Bridge wind farm monitoring data cleaning method and system based on variational autoencoder

Abstract

The present invention discloses a method for cleaning bridge wind field monitoring data based on variational autoencoder, comprising: setting an anemometer sensor, obtaining wind speed u and wind direction α data at symmetrical positions on both sides of the bridge deck and generating a training data set; marking each pair of upstream and downstream data in the data set according to the average wind direction #imgabs0#; randomly splitting the X-Y data set with data marking into a training set and a verification set; building a training model based on the variational autoencoder; generating two model instances #imgabs1# and learning u independently x ,u y ; Establish the loss function, including reconstruction loss and KL divergence loss; import the data set into the model for training until the loss function is stable; input the disturbed data into the model to obtain the interference-free data predicted by the model, and then complete the data. The data cleaning method based on variational autoencoder proposed in this patent corrects the wind field monitoring data at the bridge site of the long-span bridge, providing reliable support for the analysis of wind field characteristics.

Description

Technical Field

[0001] This patent is applicable to the field of wind field monitoring in bridge engineering projects, especially under strong wind conditions and structural interference of the bridge deck itself. The wind speed sensors on the bridge deck face serious data loss and data anomalies, and the measured data need to be cleaned. Background Art

[0002] Since the stiffness and damping of long-span bridges are relatively low, they are prone to large deformation and vibration under wind loads. Therefore, the wind resistance of long-span bridge structures usually becomes a controlling factor in structural design. During operation, understanding the changes in wind load parameters borne by the structure is also helpful for making daily maintenance decisions. Therefore, it is of great practical significance to conduct long-term monitoring of the wind environment at the bridge site or adjacent bridge site, and to extract wind field characteristics and wind parameters based on on-site monitoring data.

[0003] The health monitoring system of the bridge structure is equipped with a large number of sensors at key locations of the structure (such as the mid-span and the quarter points). The wind speed sensor can monitor and record the wind field at the bridge site in real time. Ideally, the average wind parameters (average wind speed, average wind direction, and wind attack angle, etc.) and the fluctuating wind parameters (fluctuating wind turbulence intensity, turbulence integral scale, gust factor, etc.) can be directly extracted from the original observation data.

[0004] However, due to years of exposure to the service environment, electronic components gradually age, and wind speed sensors often experience serious data loss in strong wind environments. In order to avoid a large amount of data loss caused by the failure of a sensor, the structural health monitoring system usually installs wind speed sensors at symmetrical positions on the bridge deck relative to the bridge axis for mutual comparison and supplementation. However, under the shelter of the bridge deck structure, the structure itself and the wind field interfere with each other, and the data recorded by the sensor is usually the disturbance value after the interaction between the wind and the bridge deck. This effect is particularly obvious when the wind flow direction and the sensor are located on the opposite side of the bridge axis. Therefore, in order to ensure the accuracy of the analysis of the structural wind field characteristics, before using the symmetrical anemometer data for mutual complementation, it is necessary to consider this disturbance effect and perform data cleaning on the measured data.

[0005] The variational autoencoder is a generative deep learning model that generates similar new data by learning the potential representation of data. Its structure includes an encoder (which maps input data to a potential space) and a decoder (which generates data from the potential space). The variational autoencoder can learn the mapping relationship between disturbed data and non-interfered data, thereby completing the data cleaning task. Summary of the invention

[0006] This patent aims to solve the problem of interference of the deck structure of large-span bridges on the data recorded by wind speed sensors, and proposes a data cleaning method based on variational autoencoder to correct the interfered data and obtain correct data without interference.

[0007] The present invention provides a method for cleaning bridge wind field monitoring data based on variational autoencoder, comprising the following steps:

[0008] S1. Data preparation: Set up anemometer sensors to obtain wind speed u and wind direction α data at symmetrical positions on both sides of the bridge deck, and decompose it into wind speed u along the bridge axis x and wind speed u in the direction perpendicular to the bridge axis y , generate training data set;

[0009] S2. Data labeling: For each pair of upper and lower anemometers in the S1 data set, calculate the average wind direction Label the data:

[0010] Mark downstream data as non-interference data and upstream data as interfered data;

[0011] Mark the upstream data as non-interference data and the downstream data as interference data;

[0012] The disturbed data is used as input X, and the uninterrupted data is used as target Y;

[0013] S3. Dataset division: The XY dataset with data labels in S2 is randomly split into a training set and a validation set according to a certain ratio coefficient;

[0014] S4. Model construction: According to the vector features of the input X and output Y of the data set in S3, a training model is constructed based on the variational autoencoder;

[0015] S5. Model training: Generate two model instances Study independently x ,u y ; Establish the loss function, including reconstruction loss and KL divergence loss; and import the data set in S3 into the model in S4 for training until the loss function is stable;

[0016] S6. Data cleaning: Input the disturbed data into Obtaining unobstructed data for model predictions It can also perform data completion.

[0017] Furthermore, the method for obtaining the wind speed u and wind direction α data at the symmetrical positions on both sides of the bridge deck in step S1 is to measure the measured wind speed u and wind direction α at a basic time interval of 5-15 minutes. For any data, its length is:

[0018] N=600f (1)

[0019] Where f is the sampling frequency of the sensor; missing data, that is, data with a length less than N, are directly discarded.

[0020] Furthermore, the wind speed u along the bridge axis is decomposed in step S1. x and wind speed u in the direction perpendicular to the bridge axis y The method is:

[0021] u x =u cosα (2)

[0022] u y =u sinα (3)

[0023] In the formula, the wind direction α takes the bridge axis as the zero point and the clockwise direction is the positive direction; u x and u y Represent the wind direction components along the bridge axis and perpendicular to the bridge axis respectively.

[0024] Furthermore, the average wind direction in step S2 The calculation method is:

[0025]

[0026] In the formula, u x ,u y The average value of are the average wind speed and average wind direction within the time interval respectively.

[0027] Furthermore, the variational autoencoder structure in step S4 includes:

[0028] (1) Encoder: It consists of alternating one-dimensional convolutional layers and pooling layers to extract input features. Its output is the mean vector μ and variance vector σ of the latent variable z. 2 ;

[0029] (2) Decoder: It consists of one-dimensional deconvolution layers and upsampling layers to generate interference-free data from latent variables;

[0030] (3) Skip connection: occurs after each convolutional layer of the encoder and before each deconvolutional layer of the decoder, and is used to preserve the spatial information and detail features in the input data and generate high-quality output data;

[0031] (4) Latent space: The mean and variance of the encoder output are used to generate latent variables as low-dimensional feature representations of the input data. This process is achieved through reparameterization.

[0032]

[0033] z=μ+σ⊙ε (13)

[0034] Where ε is a vector sampled from a standard normal distribution and ⊙ represents element-wise multiplication.

[0035] Furthermore, the reconstruction loss in step S5 is calculated as follows:

[0036]

[0037] Where, L recon represents the reconstruction loss, M represents the total number of data entries, N represents the number of samples in the basic time interval, and v i , Represents the actual interference-free data and the output interference-free data vector, v i,j , Represents the elements in the actual data and output data vectors.

[0038] Furthermore, the KL divergence loss in step S5 is calculated as follows:

[0039]

[0040] Where, L KL represents the divergence loss, L represents the dimension of the latent variable, μ i ,σ i represent the mean and standard deviation of the latent variables, respectively.

[0041] Furthermore, the reconstruction loss and KL divergence loss are weighted and the total loss is obtained:

[0042] L Total =L recon +βL KL (16)

[0043] In the formula, β is the weighting coefficient.

[0044] As another aspect of the present invention, it also relates to a bridge wind farm monitoring data cleaning system based on a variational autoencoder, comprising:

[0045] The data acquisition unit is used to set an anemometer sensor to obtain the wind speed u and wind direction α data at the symmetrical positions on both sides of the bridge deck, and decompose it into the wind speed u along the bridge axis x and wind speed u in the direction perpendicular to the bridge axis y , generate training data set;

[0046] Data marking unit, used to calculate the average wind direction for each pair of upper and lower anemometers in the S1 data set Label the data:

[0047] Mark downstream data as non-interference data and upstream data as interfered data;

[0048] Mark the upstream data as non-interference data and the downstream data as interference data;

[0049] The disturbed data is used as input X, and the uninterrupted data is used as target Y;

[0050] The data set division unit is used to randomly divide the XY data set with data labels in S2 into a training set and a validation set according to a certain ratio coefficient;

[0051] A model building unit, used to build a training model based on a variational autoencoder according to the vector features of the input X and output Y of the data set in S3;

[0052] Model training unit, used to generate two model instances Study independently x ,u y ; Establish the loss function, including reconstruction loss and KL divergence loss; and import the data set in S3 into the model in S4 for training until the loss function is stable;

[0053] Data cleaning unit, used to input disturbed data into Obtaining unobstructed data for model predictions It can also perform data completion.

[0054] As another aspect of the present invention, it also relates to a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program is executed by a processor for a method for cleaning bridge wind farm monitoring data.

[0055] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0056] 1. The bridge wind field monitoring data cleaning method of the present invention corrects the wind field monitoring data at the bridge site of a large-span bridge through a data cleaning method based on a variational autoencoder, thereby eliminating the interference of the shielding effect of the bridge deck structure on the measured data, and can obtain corrected data close to the actual situation, thereby providing reliable support for subsequent wind field characteristic analysis.

[0057] 2. The bridge wind field monitoring data cleaning method of the present invention adopts a variational autoencoder, which has a simple structure, is easy to program, and is convenient for software development and embedding in other wind field analysis tools. This makes the method of the present invention highly feasible and practical in practical applications, and can provide an effective technical means for wind field monitoring and analysis in the field of bridge engineering, and has good application prospects.

[0058] 3. The data cleaning method for monitoring wind field of bridges of the present invention forms a complete set of data cleaning process, from data preparation, labeling, division, to model construction, training, and finally data cleaning and completion. Each link is closely connected and cooperates with each other. It has strong versatility and can be shared in monitoring tasks on the same type of bridges. This is of great significance for wind field monitoring in the field of bridge engineering, and can provide a universal and effective data processing method for wind field analysis of different bridges. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a schematic diagram of the process of data cleaning of measured data at the bridge site as described in this patent;

[0060] Figure 2 It is a schematic diagram of the arrangement of bridge deck sensors and disturbance effects described in the embodiment of this patent;

[0061] Figure 3 Schematic diagram of the structure of the variational autoencoder network described in the embodiment of this patent;

[0062] Figure 4 This is the correction result of the wind speed component and the synthetic amount in the embodiment of this patent;

[0063] Figure 5 The power spectrum density of the wind field calculated using disturbed and undisturbed data in the embodiment of this patent;

[0064] In the figure, ANEU is a wind speed sensor located upstream of the bridge deck; ANED is a wind speed sensor located downstream of the bridge deck. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be described clearly and completely in conjunction with the drawings in the embodiments of the present invention. The described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be designed and configured in various different ways.

[0066] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] It should be noted that similar symbols and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings.

[0068] Example 1

[0069] like Figure 1 As shown, this embodiment provides a method for cleaning bridge wind farm monitoring data based on a variational autoencoder, comprising:

[0070] S1. Data preparation:

[0071] (1) Anemometer sensor setting and data collection

[0072] The bridge structure used in this embodiment is provided with wind speed sensors on both the upstream and downstream sides of the span, the upstream side is ANEU and the downstream side is ANED. Figure 2 As shown in the figure. When the wind is generated from the upstream side, it is first recorded by ANEU, and then the airflow passes through the bridge deck and interacts with the bridge deck structure before being recorded by ANED. At this time, the data of ANEU can be considered as non-interference data, and the data of ANED is interference data. When the wind is generated from the downstream side, the process is reversed.

[0073] First, the measured wind speed u and wind direction α are divided into several pieces with a basic time interval of 10 minutes. For any piece of data, its length should be:

[0074] N=600f (1)

[0075] f is the sampling frequency of the sensor. During the data collection process, missing data, that is, data with a length less than N, are directly discarded to ensure the integrity and validity of the data, thereby generating an initial data set for subsequent processing.

[0076] (2) Wind speed data decomposition

[0077] Using the collected wind speed u and wind direction α data, the wind speed is decomposed into the wind speed u along the bridge axis according to the following formula: x and wind speed u in the direction perpendicular to the bridge axis y :

[0078] u x =u cosα (2)

[0079] The wind direction α takes the bridge axis as the zero point and the clockwise direction as the positive direction. The wind speed component u along the bridge axis is calculated by this formula: x .

[0080] u y =u sinα (3)

[0081] The wind speed component u in the direction perpendicular to the bridge axis is calculated y .

[0082] Complete the wind speed decomposition work in the data preparation stage and obtain the decomposed wind speed data set as the basic data for subsequent data marking.

[0083] If the wind speed vector is used directly as input without decomposition, the model may find it difficult to accurately capture the complex changes and interference of the wind field in different directions. The decomposed wind speed data allows the model to focus more on the data characteristics in each direction, which can improve the efficiency of data utilization, thereby improving the accuracy of the model's wind field data cleaning and more effectively correcting the disturbed data.

[0084] The decomposed wind speed data helps to extract wind field characteristics more accurately. In wind field analysis, the calculation and analysis of turbulence intensity, integral scale and other fluctuating wind parameters of wind speed components in different directions require clear directional information. x and u y These parameters can provide a more comprehensive understanding of the dynamic characteristics of the wind field and provide more detailed data support for the research on wind fields in bridge engineering.

[0085] S2. Data labeling stage:

[0086] (1) Calculation of average wind direction

[0087] For each pair of anemometer data at symmetrical positions on both sides of the bridge deck obtained during the data preparation phase, the average wind direction is first calculated. The calculation process is as follows:

[0088]

[0089] In the formula, u x ,u y The average value of are the average wind speed and average wind direction within the time interval respectively.

[0090] (2) Data labeling operation

[0091] The average wind direction calculated from Data labeling:

[0092] when When , the anemometer data located downstream of the bridge deck are marked as non-interference data, and the anemometer data located upstream of the bridge deck are marked as interference data.

[0093] when , the upstream data is marked as non-interference data and the downstream data is marked as interfered data.

[0094] The marked disturbed data is used as input X in the subsequent model training and data cleaning process, and the non-interference data is used as the target Y to build a set of XY data pairs for subsequent operations such as data set partitioning and model training.

[0095] The degree of interference of bridge structure on wind field is closely related to wind direction. Average wind direction The calculation helps to consider the actual situation of the interaction between wind and bridge structure. When wind blows towards the bridge from different directions, the wind field will be affected by different degrees of shielding, reflection, bypass, etc. due to the existence of the bridge's own structure (such as the bridge deck, bridge towers, bridge piers, etc.).

[0096] By calculating the average wind direction Based on this, a set of data pairs of interfered data X and non-interfered data Y is constructed, and then imported into the model for training and learning. This can more accurately learn the specific rules of interference of this bridge structure on the wind field, making the learned model more realistic.

[0097] S3. Dataset division stage

[0098] (1) Determination of the division ratio

[0099] Determine an appropriate ratio coefficient based on actual needs and data characteristics. For example, the common 70%-80% can be used as the training set ratio, and the remaining part can be used as the validation set ratio.

[0100] (2) Random Split Operation

[0101] The XY data set obtained after the data labeling stage is split using a random algorithm according to a ratio coefficient of 80%. A part of the data (the amount determined by the ratio) is used as a training set for the model training process to learn the mapping relationship between the disturbed data and the non-interfered data; the other part of the data is used as a validation set to verify and evaluate the performance of the model during the model training process to ensure the accuracy and generalization ability of the model, complete the data set division, and obtain two data sets: the training set and the validation set.

[0102] S4. Model building stage

[0103] Establish a variational autoencoder neural network model with skip connections. The specific structure is shown in the attached Figure 3As shown in the figure, the input and output of the network are both one-dimensional vectors with a length of 600.

[0104] (1) Encoder construction: Construct an encoder consisting of alternating one-dimensional convolutional layers and pooling layers. The one-dimensional convolutional layer is used to perform feature extraction on the input data (i.e., the disturbed data X). By setting different parameters such as the convolution kernel and convolution step size, various feature information in the data is extracted. The pooling layer downsamples the feature map output by the convolutional layer to reduce the amount of data while retaining the main features. Its output is the mean vector μ and variance vector σ of the latent variable z. 2 , providing a basis for the subsequent generation of latent space.

[0105] (2) Decoder construction: Construct a decoder consisting of a one-dimensional deconvolution layer and an upsampling layer. The one-dimensional deconvolution layer can upsample the latent variables and gradually restore the dimension of the data. The upsampling layer further processes the data so that it eventually generates output data similar to the original uninterrupted data, thereby realizing the function of generating uninterrupted data from the latent space.

[0106] (3) Skip connection setting: Skip connections are set after each convolutional layer of the encoder and before each deconvolutional layer of the decoder. This connection method can directly transfer feature information of different levels in the encoder to the corresponding level of the decoder, retaining the spatial information and detail features in the input data, helping to generate high-quality output data and improving the model's reconstruction effect on data.

[0107] (4) Latent space generation: The mean vector μ and variance vector σ output by the encoder 2 , use the reparameterization method to generate the latent variable z. The specific process is:

[0108]

[0109] z=μ+σ⊙ε (21)

[0110] First, from the standard normal distribution The vector ε is sampled from the matrix, and then the latent variable z is generated according to the formula z=μ+σ⊙ε (where ⊙ represents element-level multiplication). The latent variable is used as a low-dimensional feature representation of the input data and plays a key role in the model learning and data generation process.

[0111] Following these four steps, the structure of the variational autoencoder is completed to form a complete model framework.

[0112] S5. Model training phase

[0113] (1) Model instance generation and initialization

[0114] Generate two model instances and Used to learn the wind speed along the bridge axis and wind speed perpendicular to the bridge axis The two model instances are initialized, including setting the initial parameters of the model, such as the weights and biases of the convolutional and deconvolutional layers, to prepare for model training.

[0115] (2) Loss function establishment

[0116] 2.1 Reconstruction loss calculation

[0117] Establish the reconstruction loss function L recon , and its calculation formula is:

[0118]

[0119] Where, L recon represents the reconstruction loss, M represents the total number of data entries, N represents the number of samples in the basic time interval, and v i , Represents the actual interference-free data and the output interference-free data vector, v i,j , Represents the elements in the actual data and output data vectors.

[0120] This loss function is used to measure the degree of difference between the model output data and the actual interference-free data, prompting the model to minimize this difference during training and improve the accuracy of data reconstruction.

[0121] 2.2 KL divergence loss calculation

[0122] Establish the KL divergence loss function L KL , and its calculation formula is:

[0123]

[0124] Where, L KL represents the divergence loss, L represents the dimension of the latent variable, μ i ,σ i represent the mean and standard deviation of the latent variables, respectively.

[0125] KL divergence loss is used to constrain the distribution of latent variables to make them close to the standard normal distribution, which helps the model learn more effective data representation and improve the generalization ability of the model.

[0126] 2.3 Determination of total loss function

[0127] The total loss function is obtained by weighted summing the reconstruction loss and KL divergence loss:

[0128] L Total =Lrecon +βL KL (twenty four)

[0129] Among them, β is the weighting coefficient. By adjusting the value of β, the proportion of reconstruction loss and KL divergence loss in the total loss can be balanced, thereby optimizing the training effect of the model.

[0130] (3) Model training process

[0131] Import the training set data obtained in the dataset division phase into the constructed model instance and During the training process, the model obtains the latent variable z through the encoder according to the input interfered data X, and then generates the predicted interference-free data u through the decoder. x and u y , and then calculate the total loss function L Total The back propagation algorithm is used to update and adjust the parameters of the model according to the total loss function, and the parameters of the model are continuously optimized so that the value of the total loss function gradually decreases and tends to be stable. During the training process, you can also set training parameters such as the number of training rounds (epochs) and learning rate to control the progress and speed of training. At the same time, the validation set data is used to monitor and evaluate the performance of the model during the training process, such as calculating the loss value, accuracy and other indicators on the validation set. When the performance of the model on the validation set no longer improves or reaches the preset stopping condition, stop the model training and obtain the trained model. and Used for subsequent data cleaning work.

[0132] S6. Data cleaning stage

[0133] (1) Data input and prediction

[0134] Input the disturbed data that needs to be cleaned into the trained model instance and Model According to the input disturbed wind speed data along the bridge axis, through its internal encoder-decoder structure, the uninterrupted wind speed data along the bridge axis u is predicted and output x ;Model The input vertical bridge axis wind speed data with interference is processed, and the vertical bridge axis wind speed data without interference u is predicted and output. y .

[0135] (2) Data completion and cleaning result acquisition

[0136] The model and The predicted interference-free data u x and uy The cleaned data is used to complete and correct the missing or abnormal parts in the original disturbed data. In this way, the interference of the shielding effect of the bridge deck structure on the measured data is eliminated, and the corrected data close to the actual situation is obtained, completing the cleaning of the bridge wind field monitoring data and providing reliable data support for subsequent wind field characteristic analysis and other work.

[0137] At the same time, this solution is convenient for software development and can be embedded in other wind farm analysis tools as a data cleaning method. For example, the cleaned wind farm data can be combined with the bridge structure safety early warning system to establish an early warning model based on accurate wind farm data. By setting reasonable thresholds and early warning rules, when the wind farm data changes abnormally or may threaten the safety of the bridge structure, early warning signals will be issued in time to remind relevant personnel to take corresponding measures to ensure the safe operation of the bridge.

[0138] In other preferred embodiments, visualization tools are also developed to display the cleaned data and the process and results of data cleaning in an intuitive and easy-to-understand manner. For example, by drawing the spatiotemporal distribution diagram of wind farm data, the wind speed and direction change curve, and the comparison diagram before and after data cleaning, bridge engineers and managers can more clearly understand the characteristics of the wind farm and the data quality, which is convenient for data analysis and decision-making.

[0139] Not only in the bridge field, in other preferred embodiments, the invention is applied to other fields related to wind farms, such as wind farm monitoring data cleaning of wind farms. Wind speed sensors in wind farms are often subject to various interferences, resulting in a decrease in data quality, affecting the evaluation of power generation efficiency and the control strategy of wind turbines. The method of the present invention can be applied to data cleaning of wind farms after appropriate adjustment and optimization, to improve the accuracy of wind farm data and provide support for the optimized operation of power farms.

[0140] Example 2

[0141] This embodiment provides a bridge wind farm monitoring data cleaning system based on a variational autoencoder, including:

[0142] (1) Data acquisition unit:

[0143] 1.1 Sensor layout and connection

[0144] Anemometer sensors are carefully arranged symmetrically on both sides of the bridge deck at key locations of the bridge structure, such as the mid-span and the quarter points. These sensors are connected to the data acquisition unit through a dedicated data transmission line to ensure that the data can be transmitted to the data acquisition unit stably and accurately.

[0145] The selection and installation of sensors must meet the accuracy requirements and environmental adaptability requirements of bridge wind field monitoring, be able to work normally in strong winds, complex climates and long-term exposure service environments, and reliably collect wind speed u and wind direction α data.

[0146] 1.2 Data collection and preprocessing

[0147] The data acquisition unit collects the wind speed and direction data transmitted by the anemometer sensor in real time at a set time interval (10 minutes is the basic time interval). During the collection process, for each piece of data, the required data length is calculated according to the formula N=600f (where f is the sampling frequency of the sensor). Data with a length less than N, that is, missing data, is directly eliminated to ensure the integrity and validity of the data.

[0148] The collected wind speed u and wind direction α data are then transmitted to the data processing module, where the data are processed according to the formula u x =ucosαandu y =usinα decomposes the wind speed into the wind speed u along the bridge axis x and wind speed u in the direction perpendicular to the bridge axis y , complete the preliminary preprocessing of the data and generate a training data set that can be used by subsequent units.

[0149] (2) Data marking unit:

[0150] 2.1 Average wind direction calculation module

[0151] The data labeling unit first calls the training data set generated by the data acquisition unit, and uses a special calculation module to calculate the average wind direction for each pair of anemometer data at symmetrical positions on both sides of the bridge deck. The calculation process is as follows:

[0152] First calculate u separately x and u y The average value of and

[0153] Then according to the formula Calculate the average wind speed within the time interval

[0154] Finally, the formula and Calculate the average wind direction

[0155] 2.2 Data Tagging Execution Module

[0156] The average wind direction calculated from The data tagging execution module performs data tagging operations:

[0157] when When the anemometer data at the downstream of the bridge deck is marked as non-interference data, the anemometer data at the upstream of the bridge deck is marked as interference data.

[0158] when , the upstream data is marked as non-interference data and the downstream data is marked as interfered data.

[0159] The marked disturbed data is used as input X, and the non-interference data is used as target Y to construct a set of XY data pairs and transmit them to the data set partitioning unit for the next step of data set partitioning operation.

[0160] (3) Dataset division unit

[0161] 3.1 Ratio setting module

[0162] The data set division unit includes a ratio setting module, in which a suitable ratio coefficient is set according to actual application requirements and data characteristics. For example, a common method is to set the data set to 70%-80% as the training set ratio and the remaining part as the validation set ratio.

[0163] 3.2 Data Splitting Module

[0164] The data splitting module uses a random algorithm to split the XY data set transmitted by the data labeling unit according to the ratio coefficient determined by the ratio setting module, and uses a part of the data (the number determined by the ratio) as the training set and the other part of the data as the verification set.

[0165] The training set data is transmitted to the model training unit for use in the model training process, so that it can learn the mapping relationship between the interfered data and the non-interfered data; the validation set data is transmitted to the model training unit for verifying and evaluating the performance of the model during the model training process, ensuring the accuracy and generalization ability of the model, completing the data set division work, and providing a suitable data set for subsequent model training.

[0166] (4) Model building unit

[0167] 4.1 Encoder Building Blocks

[0168] The model building unit first builds the encoder. The encoder consists of alternating one-dimensional convolutional layers and pooling layers. The one-dimensional convolutional layer extracts features from the input data by setting different parameters such as the convolution kernel size and convolution step size, which can effectively capture local feature information in the data. The pooling layer uses a suitable pooling method (such as maximum pooling or average pooling) to downsample the feature map output by the convolutional layer to reduce the amount of data while retaining the main features. Its output is the mean vector μ and variance vector σ of the latent variable z.2 , providing a basis for the generation of latent space.

[0169] 4.2 Decoder Building Blocks

[0170] The decoder building module builds a decoder consisting of a one-dimensional deconvolution layer and an upsampling layer. The one-dimensional deconvolution layer upsamples the latent variables by setting the corresponding deconvolution parameters and gradually restores the dimension of the data. The upsampling layer further processes the data so that it eventually generates output data similar to the original uninterrupted data, thereby realizing the function of generating uninterrupted data from the latent space.

[0171] 4.3 Jump connection setting module

[0172] The skip connection setting module sets a skip connection after each convolution layer of the encoder and before each deconvolution layer of the decoder. This connection method can directly transfer the feature information of different levels in the encoder to the corresponding level of the decoder, retain the spatial information and detail features in the input data, help generate high-quality output data, and improve the model's reconstruction effect on data.

[0173] 4.4 Latent Space Generation Module

[0174] The latent space generation module uses the mean vector μ and variance vector σ output by the encoder 2 , generate the latent variable z through the reparameterization method. The specific process is: first, from the standard normal distribution The vector ε is sampled from the vector ε, and then the latent variable z is generated according to the formula z=μ+σ⊙ε (where ⊙ represents element-level multiplication). The latent variable is used as a low-dimensional feature representation of the input data. It plays a key role in the model learning and data generation process, completes the structure construction of the variational autoencoder, forms a complete model framework, and transmits the constructed model to the model training unit.

[0175] (5) Model training unit

[0176] 5.1 Model instance generation and initialization module

[0177] After the model training unit receives the model framework from the model building unit, it uses the model instance generation and initialization module to generate two model instances. and Used to learn the wind speed along the bridge axis and wind speed perpendicular to the bridge axis The two model instances are initialized, including setting the initial parameters of the model, such as the weights and biases of the convolutional and deconvolutional layers, to prepare for model training.

[0178] 5.2 Loss Function Establishment Module

[0179] Reconstruction loss calculation submodule: Establish the reconstruction loss function L recon , and its calculation formula is: Where M represents the total number of data entries, N represents the number of samples in the basic time interval, and v i , Represents the actual interference-free data and the output interference-free data vector, v i,j , Represents the elements in the actual data and output data vectors. This loss function is used to measure the difference between the model output data and the actual interference-free data, so as to minimize the difference during the training process and improve the accuracy of data reconstruction.

[0180] KL divergence loss calculation submodule: Establish the KL divergence loss function L KL , the calculation formula is: Where L represents the dimension of the latent variable, μ i ,σ i Represent the mean and standard deviation of the latent variables respectively. KL divergence loss is used to constrain the distribution of latent variables to make it close to the standard normal distribution, which helps the model learn more effective data representation and improve the generalization ability of the model.

[0181] Total loss function determination submodule: The reconstruction loss and KL divergence loss are weighted and summed to obtain the total loss function L Total , the calculation formula is: L Total =L recon +βL KL , where β is the weighting coefficient. By adjusting the value, the proportion of reconstruction loss and KL divergence loss in the total loss can be balanced, thereby optimizing the training effect of the model.

[0182] 5.3 Model Training Execution Module

[0183] The model training execution module imports the training set data from the data set partitioning unit into the model instance and During the training process, the model obtains the latent variable z through the encoder according to the input interfered data X, and then generates the predicted interference-free data u through the decoder. x and u y , and then calculate the total loss function L TotalThe back propagation algorithm is used to update and adjust the parameters of the model according to the total loss function, and the parameters of the model are continuously optimized so that the value of the total loss function gradually decreases and tends to be stable. During the training process, training parameters such as the number of training rounds (epochs) and learning rate can also be set to control the progress and speed of training. At the same time, the performance of the model during the training process is monitored and evaluated using the validation set data, such as calculating the loss value, accuracy and other indicators on the validation set. When the performance of the model on the validation set no longer improves or reaches the preset stop condition, the model training is stopped, and the trained model and are obtained, and are transmitted to the data cleaning unit.

[0184] (6) Data cleaning unit

[0185] 6.1 Data Input and Prediction Module

[0186] The data cleaning unit receives the trained data from the model training unit. and Input the disturbed data that needs to be cleaned into the trained model instance and Model According to the input disturbed wind speed data along the bridge axis, through its internal encoder-decoder structure, the uninterrupted wind speed data along the bridge axis u is predicted and output x ;Model The input vertical bridge axis wind speed data with interference is processed, and the vertical bridge axis wind speed data without interference u is predicted and output. y .

[0187] 6.2 Data completion and cleaning result output model

[0188] The data completion and cleaning result output module will model and The predicted interference-free data u x and u y The cleaned data is used to complete and correct the missing or abnormal parts in the original disturbed data. In this way, the interference of the shielding effect of the bridge deck structure on the measured data is eliminated, and the corrected data close to the actual situation is obtained, completing the cleaning of the bridge wind field monitoring data and providing reliable data support for subsequent wind field characteristic analysis and other work.

[0189] The present application also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of any of the above-mentioned methods for constructing a digital twin of an ultra-large anchorage can be implemented.

[0190] The computer-readable storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0191] For an introduction to the computer-readable storage medium provided in this application, please refer to the above method embodiment, and this application will not go into details here.

[0192] The above describes in detail the preferred specific embodiments of the present invention. It should be understood that ordinary technicians in this field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in this technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the existing technology should be within the scope of protection determined by the claims.

Claims

1. A bridge wind field monitoring data cleaning method based on variational autoencoder, characterized in that: include: S1. Set up anemometer sensors to obtain wind speed u and wind direction α data at symmetrical positions on both sides of the bridge deck, and decompose it into wind speed u along the bridge axis x and wind speed u in the direction perpendicular to the bridge axis y , generate training data set; S2. For each pair of upper and lower anemometers in the S1 data set, calculate the average wind direction Label the data: Mark downstream data as non-interference data and upstream data as interfered data; Mark the upstream data as non-interference data and the downstream data as interference data; The disturbed data is used as input X, and the uninterrupted data is used as target Y; S3. Randomly split the XY data set with data labels in S2 into a training set and a validation set according to a certain ratio coefficient; S4. Build a training model based on the variational autoencoder according to the vector features of the input X and output Y of the data set in S3; S5. Generate two model instances Study independently x ,u y ; Establish the loss function, including reconstruction loss and KL divergence loss; and import the data set in S3 into the model in S4 for training until the loss function is stable; S6. Input the disturbed data into Obtaining unobstructed data for model predictions It can also perform data completion.

2. The bridge wind field monitoring data cleaning method based on variational autoencoder according to claim 1 is characterized in that: The method for obtaining the wind speed u and wind direction α data at the symmetrical positions on both sides of the bridge deck in step S1 is to measure the measured wind speed u and wind direction α at a basic time interval of 5-15 minutes. For any data, its length is: N=600f (1) Where f is the sampling frequency of the sensor; missing data, that is, data with a length less than N, are directly discarded.

3. The bridge wind field monitoring data cleaning method based on variational autoencoder according to claim 2 is characterized in that: The step S1 is decomposed into the wind speed u along the bridge axis. x and wind speed u in the direction perpendicular to the bridge axis y The method is: in x =ucosα (2) u y =usinα (3) In the formula, the wind direction α takes the bridge axis as the zero point and the clockwise direction is the positive direction; u x and u y Represent the wind direction components along the bridge axis and perpendicular to the bridge axis respectively.

4. The method for cleaning bridge wind farm monitoring data based on variational autoencoder according to any one of claims 1 to 3, characterized in that: The average wind direction in step S2 The calculation method is: In the formula, u x ,u y The average value of are the average wind speed and average wind direction within the time interval respectively.

5. The method for cleaning bridge wind farm monitoring data based on variational autoencoder according to any one of claims 1 to 3, characterized in that: The variational autoencoder structure in step S4 includes: (1) Encoder: It consists of alternating one-dimensional convolutional layers and pooling layers to extract input features. Its output is the mean vector μ and variance vector σ of the latent variable z. 2 ; (2) Decoder: It consists of one-dimensional deconvolution layers and upsampling layers to generate interference-free data from latent variables; (3) Skip connection: occurs after each convolutional layer of the encoder and before each deconvolutional layer of the decoder, and is used to preserve the spatial information and detail features in the input data and generate high-quality output data; (4) Latent space: The mean and variance of the encoder output are used to generate latent variables as low-dimensional feature representations of the input data. This process is achieved through reparameterization. z=μ+σ⊙ε (5) Where ε is a vector sampled from a standard normal distribution and ⊙ represents element-wise multiplication.

6. The method for cleaning bridge wind farm monitoring data based on variational autoencoder according to any one of claims 1 to 3, characterized in that: The reconstruction loss in step S5 is calculated as follows: Where, L recon represents the reconstruction loss, M represents the total number of data entries, N represents the number of samples in the basic time interval, and v i , Represents the actual interference-free data and the output interference-free data vector, v i,j , Represents the elements in the actual data and output data vectors.

7. The method for cleaning bridge wind farm monitoring data based on variational autoencoder according to claim 6 is characterized in that: The KL divergence loss in step S5 is calculated as follows: Where, L KL represents the divergence loss, L represents the dimension of the latent variable, μ i ,σ i represent the mean and standard deviation of the latent variables, respectively.

8. The bridge wind field monitoring data cleaning method based on variational autoencoder according to claim 7 is characterized in that: The reconstruction loss and KL divergence loss are weighted and the total loss is obtained: L Total =L recon +βL KL (8) In the formula, β is the weighting coefficient.

9. A method and system for cleaning bridge wind field monitoring data based on variational autoencoder, characterized in that: include: The data acquisition unit is used to set an anemometer sensor to obtain the wind speed u and wind direction α data at the symmetrical positions on both sides of the bridge deck, and decompose it into the wind speed u along the bridge axis x and wind speed u in the direction perpendicular to the bridge axis y , generate training data set; Data marking unit, used to calculate the average wind direction for each pair of upper and lower anemometers in the S1 data set Label the data: Mark downstream data as non-interference data and upstream data as interfered data; Mark the upstream data as non-interference data and the downstream data as interference data; The disturbed data is used as input X, and the uninterrupted data is used as target Y; The data set division unit is used to randomly divide the XY data set with data labels in S2 into a training set and a validation set according to a certain ratio coefficient; A model building unit, used to build a training model based on a variational autoencoder according to the vector features of the input X and output Y of the data set in S3; Model training unit, used to generate two model instances Study independently x ,u y ; Establish the loss function, including reconstruction loss and KL divergence loss; and import the data set in S3 into the model in S4 for training until the loss function is stable; Data cleaning unit, used to input disturbed data into Obtaining unobstructed data for model predictions It can also perform data completion.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is executed by a processor to implement the method for cleaning bridge wind farm monitoring data as described in claims 1-8.

Images