the main correlation parameters of the bridge wind vibration response are determined to minimize the redundant input variables of the subsequent neural network modeling, remove the input wind characteristics with the absolute value of the correlation less than 0.3, and improve the training and prediction efficiency of the model;

[0084] Step 5, as shown in Figure 2(a), the bridge wind-induced vibration response prediction subsystem establishes a bridge wind-induced vibration response prediction model based on the TCN network.

[0085] In multivariate wind speed prediction, multiple factors that affect wind speed need to be considered, each of which has its own unique physical properties and units of measurement. The TCN network model first normalizes the data to ensure the accuracy and stability of the model. Through normalization, different variables are unified into a similar scale range to avoid certain variables having a greater impact on model training. Normalization helps to simplify the model's data processing flow, thereby enhancing the model's processing efficiency and accelerating convergence.

[0086] The TCN network model uses the minimum-maximum normalization technique for the data, which converts the initial data into the range of 0 to 1 through linear transformation.

[0087] The normalized calculation method is as follows:

[0088]

[0089] Among them, x is the original data, x' represents the normalized data, and x min is the minimum value of all data in the variable to which x belongs. max It is the maximum value of all data in the variable to which x belongs.

[0090] The TCN network model establishes a causal relationship between input and output through causal convolution. The schematic diagram of one-dimensional causal convolution is shown in Figure 2(b). The predicted value y at a certain time t is t Only with the data x at time t t and the data before it (x0, x1, L, xt-1 ), that is, the current information is not affected by future information. At the same time, causal convolution performs zero padding at the beginning of the data sequence to ensure that the input and output lengths are equal. Specifically, suppose there is an input sequence X = (x0, x1, L, x t-1 ) and the output sequence Y = (y0, y1, L, y t ),H=(h0,h1,L,h k ), then in x t The causal convolution result at can be expressed as:

[0091]

[0092] Among them, * represents the causal convolution operation, h k is the weight of the convolution kernel, and K is the size of the convolution kernel.

[0093] Its receptive field can be expressed as:

[0094] RF=K+(K-1)(L-1) (17)

[0095] Among them, RF is the receptive field size, and L is the total number of network layers.

[0096] In order to capture information of a longer time scale, the TCN network model is expanded on the basis of causal convolution. Its main principle is shown in Figure 2(c). By introducing a dilation rate in the convolution kernel, the input data of the convolution is sampled at intervals to expand the receptive field. t The dilated convolution can be expressed as:

[0097]

[0098] Among them, the expansion rate d usually increases exponentially with the increase of the depth of the network, specifically:

[0099] d=2 i (19)

[0100] Among them, i is the number of network layers.

[0101] The size of the receptive field after expansion can be expressed as:

[0102] RF=K+(K-1)(L-1)(D-1) (20)

[0103] Where D is the dilation rate of the last layer. The receptive field can be flexibly adjusted by changing the size of the convolution kernel, the number of network layers, and the dilation rate.

[0104] While the TCN model expands the receptive field by increasing the number of hidden layers, a network that is too deep will affect the stability of model training, resulting in problems such as gradient vanishing and gradient explosion. For this reason, residual connections are introduced into the TCN model, which use residual blocks to increase the depth of the network layer while ensuring its accuracy. Its specific structure is shown in Figure 2(d). With the help of residual connections, the model can learn the changes in the input sequence, and can more easily transfer gradients and alleviate problems such as gradient vanishing. Its formula is expressed as:

[0105] y=x+F(x) (21)

[0106] in, F(x) is the output of the network layer, x is the input, and y is the final output.

[0107] The advantage of residual connection is that it can directly add cross-layer connections to allow the TCN network to learn residuals, that is, residual information that the model needs to learn. This helps the TCN model to transfer information more effectively in deep networks and long sequence modeling, so that it can better capture long-term dependencies in sequences in actual time series tasks.

[0108] Step 6: Based on the TCN deep learning network, the wind characteristics with high correlation after the later data processing and calculation subsystem analysis are used as input vectors, and the longitudinal, transverse, vertical and torsional accelerations of the bridge are used as output vectors to predict the bridge vibration under different environments;

[0109] Step seven, the three evaluation indicators of mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are used to comprehensively evaluate the prediction ability of the TCN model, and the results are compared with the measured results of wind-induced vibration response of bridges during the prediction period, and the TCN network prediction model parameters are updated as feedback.

[0110] Example verification:

[0111] The actual data collection of a large-span rigid frame bridge with a very large high pier was selected to verify the accuracy of the wind-induced vibration response prediction of each model. The total length of the bridge is 2292m, the width of the bridge deck is 19m, the bridge height is 123.8m, and the main body is east-west. The main bridge is a 82+4×150+82m continuous rigid frame cast-in-place box girder. The lower structure adopts pile foundation. The pier column is a variable-section double-limb hollow pier. The maximum pier height is 114.8m and the bridge deck width is 19m. This study is for the mid-span section on both sides of a main pier. Multiple three-dimensional anemometers and two-dimensional accelerometers are used to form the front-end data acquisition and transmission subsystem. The instrument layout is as follows: Figure 3 The acquisition instrument remotely transmits the wind environment data and bridge vibration response data at the bridge site to the cloud database through the Internet of Things card. The measured wind speed time history curve collected is shown in Figure 4 As shown in the figure, the measured bridge vibration time history curve is as follows Figure 5 shown.

[0112] The later data processing and calculation subsystem calls the data in the cloud database for processing and analysis, and analyzes the correlation between the input wind characteristics and the output vibration response characteristics. The analysis results are as follows: Figure 6 As shown. In order to verify the prediction accuracy and generalization ability of the TCN model, this paper selects recurrent neural network (RNN), long short-term memory network (LSTM) and gated unit network (GRU) to predict the acceleration in four directions, and compares the experimental results with the TCN model. In order to ensure the fairness of the experimental comparison, all models use the same parameter configuration for acceleration prediction in each direction, and the Adam algorithm is used as the optimizer for model training. The acceleration prediction results of different models in different directions are plotted Figure 7(a) to Figure 7(d) .

[0113] The prediction results show that the prediction effect of the TCN model is better than that of other comparison models, the prediction effect of the RNN model is the worst, and the prediction results of the LSTM and GRU models are between the TCN and RNN models. In order to more accurately compare the prediction effects of different models, the error indicators of each model are calculated separately, as shown in Table 1:

[0114] Table 1 Comparison of prediction errors of different deep learning models

[0115]

[0116] In the prediction of torsional acceleration and vertical acceleration, the TCN model achieved an error of less than 5%, which shows that TCN has a good effect in processing sequential tasks due to its unique model structure. In addition, the MAE and RMSE of the TCN model are also the lowest among the four models. In the prediction of transverse acceleration, the MAE, RMSE and MAPE of the TCN model are 16.78, 23.13 and 13.6%, respectively, which are 3.00%, 5.55% and 2.86% higher than the GRU model, and 17.54%, 16.59% and 14.47% higher than the RNN model. In the prediction of longitudinal acceleration, the MAE, RMSE and MAPE of the TCN model are 9.98, 13.46 and 13.6%, respectively, which is the best among the four models.

[0117] The main reason for the above phenomenon is that the wind vibration response of the bridge has a memory effect. When processing long sequences, RNN often encounters the problem of gradient vanishing or gradient exploding, which makes it difficult to capture long-distance dependencies; LSTM introduces a gating mechanism that can well capture long-term dependencies, thereby effectively processing long sequence data; GRU has fewer parameters and a simpler structure, and can be applied to various complex time series patterns and has slightly better effects than LSTM; TCN combines causal convolution, dilated convolution, and residual connection to achieve flexible adjustment of the receptive field, high parallelism, and gradient stability, and also has significant advantages in memory utilization.

[0118] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A wind-induced vibration response prediction system for large-span bridges based on TCN network, characterized by: It includes front-end data acquisition and transmission subsystem, post-data processing and calculation subsystem, and bridge wind-induced vibration response prediction subsystem; among which: The front-end data acquisition and transmission subsystem is a bridge monitoring system composed of multiple three-dimensional anemometers and acceleration sensors, which can continuously monitor the wind environment data and bridge vibration response data at the bridge site in real time and transmit them to the database; The post-data processing and calculation subsystem can pre-process the wind environment data at the bridge site and the bridge vibration response data in the database, calculate the corresponding wind characteristic data and bridge vibration response characteristics, and perform wind vibration response analysis on the measured bridge; The bridge wind-induced vibration response prediction subsystem is constructed based on the TCN network prediction model, with the wind characteristic data obtained in the later data processing and calculation subsystem as input, and the bridge vibration response data as output to train the TCN algorithm, and derive the wind-induced vibration response prediction data to provide information reference for bridge construction and operation decision-making; The bridge wind-induced vibration response prediction subsystem can evaluate the prediction effect, compare it with the actual measured results of the bridge wind-induced vibration response during the prediction period, and provide feedback to update the TCN network prediction model parameters.

2. According to claim 1, a TCN network-based large-span bridge wind-induced vibration response prediction system is characterized by: The post-data processing and calculation subsystem processes the wind environment data at the original bridge site with a basic time interval of 10 minutes to obtain the measured wind characteristics, including: average wind speed, wind direction angle, wind attack angle, turbulence intensity, turbulence integral scale, and fluctuating wind power spectrum density; The calculation formula of the wind characteristics is as follows: u=xcos(θ)+ysin(θ)-U (4) v=-xsin(θ)+ycos(θ) (5) AND i =σ i / U (i=u,v,w) (7) In the formula, x, y, and z are the wind speeds along the bridge, across the bridge, and vertically, respectively; are the average wind speeds along the bridge, across the bridge and vertically in the basic time interval; U, θ, α are the average wind speed, wind direction angle and wind attack angle in the basic time interval; u, v, w are the along wind direction, across wind direction and vertical pulsating wind respectively; I i is the turbulence degree in each wind direction; L i is the turbulence integral scale of each wind direction; S i (f) is the power spectrum density of fluctuating wind in each wind direction.

3. The wind-induced vibration response prediction system for a large-span bridge based on a TCN network according to claim 1 is characterized by: The post-processing data and calculation subsystem, when processing the original bridge vibration response data, uses a second-order Butterworth-based low-pass filter to remove unnecessary high-frequency noise, subtracts the signal mean to avoid drift, and converts the vibration time domain signal into a frequency domain signal through a fast Fourier transform (FFT) to obtain the bridge modal frequency and damping ratio; The damping ratio ζ can be calculated by the logarithmic decay rate of the free decay vibration, assuming that the two adjacent peaks are x1 and x2: Damping ratio calculation formula:

4. The wind-induced vibration response prediction system for a large-span bridge based on a TCN network according to claim 1 is characterized by: The post-data processing and calculation subsystem can use the measured wind characteristic data as input features and the measured bridge vibration response characteristics as output features to construct a TCN machine learning data set, analyze the correlation between wind characteristics and vibration response characteristics, and determine the main correlation parameters of the bridge wind vibration response, so as to minimize the redundant input variables of subsequent neural network modeling and improve the training and prediction efficiency of the model.

5. The wind-induced vibration response prediction system for a large-span bridge based on a TCN network according to claim 1 is characterized by: The bridge wind-induced vibration response prediction subsystem uses the wind characteristics with high correlation after analysis by the later data processing and calculation subsystem as input vectors, and the longitudinal, transverse, vertical and torsional accelerations of the bridge as output vectors, and predicts bridge vibrations under different environments based on the TCN deep learning network.

6. The wind-induced vibration response prediction system for a large-span bridge based on a TCN network according to claim 1 is characterized by: The bridge wind vibration response prediction subsystem uses three evaluation indicators, namely mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE), to achieve a comprehensive evaluation of the prediction ability of the TCN model, compare it with the actual measured results of the bridge wind vibration response during the prediction period, and provide feedback to update the TCN network prediction model parameters. The calculation formulas for each evaluation indicator are as follows: Among them, y t represents the measured value of acceleration at time t; Represents the predicted value of acceleration at time t; N is the total number of samples.

7. A method for predicting wind-induced vibration response of a long-span bridge based on a TCN network, characterized in that: The following steps are involved: S1, the front-end data acquisition and transmission subsystem collects wind environment data and bridge vibration response data at the bridge site and transmits them to the database; S2, the post-data processing and calculation subsystem analyzes and processes the measured data of wind environment and wind-induced vibration response of bridges; S2-1, call the original wind environment measured data in the database and process it with 10 minutes as the basic time interval to obtain the measured wind characteristics, including: average wind speed, wind direction angle, wind attack angle, turbulence intensity, turbulence integral scale, and fluctuating wind power spectrum density; S2-2, call the measured data of wind-induced vibration response of bridges in the database, use a low-pass filter based on second-order Butterworth to remove unnecessary high-frequency noise, subtract the signal mean to avoid drift, and convert the vibration time domain signal into a frequency domain signal through fast Fourier transform (FFT) to obtain the bridge modal frequency and damping ratio; S2-3, using the measured wind characteristic data as input features and the measured bridge vibration response characteristics as output features, a TCN machine learning data set was constructed to analyze the correlation between wind characteristics and vibration response characteristics, and determine the main correlation parameters of the bridge wind vibration response, so as to minimize the redundant input variables of the subsequent neural network modeling and improve the training and prediction efficiency of the model; S3, the bridge wind-induced vibration response prediction subsystem realizes the bridge wind-induced vibration response prediction based on the TCN network and compares it with the measured data to update the improved algorithm; S3-1, based on the TCN deep learning network, the wind characteristics with high correlation after post-data processing and calculation subsystem analysis are used as input vectors, and the longitudinal, transverse, vertical and torsional accelerations of the bridge are used as output vectors to predict bridge vibrations under different environments; S3-2, using the three evaluation indicators of mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) to realize the comprehensive evaluation of the prediction ability of the TCN model, and comparing it with the measured results of the wind-induced vibration response of the bridge during the prediction period, and providing feedback to update the parameters of the TCN network prediction model.

8. The early warning method for wind-induced vibration response of a long-span bridge according to claim 7 is characterized by: In step S2, the calculation formula of the wind characteristics is as follows: u=xcos(θ)+ysin(θ)-U (4) v=-xsin(θ)+ycos(θ) (5) AND i =σ i / U (i=u,v,w) (7) In the formula, x, y, and z are the wind speeds along the bridge, across the bridge, and vertically, respectively; are the average wind speeds along the bridge, across the bridge and vertically under the basic time interval; U, θ, α are the average wind speed, wind direction angle and wind attack angle under the basic time interval; u, v, w are the along wind direction, across wind direction and vertical pulsating wind respectively; I i is the turbulence degree in each wind direction; L i is the turbulence integral scale of each wind direction; S i (f) is the power spectrum density of fluctuating wind in each wind direction.

9. The early warning method for wind-induced vibration response of a long-span bridge according to claim 7, characterized in that: In step S2, the damping ratio ζ can be calculated by the logarithmic decay rate of the free decay vibration. Assuming that the two adjacent peaks are x1 and x2, then: Damping ratio calculation formula:

10. The early warning method for wind-induced vibration response of a long-span bridge according to claim 7, characterized in that: In step S2, the features whose absolute values ​​of the calculated wind features and bridge vibration response features are less than 0.3 will not be used as input features of the machine learning dataset.

11. The early warning method for wind-induced vibration response of a long-span bridge according to claim 7, characterized in that: The calculation formulas for each evaluation index in step S3 are as follows: Among them, y t represents the measured value of acceleration at time t; Represents the predicted value of acceleration at time t; N is the total number of samples.

Images