CN117910120B - Buffeting response prediction method for wind-bridge system based on lightweight Transformer - Google Patents
Buffeting response prediction method for wind-bridge system based on lightweight Transformer Download PDFInfo
- Publication number
- CN117910120B CN117910120B CN202410321437.9A CN202410321437A CN117910120B CN 117910120 B CN117910120 B CN 117910120B CN 202410321437 A CN202410321437 A CN 202410321437A CN 117910120 B CN117910120 B CN 117910120B
- Authority
- CN
- China
- Prior art keywords
- wind speed
- model
- bridge
- buffeting
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 68
- 230000004044 response Effects 0.000 title claims abstract description 64
- 238000004364 calculation method Methods 0.000 claims abstract description 26
- 238000012545 processing Methods 0.000 claims abstract description 24
- 239000000725 suspension Substances 0.000 claims abstract description 17
- 230000006399 behavior Effects 0.000 claims abstract description 6
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 6
- 238000012549 training Methods 0.000 claims description 65
- 239000011159 matrix material Substances 0.000 claims description 26
- 238000000354 decomposition reaction Methods 0.000 claims description 22
- 230000006870 function Effects 0.000 claims description 22
- 230000000694 effects Effects 0.000 claims description 19
- 238000013136 deep learning model Methods 0.000 claims description 18
- 230000007246 mechanism Effects 0.000 claims description 17
- 230000004913 activation Effects 0.000 claims description 13
- 238000010606 normalization Methods 0.000 claims description 13
- 230000008569 process Effects 0.000 claims description 13
- 238000013135 deep learning Methods 0.000 claims description 11
- 238000004458 analytical method Methods 0.000 claims description 7
- 238000009826 distribution Methods 0.000 claims description 7
- 230000003068 static effect Effects 0.000 claims description 6
- 238000012360 testing method Methods 0.000 claims description 5
- 238000013016 damping Methods 0.000 claims description 4
- 238000006073 displacement reaction Methods 0.000 claims description 4
- 238000011156 evaluation Methods 0.000 claims description 4
- 238000000605 extraction Methods 0.000 claims description 4
- 230000010354 integration Effects 0.000 claims description 4
- 238000010200 validation analysis Methods 0.000 claims description 4
- 238000012795 verification Methods 0.000 claims description 4
- 230000001133 acceleration Effects 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 2
- 238000003062 neural network model Methods 0.000 claims description 2
- 238000005070 sampling Methods 0.000 claims description 2
- 230000008676 import Effects 0.000 claims 1
- 238000010276 construction Methods 0.000 abstract description 6
- 238000013461 design Methods 0.000 abstract description 6
- 238000012423 maintenance Methods 0.000 abstract description 4
- 230000000541 pulsatile effect Effects 0.000 description 8
- 238000010586 diagram Methods 0.000 description 7
- 238000004088 simulation Methods 0.000 description 7
- 230000008901 benefit Effects 0.000 description 6
- 238000011160 research Methods 0.000 description 6
- 230000003595 spectral effect Effects 0.000 description 4
- 238000001228 spectrum Methods 0.000 description 3
- 230000009471 action Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 2
- 230000007613 environmental effect Effects 0.000 description 2
- 230000005284 excitation Effects 0.000 description 2
- 239000012530 fluid Substances 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 238000013508 migration Methods 0.000 description 2
- 230000005012 migration Effects 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 230000010349 pulsation Effects 0.000 description 2
- 238000007634 remodeling Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000012938 design process Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 239000013013 elastic material Substances 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000015654 memory Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 230000002035 prolonged effect Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/13—Architectural design, e.g. computer-aided architectural design [CAAD] related to design of buildings, bridges, landscapes, production plants or roads
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/23—Design optimisation, verification or simulation using finite element methods [FEM] or finite difference methods [FDM]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/28—Design optimisation, verification or simulation using fluid dynamics, e.g. using Navier-Stokes equations or computational fluid dynamics [CFD]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0495—Quantised networks; Sparse networks; Compressed networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0499—Feedforward networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/10—Numerical modelling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/08—Fluids
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/12—Timing analysis or timing optimisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/14—Force analysis or force optimisation, e.g. static or dynamic forces
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Computational Mathematics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Geometry (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Computer Hardware Design (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Operations Research (AREA)
- Fluid Mechanics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Probability & Statistics with Applications (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Architecture (AREA)
- Civil Engineering (AREA)
- Structural Engineering (AREA)
- Aerodynamic Tests, Hydrodynamic Tests, Wind Tunnels, And Water Tanks (AREA)
Abstract
The invention relates to the technical field of wind-resistant design of bridges, and discloses a buffeting response prediction method of a wind-bridge system based on a lightweight converter. Selecting a proper fluctuating wind speed model according to the set average wind speed, generating a fluctuating wind speed time sequence generated by superimposing the average wind speed after fluctuating wind speed, and converting the fluctuating wind speed time sequence into buffeting force time interval data after aerodynamic theory and Fourier transform calculation; then, establishing a finite element model of the suspension bridge, inputting the generated buffeting force time course data into the suspension bridge model, and integrating the generated data with the fluctuating wind speed time course data to obtain buffeting response samples of the suspension bridge; and constructing a seven-layer convolutional neural network model of a transducer structure for processing sequence data, and capturing the inherent characteristics and dynamic behaviors of the fluctuating wind speed and the buffeting response. The invention can efficiently predict buffeting response, provide information reference for construction and maintenance of bridge engineering, prolong the service life of the structural connecting piece of the bridge and save maintenance cost.
Description
Technical Field
The invention relates to the technical field of wind-resistant design of bridges, in particular to a buffeting response prediction method of a wind-bridge system based on a lightweight converter.
Background
In the past few decades, the construction of highway bridges in China has progressed rapidly. Due to its significantly high efficiency, convenience and large capacity for transportation efficiency, highways have become one of the main options for freight and passenger traffic. With the advancement of time, the length ratio of the large-span bridge is continuously increased, so that the large-span bridge is connected with various areas, and the development of regional economy is promoted. Therefore, the highway network must face unavoidable threat of buffeting response in construction and operation, and because buffeting response occurs frequently, when buffeting response occurs, probability of a vehicle passing through a bridge is remarkably increased, and riding comfort is reduced. In the last two decades, several bridge structural damage accidents caused by wind loads have occurred, and therefore it has become critical to ensure the safety of the running of vehicles on bridges under wind load excitation.
Particularly, the large-span bridge is widely used in areas with large wind load influence such as large-span scenes such as rivers, straits and the like. However, with the increasing span of the bridge, the structure has the characteristics of large flexibility, light weight, low damping and the like, so that the structure becomes very sensitive to wind load, and the wind resistance problem of the bridge also becomes more complex and severe. Among limited vibrations caused by wind load, buffeting is forced vibration under the action of a pulsating wind speed, is limited-amplitude vibration with forced vibration characteristics, and has smaller vibration amplitude and insufficient vibration amplitude to cause direct damage to a bridge compared with vortex vibration and wind-rain excitation. However, because the wind speed of the buffeting response is low and the efficiency is high, the buffeting frequency is very high, and bridges in the natural environment are difficult to avoid; and long-time buffeting can cause fatigue damage to the connecting members of the bridge, and meanwhile driving experience can be influenced. In general, in the study of buffeting response, the investigation of bridge system safety under wind load is relatively limited.
Currently, many studies focus on the prediction of buffeting response of train-bridge systems under wind loading. In these conventional research methods, there are limitations in that the computational complexity is high, and a large-scale structure requires a large amount of computational resources when faced with complex, highly nonlinear, and large-scale data sets. Deep learning is used as an emerging learning model, and the deep learning model automatically learns the characteristics and the representation of data without relying on complicated characteristic engineering. The multi-level, highly nonlinear structure enables it to better capture complex patterns in data and exhibit good performance over large data sets. The deep learning also has the capability of end-to-end learning, and the learning is directly performed from the original input to the output, so that the flow of model design is simplified. These advantages make deep learning excellent in handling complex, large-scale, high-dimensional data in time series problems, becoming a popular choice for intelligent prediction methods of buffeting responses.
At present, a plurality of researches fully prove that the deep learning has good prediction accuracy and subsequent research value on wind-bridge system prediction. The success of LSTM (Long Short-Term Memory) in time series prediction is widely applied to the fields with strong relevance such as price prediction, stock trend, bridge wind vibration displacement prediction and the like, but because of the structural problem of the model, when the model is used for processing Long time series, the calculation bottleneck may exist due to the characteristic of serial calculation. And the number of the super parameters is small, so that the training requirements of a large-scale data set are difficult to adapt well. In contrast, a transducer has greater expressive power and adaptability in processing long sequences, global dependencies, and large-scale datasets, becoming a powerful tool in processing sequence data.
Disclosure of Invention
Aiming at the problems, the invention aims to provide a buffeting response prediction method of a wind-bridge system based on a lightweight transducer, which utilizes a deep learning network technology to establish a relation model of the pulsating wind speed and the vertical bridge deformation of a bridge, and samples a novel transducer model. And the buffeting response prediction is efficiently carried out, information reference is provided for the construction and maintenance of bridge engineering, the service life of the structural connecting piece of the bridge is prolonged, and the maintenance cost is saved. The technical proposal is as follows:
A buffeting response prediction method of a wind-bridge system based on a lightweight transducer comprises the following steps:
step 1: selecting a fluctuating wind speed model according to the average wind speed, and determining parameters related to the fluctuating wind speed model; simulating the fluctuating wind speed by adopting a harmonic superposition method;
step 2: generating a pulsating wind speed time series data superimposed with the average wind speed as a pulsating wind speed sample by using the selected pulsating wind speed model and parameters;
step 3: converting the pulse wind speed time series data generated in the step 2 into buffeting force time series data serving as a bridge buffeting response sample;
Step 4: establishing a bridge motion model through commercial finite element software, inputting buffeting force time sequence data, and extracting vertical bridge deformation of the bridge;
Step 5: taking the pulsating wind speed time series data generated in the step 2 as an input set, taking the vertical bridge deformation of the bridge obtained in the step 4 as an output set, forming a plurality of data sets with different time series lengths, and dividing the data sets into a training set and a testing set;
Step 6: constructing a transducer deep learning network, verifying the prediction correctness of the network, and performing super-parameter setting on the network before training;
Step 7: inputting the training set into a transducer deep learning network for predictive training to obtain a predictive model of the vertical deformation of the bridge structure, and then storing the trained model;
And (3) carrying out data remodeling and normalization on the buffeting response sample and the fluctuating wind speed sample by adopting a transducer deep learning model: the transducer deep learning model comprises an input layer, a plurality of TransformerBlock layers and an output layer; utilizing the input layer part to adjust the format of the data into a proper transducer model; the transducer model consists of a plurality of TransformerBlock layers, and the output layer comprises a plurality of layers and is used for mapping learned potential distribution to output and predicting to obtain a prediction model of the vertical deformation of the bridge structure;
step 8: different time sequence lengths and different sample orders of magnitude buffeting response prediction samples are obtained through a transducer model, and the average value and standard deviation of the transducer model are calculated according to the buffeting response prediction samples.
Further, the step 2 specifically includes:
Step 2.1: generating fluctuating wind speed fluctuation at different positions based on specified parameters by adopting a harmonic superposition method, wherein a wind field is decomposed into harmonic components; loading node coordinate information and setting basic parameters including wind speed, ground roughness and frequency resolution; a basic formula for pulsating wind speed is employed, wherein the vertical distribution of wind speed is represented by an exponential function:
（1）；
in the formula, u z (i) is the average wind speed; v 10 is the base wind speed; z i is the vertical position of the node; alpha is an exponential parameter;
step 2.2: the friction wind speed at each node is calculated as follows:
（2）；
In the formula, u z0 (i) is friction wind speed, and K is a constant; z 0 is the ground roughness;
Step 2.3: introducing a direct decomposition frequency; through the decomposition, the fluctuating wind speed signal is converted from a time domain to a frequency domain, and the distance between adjacent nodes is calculated, so that a foundation is provided for the establishment of a subsequent cross spectrum density matrix; in the process of the pulsating wind speed signal, when the Fourier decomposition is adopted for frequency analysis, the numerical method is adopted for calculating the Fourier transformation; the numerical method approximately calculates the frequency components of the signals, namely a Fourier decomposition numerical solution; decomposing the cross spectral density matrix by using a frequency component of a Fourier decomposition numerical solution through a cholesky decomposition method; wherein, the elements of the cross spectral density matrix are calculated by the correlation of the space distance and the frequency;
Step 2.4: the interpolation point calculation and interpolation method is used, wind speed components of different nodes are obtained through interpolation of the frequency domain representation, so that gaps among discrete frequency points are filled, and the smoothness of frequency domain components is ensured;
Step 2.5: converting the frequency domain representation into a time domain pulsating wind speed signal by adopting inverse fast Fourier transform; and (3) superposing the time domain representation of each frequency component by circularly traversing each node and frequency to obtain the final fluctuating wind speed time course of the node position, wherein the formula is expressed as follows:
（3）；
（4）；
In the formula, X is a fluctuating wind speed frequency domain sequence; n is the length of the sequence of fluctuating wind speeds; x n is the signal in the pulsatile wind speed time domain; x k is an element in the pulsatile wind speed frequency domain sequence, subscript k=0, 1,..; changes in frequency components over time and frequency index are described; k is an index of the frequency domain; n is the index of the time domain.
Further, the specific process in the step 3 is as follows:
Step 3.1: loading average wind speed U mean, transverse wind speed U 0, vertical wind speed w 0 and girder node position x pos from the simulated wind field data; the spatial variation of the transverse wind speed and the vertical wind speed is calculated by a finite difference method;
Step 3.2: obtaining the length L of the beam section and the distance between adjacent nodes through differential calculation; in the static wind load calculation stage, components at all nodes on the main beam section are calculated by adopting a formula based on aerodynamic theory, wherein the formula is as follows:
（5）；
（6）；
（7）；
in the formula, F D、FL and M M are respectively a drag component, a lift component and a moment component; Is air density, U mean is average wind speed, H is influence height in wind speed vertical direction, B is bridge width, L is length of main girder section, C D、CL and C M are resistance coefficient, lift coefficient and moment coefficient respectively;
step 3.3: the dynamic wind load time course in the time domain is obtained by processing the transverse wind speed u 0 and the vertical wind speed w 0; converting the fluctuating wind speed of the frequency domain into a buffeting force time course of the time domain by adopting a Fourier transform method; the dynamic wind load time course formula is expressed as follows:
（8）；
（9）；
（10）；
In the formula, F D（t）、FL (t) and M M (t) are respectively a drag component time interval, a lift component time interval and a moment component time interval; u (t) is the lateral wind speed in the time domain, w (t) is the vertical wind speed in the time domain, and t represents time.
Further, in the step 4, the bridge motion model is as follows:
（11）；
（12）；
In the formula, M xx and M yy are respectively the quality matrixes of the bridge in the transverse direction and the bridge in the vertical direction, AndThe transverse and vertical accelerations of the bridge are respectively, C xx and C yy are respectively the damping matrices of the bridge in the transverse and vertical directions,AndThe transverse and vertical speeds of the bridge are respectively, K xx and K yy are respectively stiffness matrixes of the bridge in the transverse and vertical directions, u and v are respectively transverse and vertical displacements of the bridge, and F gx (t) and F gy (t) are respectively transverse and vertical wind loads caused by buffeting force.
Further, the step 6 specifically includes:
Pre-training based on buffeting response of a simulated suspension bridge, and constructing a seven-layer convolutional neural network model of a transducer structure for processing sequence data, wherein the model comprises a plurality of TransformerBlock layers, and each layer comprises a self-attention mechanism, a feedforward network and layer normalization; the system is used for capturing the intrinsic characteristics and dynamic behaviors of the pulsating wind speed and buffeting response; taking the pulsating wind speed sample generated in the step (2) simulation as input data, and taking the bridge buffeting response sample generated in the step (3) simulation as output data; the input data is subjected to feature extraction through three TransformerBlock, then feature integration is performed through two full-connection layers, and finally a pre-trained transducer model is output.
Further, the step 7 specifically includes:
step 7.1: the input layer of the transducer deep learning model is a layers layer, and the input layer is used as the input of the model during training; defining the shape of a buffeting force time sequence and a fluctuating wind speed time sequence of a neural network model, allowing a transducer model to receive input with a set structure, and receiving training data for forward propagation of the model so as to ensure that the transducer model can read the data normally;
Step 7.2: the TransformerBlock layers contain self-attention mechanism, feedforward network and layer normalization; wherein the self-attention mechanism selects a multi-headed sub self-attention for learning feature associations in the input sequence in each TransformerBlock; in each layer, the model further learns buffeting response characteristic relation through a feedforward network on the basis of self-attention; calculating correlations of the buffeting response time sequence TransformerBlock with all keys, and using the correlations to perform weighted average on the values for layer normalization; parameter calculations are normalized in matrix form and layer as:
（13）；
（14）；
In the formula, Q, K and V respectively represent three parameter matrixes, X input is an input matrix, and W Q、K、V is a weight matrix corresponding to the three parameter matrixes; z is an output matrix; d k is the scaling factor; softmax is the activation function used to calculate the attention weight; t is a transposed symbol; attention is self-Attention mechanism operation;
The feedforward network comprises two full-connection layers, wherein the activation function of the first layer is Relu, the second layer does not use the activation function, and the structure of the feedforward network introduces nonlinear characteristics through the activation function; generating a coded information matrix C after the calculation of the multiple layers TransformerBlock, wherein the coded information matrix C is applied to a subsequent output layer;
Step 7.3: the output layer of the transducer deep learning model comprises a full-connection layer and a remodelling layer; the method has the effects that the modeled characteristics are mapped to an output space, and then the shape of the output is adjusted to be matched with the shape of the target output through a remodelling operation; during training, the output of the transform deep learning model is used to calculate the loss function for back propagation and weight updating.
Further, before the step 8, the method further includes: the model is evaluated by using a separate validation dataset, in particular:
step a: simulating the fluctuating wind speeds of N 0 suspension bridges under the average wind speed, and generating a fluctuating wind speed sample after superposition of the average wind speeds;
Step b: simulating to obtain bridge deformation under the given parameter condition according to the finite element model; the method comprises the steps of extracting a verification sample of a training sample set from a sample by using a random sampling method;
Step c: taking a pulsating wind speed sample as input, and taking the corresponding vertical deformation of the bridge as output;
step d: after each training period is finished, evaluating the performance of the model by using a verification set, and selecting a mean square error, an average absolute error, a decision coefficient and variance thereof as model evaluation indexes to evaluate the prediction performance and generalization capability of the transducer model;
Step e: if the model prediction effect is not ideal, optimizing the transducer model, including adjusting the architecture of the transducer model, the setting of an optimizer, the learning rate and the super parameters, and obtaining and storing the model with optimal performance in the training process.
Compared with the prior art, the invention has the beneficial effects that:
1) The invention provides an effective bridge buffeting response prediction method, which realizes data support for real-time and accurate judgment of a bridge by predicting the bridge buffeting response based on a lightweight transducer model and provides powerful support for timely finding and processing potential problems for operation management staff.
2) The invention provides a time sequence processing method for the influence of pulsating wind on a structure, which has excellent processing on long-range dependency relationship because a transducer model has the advantage of parallel computation, and can better capture the space-time dynamics of structural vibration caused by the pulsating wind.
3) The invention is based on a transducer model, because the self-attention mechanism allows the network to automatically focus on information at different positions in the input sequence, making it more adaptive. This provides insight into the subsequent study of process model migration, etc.
4) The invention is based on a transducer model, the self-attention mechanism of which makes it to a certain extent interpretable. Engineers can understand the specific input features of the model's interest in the predictions by analyzing the model's attention weights to better understand the influencing factors of the structural vibrations.
Drawings
FIG. 1 is a buffeting response prediction system analysis process.
FIG. 2 is a schematic diagram of a sample of fluctuating wind speeds.
Fig. 3 is a diagram of a suspension bridge arrangement.
Fig. 4 is a schematic diagram of a vertical bridge vibration sample.
Fig. 5 is a construction diagram of a transducer deep learning model.
Fig. 6 (a) shows the loss values for the training phase for two models of different sequence lengths (200).
Fig. 6 (b) shows the loss values for the training phase for two models of different sequence lengths (500).
Fig. 6 (c) shows the loss values for the training phase for two models of different sequence lengths (800).
Fig. 6 (d) shows the loss values for the training phase for two models of different sequence lengths (1000).
Fig. 7 (a) shows the loss values for the training phase under two different training sample sets (200) of the model.
Fig. 7 (b) shows the loss values for the training phase for two models with different training sample sets (150).
Fig. 7 (c) shows the loss values for the training phase for two models with different training sample sets (100).
Fig. 7 (d) shows the loss values for the training phase for two models with different training sample sets (50).
Fig. 8 (a) shows the prediction effect of the prediction phase for two models of different sequence lengths (200).
Fig. 8 (b) shows the prediction effect of the training prediction phase for two models of different sequence lengths (500).
Fig. 8 (c) shows the prediction effect of the training prediction phase for two models of different sequence lengths (800).
Fig. 8 (d) shows the prediction effect of the training prediction phase for two models of different sequence lengths (1000).
Fig. 9 (a) shows the prediction effect of the prediction phase under two different training sample sets (200).
Fig. 9 (b) shows the prediction effect of the prediction phase under two different training sample sets (150).
Fig. 9 (c) shows the prediction effect of the prediction phase under two different training sample sets (100).
Fig. 9 (d) shows the prediction effect of the prediction phase under two different training sample sets (50).
Detailed Description
The invention is described in further detail below with reference to the drawings and the detailed description.
The invention relates to the technical field of wind-resistant design of bridges, in particular to a buffeting response prediction method of a wind-bridge system based on a lightweight converter. The method can be divided into two parts, namely a simulation of a buffeting response sample and a transducer deep learning model. Firstly, selecting a proper fluctuating wind speed model according to a specific average wind speed, generating a fluctuating wind speed, overlapping the fluctuating wind speed, and converting a fluctuating wind speed time sequence generated after the average wind speed is superimposed into buffeting force time-course data after calculation of an aerodynamic theory and a Fourier transform mathematical tool; and then, establishing a finite element model of the suspension bridge, inputting the generated buffeting force time course data into the suspension bridge model, and integrating the generated data with the fluctuating wind speed time course data to obtain buffeting response samples of the suspension bridge.
After data collection is completed, a seven-layer transform structure convolutional neural network model for processing sequence data is constructed, wherein the model comprises a plurality of TransformerBlock layers, and the aim is to capture the intrinsic characteristics and dynamic behaviors of the pulsating wind speed and buffeting response; and taking the time series data of the pulsating wind speed as input data and taking the buffeting response of the bridge as output data.
Finally, after the model is trained and optimized, training data and test data are input into the constructed network model in batches in the training process of the transducer model, and in order to be capable of fully knowing the advantages of the model on processing long sequences, six model evaluation indexes of MSE, MAE, R 2 and variances thereof are introduced to evaluate the prediction performance and generalization capability of the transducer model. The flow is shown in figure 1, and the specific technical scheme is as follows:
step 1: based on the average wind speed, an appropriate pulsatile wind speed model is selected, and parameters associated with the pulsatile wind speed model, such as the number of frequency segments, cut-off frequency, etc., are determined. The invention adopts a harmonic superposition method to simulate the fluctuating wind speed.
A model is generated by adopting the fluctuating wind speed based on Fourier decomposition, the model has spectrum controllability, and the frequency components of the generated fluctuating wind speed signal can be flexibly adjusted by virtue of the Fourier decomposition, so that the characteristics of an actual scene can be better matched. Meanwhile, the Fourier decomposition can capture the contribution of various frequency components to pulsation; this helps to generate a more realistic and complex pulsating wind speed signal. Therefore, the Fourier decomposition-based pulsating wind speed generation model is more suitable for the pulsating wind speed research of the large-span bridge.
Step 2: programming a program ComputeWind in programming software Matlab; the program generates a time series of fluctuating wind speeds superimposed with the average wind speed as a sample of fluctuating wind speeds using the selected model of fluctuating wind speeds and parameters.
Code ComputeWind generates the fluctuating wind speed fluctuations at different locations based on specified parameters, and the code adopts a harmonic superposition method in which the wind field is decomposed into individual harmonic components. First, node coordinate information is loaded and basic parameters including wind speed, ground roughness, frequency resolution, etc. are set. The code uses a basic formula for fluctuating wind speeds, where the vertical distribution of wind speeds is represented by an exponential function:
（1）；
In the formula, v 10 is the basic wind speed; z i is the vertical position of the node; alpha is an exponential parameter.
The friction wind speed at each node is calculated after the average wind speed u z (i) is obtained, by the following formula:
（2）；
In the formula, K is a constant; z 0 is the roughness of the ground.
After the calculation is completed, the direct decomposition frequency is imported. Through the decomposition, the pulsating wind speed signal can be converted from a time domain to a frequency domain, so that the frequency characteristic of the signal can be better understood, the contributions of different frequency components can be identified, or operations such as filtering and the like can be performed; meanwhile, the distance between adjacent nodes is calculated, and a foundation is provided for the establishment of a subsequent cross spectrum density matrix; in the case of performing frequency analysis by fourier decomposition in the pulsating wind speed signal processing, since the signal is discrete, the fourier transform is calculated by a numerical method, and the frequency component in the analysis form is not obtained by the analysis method. Such numerical methods may be Discrete Fourier Transforms (DFT) or Fast Fourier Transforms (FFT) that calculate approximately the frequency content of the signal by means of computer algorithms, i.e. fourier decomposition numerical solutions. Decomposing the cross spectral density matrix by a Cholesky decomposition method by adopting frequency components of a Fourier decomposition numerical solution; wherein, the elements of the cross spectral density matrix are calculated by the correlation of the space distance and the frequency; the Matlab program is designed to cycle through each node and each frequency to obtain a frequency domain representation of the pulsating wind field.
Then, by interpolating the frequency domain representation, wind speed components of different nodes are obtained. The step uses interpolation point calculation and interpolation methods to fill the gap between discrete frequency points and ensure the smoothness of frequency domain components.
Finally, an Inverse Fast Fourier Transform (IFFT) is used to convert the frequency domain representation into a time domain pulsating wind speed signal. And (3) superposing the time domain representation of each frequency component by circularly traversing each node and frequency to obtain the final fluctuating wind speed time course of the node position, wherein the formula is expressed as follows:
（3）；
（4）；
In the formula, X is a fluctuating wind speed frequency domain sequence; n is the length of the sequence of fluctuating wind speeds; x n is the signal in the pulsatile wind speed time domain; x k is an element in the pulsatile wind speed frequency domain sequence, subscript k=0, 1,..; changes in frequency components over time and frequency index are described; k is an index of the frequency domain; n is the index of the time domain.
FIG. 2 is a sample of the generated pulsatile wind speed.
Step 3: programming a program ComputeForceBridge in programming software Matlab; the program converts the pulse wind speed time series data generated in the step 2 into buffeting force time series data, and the buffeting force time series data is a buffeting response sample.
The ComputeForceBridge code is based on wind field simulation data, and through the application of aerodynamic theory and a Fourier transform mathematical tool, accurate modeling of static and dynamic wind load time courses on the bridge structure is realized, and buffeting force on a target structure is calculated. According to the basic principle that the elastic structure is subjected to wind load, the values of parameters such as air density, reference area of the structure, resistance coefficient and the like are determined, and the wind speed is converted into buffeting force. The procedure is described below:
First, the average wind speed U mean, the transverse wind speed U 0, the vertical wind speed w 0, and the girder-node position x pos are loaded from the simulated wind field data. The spatial variation of the transverse wind speed and the vertical wind speed is calculated by a finite difference method. Next, the length L of the beam section and the distance between adjacent nodes are obtained by differential calculation. In the static wind load calculation stage, components at all nodes on the main beam section are calculated by adopting a formula based on aerodynamic theory, wherein the formula is as follows:
（5）；
（6）；
（7）；
In the formula (i), Is the air density, U mean is the average wind speed, H is the height of influence in the vertical direction of the wind speed, B is the bridge width, L is the length of the main girder section, and C D、CL and C M are the drag coefficient, lift coefficient and moment coefficient, respectively.
The dynamic wind load time course in the time domain is obtained by processing the transverse wind speed u 0 and the vertical wind speed w 0. And converting the fluctuating wind speed of the frequency domain into a buffeting force time course of the time domain by adopting a Fourier transform method. The dynamic wind load time course formula is expressed as follows:
（8）；
（9）；
（10）；
In the formula, F D（t）、FL (t) and M M (t) are respectively a drag component time interval, a lift component time interval and a moment component time interval; u (t) is the lateral wind speed in the time domain, w (t) is the vertical wind speed in the time domain, and t represents time.
And outputting time course data to corresponding files after the static and dynamic wind load calculation is completed, so as to be used for subsequent transducer and other model training.
Step 4: and establishing a bridge motion model through commercial finite element software Ansys, inputting buffeting force time course data, and extracting vertical bridge deformation of the bridge.
The Ansys establishes a finite element model of the bridge, and firstly, geometric information of a main structural part of the bridge needs to be determined; then, determining material characteristics of the bridge, such as elastic modulus, poisson ratio, density and the like, based on the actual performance parameters of the bridge material; and finally, creating a finite element model of the bridge in ANSYS.
In the finite element model design process, complex environmental factors possibly existing in the bridge under the actual working condition, such as wind load, temperature change, earthquake and the like, need to be considered. These environmental factors can be modeled by setting corresponding loads and boundary conditions in the finite element model. The method comprises the steps of setting fluid-structure interaction conditions, including establishing a geometric model, dividing grids, setting a structure and a fluid model, and realizing coupling of fluid and the structure through an FSI module, so as to finally solve and analyze the influence of interaction on the structure. Fig. 3 is a diagram of a suspension bridge.
Importing the buffeting force time series data obtained in the step 3 and the created finite element model of the suspension bridge in the operation stage into ANSYS software; then, a simulation function of ANSYS software is operated, the buffeting force is simulated to generate wind-induced vibration to the suspension bridge within a certain time, and a buffeting force-bridge model solution is obtained, so that buffeting response data of the bridge are obtained.
And the motion equation of the bridge under the effect of buffeting force:
（11）；
（12）；
In the formula, M xx and M yy are respectively the quality matrixes of the bridge in the transverse direction and the bridge in the vertical direction, AndThe transverse and vertical accelerations of the bridge are respectively, C xx and C yy are respectively the damping matrices of the bridge in the transverse and vertical directions,AndThe transverse and vertical speeds of the bridge are respectively, K xx and K yy are respectively stiffness matrixes of the bridge in the transverse and vertical directions, u and v are respectively transverse and vertical displacements of the bridge, and F gx (t) and F gy (t) are respectively transverse and vertical wind loads caused by buffeting force.
After the result is calculated by the motion equation of the bridge under the effect of the buffeting force, the deformation of the bridge is stored, the result with larger deformation is tidied, and the data is stored for subsequent use. Fig. 4 is a schematic diagram of a vibration sample generated by the Ansys finite element software.
Step 5: taking the pulsating wind data generated in the step 2 as an input set and taking the bridge deformation calculated by the Ansys finite element model in the step 4 as an output set to form a data set; wherein the first 80% is used as training set and the last 20% is used as test set
And (3) sorting and summarizing the transverse bridge deformation data and the vertical bridge deformation data generated by the Ansys finite element model in the step (4), comparing the transverse bridge deformation data and the vertical bridge deformation data, and extracting the vertical bridge deformation as an output set of training data.
In the training process, training data and test data are input into a constructed network model in batches, and in order to be able to more fully understand the advantages of the model on processing long sequences, in this embodiment, time sequence data with a sequence length of 200-500-800-1000 are used as training data sets 1-4 respectively; transformer has stronger expressive power and adaptability in processing long sequences, global dependencies, and large-scale datasets; therefore, as the length of the sequence expands, the prediction accuracy of the bridge buffeting response slowly decreases.
When the same time sequence length is the same, the smaller the order of magnitude is, the influence of the transducer model on the learning degree of the internal connection of the data is generated, so that the prediction accuracy of the model is influenced, and therefore, the time sequence data with the sample order of magnitude of 50-100-150 and the general length of 200 under the condition of the sequence length of 1000 are respectively used as training data sets in the embodiment.
Step 6: constructing a transducer deep learning network, verifying the prediction correctness of the network, and performing super-parameter setting on the network before training;
Pre-training based on buffeting response of a simulated suspension bridge, and constructing a seven-layer convolutional neural network model of a transducer structure for processing sequence data, wherein the model comprises a plurality of TransformerBlock layers, and each layer comprises a self-attention mechanism, a feedforward network and layer normalization; the method aims at capturing the inherent characteristics and dynamic behaviors of the fluctuating wind speed and buffeting response; taking the pulsating wind speed sample generated in the step (2) simulation as input data, and taking the bridge buffeting response sample generated in the step (3) simulation as output data; the input data is subjected to feature extraction through three TransformerBlock, then feature integration is performed through two full-connection layers, and finally a pre-trained transducer model is output. The architecture of the transducer deep learning model is shown in FIG. 5.
Step 7: inputting the training set into a transducer deep learning network for predictive training to obtain a predictive model of the vertical deformation of the bridge structure, and then storing the trained model.
And (3) carrying out data remodeling and normalization on the buffeting response sample and the fluctuating wind speed sample by adopting a transducer deep learning model: the transducer deep learning model comprises an input layer, a plurality of TransformerBlock layers and an output layer; utilizing the input layer part to adjust the format of the data into a proper transducer model; the transducer model consists of a plurality of TransformerBlock, and is a key part for learning characteristic connection in a buffeting response sequence; the output layer comprises a plurality of layers, and the aim of the layers is to map the learned potential distribution to the output and predict to obtain a prediction model of the vertical deformation of the bridge structure.
The input layer of the transducer deep learning model is a layers layer, and during training, the input layer is used as the input of the model, the process defines the shape of pulsating wind data of the transducer, the transducer model is allowed to receive the input with a specific structure, training data is accepted for forward propagation of the model, and the process ensures that the transducer model can read the data normally.
The TransformerBlock layers contain a self-attention mechanism, a feed forward network, and layer normalization, where the self-attention mechanism selects multi-headed sub self-attention, which is used to learn the feature connections in the input sequence in each TransformerBlock. In each layer, the model further learns the characteristic relation between the time sequence of the fluctuating wind speed and the buffeting response through a neural network layer on the basis of self-attention; among them, self-Attention (Self-Attention) is the main object of study and is the focus of the transducer model. For the pulsatile wind speed time series TransformerBlock, the correlations with all keys are calculated, and then the values are weighted averaged using these correlations, layer normalization is performed, and the time correlation characteristics of the wind speed time course are deeply mined. The transducer model may better focus on the portions of the buffeting response time series data that are relevant to the current query. Parameter calculations are normalized in matrix form and layer as:
（13）；
（14）；
In the formula, Q, K and V respectively represent three parameter matrixes, X input is an input matrix, and W Q、K、V is a weight matrix corresponding to the three parameter matrixes; z is the output matrix of the self-attention layer; d k is a scaling factor to ensure that dot product does not cause problems of gradient explosion or gradient disappearance when the dimension is large; softmax is an activation function used to calculate the attention weights, ensuring that the sum of the attention weights of the model at different positions is 1, so that the model can focus on information at different positions in the input sequence; attention is a self-Attention mechanism operation.
The Feed Forward network (Feed Forward) is a two-layer fully connected layer composition, the first layer has an activation function Relu, the second layer does not use an activation function, and its structure introduces nonlinear characteristics through the activation function, allowing the model to learn more complex features and representations. The combination of linear transformation and nonlinear activation functions enables the model to better adapt to complex data distributions. The coding information matrix C is generated after the calculation of the layers TransformerBlock, and the coding information matrix C is applied to the subsequent output layer.
Step 8: and obtaining a buffeting response prediction result through a transducer model by adopting a training set sample, and calculating the mean value and standard deviation of the transducer model prediction result.
The performance of the model is evaluated using a validation set, and six model evaluation indices, mean square error (Mean Squared Error), mean absolute error (Mean Absolute Error), decision coefficient (R2), and variance thereof, are selected to evaluate the transform model predictive performance and generalization ability.
In summary, the method provides an effective bridge buffeting response prediction method, and the method realizes data support for real-time and accurate judgment of the bridge by predicting the bridge buffeting response based on a lightweight transducer model, and provides powerful support for timely finding and processing potential problems for operation management staff. The invention provides a time sequence processing method for the influence of pulsating wind on a structure, which has excellent processing on long-range dependency relationship because a transducer model has the advantage of parallel computation, and can better capture the space-time dynamics of structural vibration caused by the pulsating wind. Second, the method is based on a transducer model, because the self-attention mechanism allows the network to automatically focus on information at different locations in the input sequence, making it more adaptive. This provides insight into the subsequent study of process model migration, etc. Finally, because the method is based on a transducer model, its self-attention mechanism makes it somewhat interpretable. Engineers can understand the specific input features of the model's interest in the predictions by analyzing the model's attention weights to better understand the influencing factors of the structural vibrations. Fig. 5 is a diagram showing a construction of a transducer deep learning model.
Example verification:
And selecting a certain extra-large-span suspension bridge, and comparing the processing capacity and training effect of different models on the long sequence. The bridge has a total length of 2470.58 meters and a main span of (320+1196+320) meters. The longer the span of the bridge is, the more sensitive the bridge is to wind load, so that the research object is a midspan section, and the research wind speed is the average wind speed at the position 10.0 meters away from the bridge deck. The overall arrangement of the main bridge is shown in fig. 3. The bridge deck width is 33.44 meters, the design is two-way four lanes, and the road design grade is road grade I. The numerical model of the bridge was developed with ANSYS, all girders, piers and main towers were modeled by 3D girder elements with 6 degrees of freedom per node, while the tension ropes were modeled by 3D rod elements with 3 degrees of freedom per node. The cross-sectional area of the sling was 0.001283 x 4m 2, and the elastic modulus of the elastic material of the sling was 1.1e 11.
Setting the actual roughness as B-class z 0 =0.05, and the ground roughness angle or slope alfa=0.016 according to the technical specification of the actual project; c D、CL and C M are respectively a drag coefficient, a lift coefficient and a moment coefficient, and the values in the embodiment are respectively 1.4557, -0.0853 and 0.0419; the influence height h=3 in the wind speed vertical direction; setting the wind speed gradient to be 0.25s; k is 0.4.
Calculating: verifying buffeting response prediction accuracy of each model under action of pulsating wind of certain bridge
The present example considers the wind load to buffeting force as a deterministic process regardless of the influence of the uncertainty parameters. And the calculation example compares the results of the LSTM model and the transducer model to show the difference in accuracy of the two models in terms of calculation results.
Through analysis of the pulsating wind speed at a certain bridge, 200 groups of prepared pulsating wind speed data are calculated and converted into buffeting force, so that time-course data output after static and dynamic wind load calculation are obtained. The calculated step size of the pulse wind schedule data is 0.25s. Because the data length of the selected comparison is 200-500-800-1000, interpolation operation is used to ensure that the interpolation pulsation wind speed time sequence of the main span is the uniform length of 50s-125s-200s-250s and is arranged into a data set 1-4; meanwhile, in order to compare the prediction precision of two models under different sample number sets, the data sets with the sample number sets of 50-100-150-200 under the condition of the data length of 1000 are adopted in the calculation example, and are arranged into data sets of 4-7. In order to compare the processing capacity and convergence speed of the transducer model with that of the LSTM model over a long time sequence, the epoch super-parameter values of the model in this example are uniformly 150. Fig. 6 (a) -6 (d) and fig. 7 (a) -7 (d) show loss values of training phases of two models under different sequence lengths and different training sample sets, and fig. 8 (a) -8 (d) and fig. 9 (a) -9 (d) show prediction effects of two models under different sequence lengths and different training sample sets. It is noted that, overall, the LSTM model converges more rapidly, but the convergence effect is far less than the contemporary fransformer model; furthermore, when the sequence length is smaller, the training effect of the transducer model is equivalent to that of the LSTM model, and as the sequence length is increased, the LSTM model is slower in convergence of the loss value of the training stage of the long-time sequence, so that the inherent relation of the model can not be well learned, and the situation of memory forgetting occurs in the prediction of the actual value. In contrast, the transducer model has a faster convergence rate and a better prediction result when processing a long-term sequence.
Table 1 shows the mean and variance of MAE, mean Square Error (MSE) and decision coefficient (R 2) between all predicted data and corresponding real data for data sets 1-4 in the dataset. At a sequence length of 200, the error between the two models is not significant. When the sequence length is 500 or more, for the LSTM model, the average value of R 2 has negative number, generally, the value range of the R 2 determination coefficient is [0,1], and when R2 is 1, the model perfectly predicts the data; when R 2 is 0, the representation model cannot interpret the data variance. In this example, the R 2 decision coefficient is used to compare the performance of two different models, the closer the value is to 1, the more the variance of the data representing model interpretation, the better the performance; the reason for the occurrence of the negative number is that the prediction error of the obtained fitting function is larger than the prediction error of the function, namely the prediction error of the average value of all sample points, namely the prediction model is invalid, and the inherent relation of the data cannot be learned. Meanwhile, for the MAE variance and MSE variance transducer model, the values are more than ten times smaller than that of the LSTM model, and the transducer model is more stable.
TABLE 1 loss values at different sequence lengths
。
Table 2 shows the mean and variance of MAE, mean Square Error (MSE) and decision coefficient (R 2) between all predicted data and corresponding real data for different number set sizes of data sets 4-7 in the data set. From the table, it can be seen that the transducer model can still predict long-time sequences well under the condition of different quantity sets, and the prediction effect is also better and better.
Table 2 loss values at different number set sizes
。
In summary, the method performs pre-training based on buffeting response of the simulated suspension bridge, and constructs a seven-layer convolutional neural network model of a transducer structure for processing sequence data, wherein the model comprises a plurality of TransformerBlock layers, and each layer comprises a self-attention mechanism, a feedforward network and layer normalization; the method aims at capturing the inherent characteristics and dynamic behaviors of the fluctuating wind speed and buffeting response; taking the time series data of the pulsating wind speed as input data and taking the buffeting response of the bridge as output data; the input data is subjected to feature extraction through three TransformerBlock, then feature integration is performed through two full-connection layers, and finally a pre-trained transducer model is output. According to practical calculation examples, under the conditions of different time sequence lengths and the same sample number set, the prediction capability of the transducer model is gradually stronger than that of the LSTM model along with the continuous increase of the time sequence lengths, and all indexes of the transducer model have obvious advantages. Under the condition of the same time sequence length and different sample number sets, as the sample number sets are continuously increased, the prediction capability of the transducer model is stronger than the prediction effect of LSTM. In addition, compared with the LSTM model, the transducer model has faster convergence speed and finds the internal relation between data, which shows that the transducer model has higher efficiency under the condition of adapting and accurately predicting the variable condition.
Claims (5)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410321437.9A CN117910120B (en) | 2024-03-20 | 2024-03-20 | Buffeting response prediction method for wind-bridge system based on lightweight Transformer |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410321437.9A CN117910120B (en) | 2024-03-20 | 2024-03-20 | Buffeting response prediction method for wind-bridge system based on lightweight Transformer |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117910120A CN117910120A (en) | 2024-04-19 |
| CN117910120B true CN117910120B (en) | 2024-06-28 |
Family
ID=90682408
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410321437.9A Active CN117910120B (en) | 2024-03-20 | 2024-03-20 | Buffeting response prediction method for wind-bridge system based on lightweight Transformer |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117910120B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118153176B (en) * | 2024-05-09 | 2024-07-12 | 西华大学 | Tie bar tension optimization method based on transducer model and GWO algorithm |
| CN118940387A (en) * | 2024-10-12 | 2024-11-12 | 浙江大学 | Structural displacement response prediction method, device and medium based on convolutional neural network |
Family Cites Families (28)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP3884260B2 (en) * | 2001-10-16 | 2007-02-21 | 株式会社日立製作所 | Wind power generator |
| DE102007021632A1 (en) * | 2007-05-09 | 2008-11-20 | Ssb-Antriebstechnik Gmbh & Co. Kg | Electric drive |
| WO2011016857A2 (en) * | 2009-08-05 | 2011-02-10 | Elliott James C | Equipment and system for structure inspection and monitoring |
| CN105138782A (en) * | 2015-09-02 | 2015-12-09 | 上海大学 | Non-stationary pulse wind speed high-precision prediction method based on EEMD-ELM |
| CN105243259A (en) * | 2015-09-02 | 2016-01-13 | 上海大学 | Extreme learning machine based rapid prediction method for fluctuating wind speed |
| CN105205495A (en) * | 2015-09-02 | 2015-12-30 | 上海大学 | Non-stationary fluctuating wind speed forecasting method based on EMD-ELM |
| CN205595806U (en) * | 2016-04-26 | 2016-09-21 | 杨磊 | Wind -powered electricity generation field frequency division transmission of electricity controlling means |
| CN107330056B (en) * | 2017-06-29 | 2020-10-16 | 河北工业大学 | Wind power plant SCADA system based on big data cloud computing platform and operation method thereof |
| CN112020388B (en) * | 2018-02-26 | 2023-10-24 | 瑟莫特拉有限公司 | Cooling, heating and stabilizing humidity by humidity fluctuation |
| CN110046379B (en) * | 2019-03-04 | 2021-01-19 | 浙江大学 | Structure overall damage identification method based on space-frequency information |
| CN111337210B (en) * | 2020-02-22 | 2021-10-15 | 中南大学 | An online monitoring method for bridge displacement under wind |
| AU2020103212A4 (en) * | 2020-11-04 | 2021-01-14 | Azam, Farooque MR | IoT AND MACHINE LEARNING-BASED POWER DISTRIBUTION MANAGEMENT SYSTEM |
| CN113642414A (en) * | 2021-07-19 | 2021-11-12 | 北京航空航天大学 | A Transformer Model-Based Method for Predicting the Remaining Service Life of Rolling Bearings |
| CN119629197A (en) * | 2021-09-05 | 2025-03-14 | 汉熵通信有限公司 | Method and system for adjusting carrier frequency point, carrier bandwidth and the like |
| CN114067818B (en) * | 2021-11-08 | 2024-09-03 | 中国人民解放军火箭军士官学校 | Time domain flexible vibration sensor voice enhancement method and system |
| CN114611365A (en) * | 2022-03-22 | 2022-06-10 | 国网浙江省电力有限公司电力科学研究院 | Dynamics analysis method and system based on transmission tower structure under influence of pulsating wind |
| CN114937016A (en) * | 2022-05-25 | 2022-08-23 | 广西大学 | Bridge concrete crack real-time detection method and device based on edge calculation and Transformer |
| CN115292780B (en) * | 2022-07-14 | 2023-06-30 | 哈尔滨工业大学 | Prediction method of buffeting response of long-span bridges based on deep learning |
| CN115392360B (en) * | 2022-08-11 | 2023-04-07 | 哈尔滨工业大学 | Transformer-based temperature-response correlation pattern recognition and health diagnosis method for large bridges |
| CN115577436B (en) * | 2022-11-17 | 2023-03-24 | 成都理工大学 | A Combined Deep Learning Method for Solving Uncertain Structure Wind-Induced Vibration Response |
| CN116049638A (en) * | 2023-02-28 | 2023-05-02 | 广东电网有限责任公司 | Transformer vibration detection method, system, equipment and storage medium |
| CN116522600A (en) * | 2023-04-04 | 2023-08-01 | 西华大学 | Complex wind field wind speed simulation method, device, equipment and medium |
| CN116542912A (en) * | 2023-04-13 | 2023-08-04 | 昆明理工大学 | Flexible body bridge vibration detection model with multi-target visual tracking function and application |
| CN116757086A (en) * | 2023-06-28 | 2023-09-15 | 南京工业大学 | Impact load inversion method |
| CN117009770A (en) * | 2023-08-30 | 2023-11-07 | 安徽大学 | Bearing fault diagnosis method based on SDP and visual transducer codes |
| CN117435988A (en) * | 2023-10-13 | 2024-01-23 | 国网湖北省电力有限公司黄冈供电公司 | EEMD-GRU-transducer-based energy storage power station water motor unit vibration degradation prediction method |
| CN117172136B (en) * | 2023-11-02 | 2024-01-05 | 成都理工大学 | Vibration response prediction method for windmill bridge system based on SSA-LSTM algorithm |
| CN117669389B (en) * | 2024-01-31 | 2024-04-05 | 西华大学 | Random vibration analysis method of earthquake-vehicle-bridge system based on deep learning |
-
2024
- 2024-03-20 CN CN202410321437.9A patent/CN117910120B/en active Active
Non-Patent Citations (3)
| Title |
|---|
| 基于深度学习的桥梁技术状况预测研究;孙斌 等;中国交通信息化;20240313(第S1期);全文 * |
| 斜拉桥风致振动在Ansys 中的实现;高能祥;城市道桥与防洪;20200531(第5期);全文 * |
| 雷暴风作用下大跨度桥梁抖振响应智能预测研究;陶天友 等;中国公路学报;20230831;第36卷(第8期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117910120A (en) | 2024-04-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN117910120B (en) | Buffeting response prediction method for wind-bridge system based on lightweight Transformer | |
| CN110580328B (en) | A Restoration Method for Missing Monitoring Value of Groundwater Level | |
| Soleimani et al. | Artificial neural network application in predicting probabilistic seismic demands of bridge components | |
| Li et al. | A convolutional neural network‐based full‐field response reconstruction framework with multitype inputs and outputs | |
| CN114693005A (en) | Three-dimensional underground reservoir dynamic prediction method based on convolutional Fourier neural network | |
| Yao et al. | Wave height forecast method with multi-step training set extension LSTM neural network | |
| Fei et al. | Research on tunnel engineering monitoring technology based on BPNN neural network and MARS machine learning regression algorithm | |
| CN114662414B (en) | Oil reservoir production prediction method based on graph wavelet neural network model | |
| CN115577436B (en) | A Combined Deep Learning Method for Solving Uncertain Structure Wind-Induced Vibration Response | |
| Ni et al. | Traffic-induced bridge displacement reconstruction using a physics-informed convolutional neural network | |
| Li et al. | A k-nearest neighbor locally weighted regression method for short-term traffic flow forecasting | |
| CN114997027B (en) | A method of intelligently solving the random signal of vehicle-axle system | |
| Xia et al. | Hybrid approach to seismic reliability assessment of engineering structures | |
| CN115471679A (en) | Method and intelligent system for synchronously assimilating water level and flow of natural river | |
| CN112836789A (en) | A Dynamic Prediction Method of Ground Connecting Wall Deformation Based on Composite Neural Network Algorithm | |
| Xia et al. | Comparative analysis of typical mathematical modelling methods through model updating of a real-life bridge structure with measured data | |
| CN116956440A (en) | A method for optimizing the arrangement of concrete performance monitoring points in complex environments | |
| Peng et al. | A seismic response prediction method based on a self-optimized Bayesian Bi-LSTM mixed network for high-speed railway track-bridge system | |
| CN112199887A (en) | Semi-supervised resistivity inversion method and system based on adversarial generative network and pseudo-labeling | |
| CN102279852A (en) | Automatic oil deposit history fitting system and method based on genetic algorithm | |
| Li et al. | A weighted pattern recognition algorithm for short-term traffic flow forecasting | |
| Inzunza-Aragón et al. | Use of artificial neural networks and response surface methodology for evaluating the reliability index of steel wind towers | |
| CN114880767B (en) | Aero-engine residual service life prediction method based on attention mechanism Dense-GRU network | |
| CN106548259B (en) | The method of geometry that rotating machinery health status is predicted under completely cut off data qualification | |
| Zhong et al. | Multi-fidelity enhanced few-shot time series prediction model for structural dynamics analysis |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |