A method for fine reconstruction of wind field in radar detection blind area by anemometer cooperative networking

Abstract

The application discloses a radar detection blind area wind field fine reconstruction method of anemometer cooperative networking, and specific steps are as follows: firstly, real-time separation of vibration displacement caused wind speed measurement deviation eliminates vibration interference, data automatic diagnosis and repair are carried out based on machine learning, and wind speed data integrity and measurement continuity are optimized; secondly, a supersonic anemometer array optimization arrangement method facing global wind field perception is established, time and space consistency calibration of multi-sensor data is realized, and the precision and reliability of wind speed measurement are improved; finally, data-physical fusion bridge flow field high-resolution intelligent reconstruction is carried out, and the data integrity of a complex flow field region is improved. The reconstruction method effectively cooperates the wind measurement laser radar and the anemometer, can play the advantages of high-frequency single-point measurement and wide-area space scanning of the two, and is expected to realize high-precision reconstruction of the global flow field of the bridge, and provide reliable support for guaranteeing the wind resistance safety of the long-span bridge.

Description

Technical Field

[0001] This invention belongs to the field of bridge wind engineering measurement, and specifically relates to a method for fine reconstruction of wind field in radar detection blind zone using anemometer collaborative networking. Background Technology

[0002] In the safety monitoring of long-span bridges, accurate measurement of wind field distribution is crucial for bridge wind-resistant design, health monitoring, and traffic safety management. Existing wind speed measurement methods mainly include fixed anemometers and LiDAR (LiDAR) systems. However, single measuring devices have many limitations in practical applications, making it difficult to comprehensively and accurately acquire information about the complex flow field around the bridge. Fixed anemometers are typically deployed at key locations on the bridge to provide local wind speed data. These devices offer high measurement accuracy and are less affected by the environment, but their measurement range is limited, failing to fully reflect the changing trends of the three-dimensional flow field around the bridge. Furthermore, fixed anemometers cannot effectively capture local turbulence, wind shear, and sudden wind speed changes in large-scale wind fields. LiDAR systems can provide wind speed measurement information over a wide area, suitable for global flow field monitoring. However, because LiDAR relies on the reflection characteristics of a light beam, it has blind spots around the bridge structure, especially in obstructed areas around bridge towers, main beams, and the anemometer itself. In addition, LiDAR's measurement frequency is relatively low, making it difficult to capture high-frequency wind speed fluctuations in real time.

[0003] Current research has gradually developed the concept of multi-sensor collaborative measurement, which involves using anemometer array networking technology combined with LiDAR data to optimize wind field reconstruction accuracy. However, in practical applications, the rational deployment of anemometers, spatiotemporal consistency of networking, and data fusion still face challenges. Traditional deployment methods often rely on experience-based site selection, lacking scientific basis and failing to adapt to the wind field characteristics of different bridge environments. Furthermore, due to the difference in measurement scale between anemometers and LiDAR, establishing a spatial matching relationship between the two to compensate for radar detection blind spots and improve the completeness of flow field reconstruction remains a critical issue that urgently needs to be addressed. Summary of the Invention

[0004] The purpose of this invention is to provide a method for fine reconstruction of wind field in radar detection blind zone using anemometers in collaborative networking, thereby achieving high-precision wind field reconstruction and improving the wind resistance safety of long-span bridges.

[0005] To achieve the above objectives, the solution of the present invention is:

[0006] A method for fine reconstruction of wind field in radar detection blind zone using anemometer collaborative networking includes the following steps:

[0007] Step 1: Preprocess the anemometer data and repair any abnormal data.

[0008] Step 2: Based on the flow characteristics of the bridge surface and the surrounding wind field, design the optimal ultrasonic anemometer array scheme.

[0009] Step 3: Establish a high-resolution intelligent reconstruction model of the bridge flow field based on data-physics fusion.

[0010] In step 1 above, a wavelet packet-Kalman filter cascade algorithm is used to separate the wind speed measurement deviation caused by vibration displacement;

[0011] First, the wind speed signal S(t) is subjected to J-order wavelet packet decomposition to obtain the sub-band signal.

[0012]

[0013] in These are wavelet packet basis functions;

[0014] Translate wavelet packet signal Further decomposed into several modal components u i (t) represents the signal components in different frequency bands:

[0015]

[0016] Where H(.) denotes the Hilbert transform, ω i The center frequency of each mode;

[0017] Introducing non-Gaussian noise modeling and using Kalman filtering correction, a state-space model is constructed:

[0018] X t+1 =AX t +W t Z t =HX t +V t

[0019] Among them, X t+1 For actual wind speed conditions, Z t To observe wind speed data, W t V t It is Gaussian white noise.

[0020] In step 1 above, abnormal data is identified using the following method:

[0021] The wind speed signal S(t) is converted into a Grammy angle field image, and a generative adversarial network is used to learn the normal wind speed data distribution, including:

[0022] Generator network G: Input noise vector z, output generated samples

[0023]

[0024] Discriminator Network D: Input is a Grammy angle field image transformed from the real wind speed distribution or generated samples. Output the probability of being true or false:

[0025] D(S)∈[0,1)

[0026] Through the optimal adversarial loss function:

[0027]

[0028] Train a generative adversarial network to generate samples that conform to the wind speed measurement distribution;

[0029] Construct an encoder-decoder network, input wind speed signal S(t), output reconstructed signal:

[0030]

[0031] Calculate reconstruction error If E(t) exceeds the control chart threshold, the wind speed is determined to be abnormal.

[0032] In step 1 above, the abnormal data is repaired based on the improved Bayesian dynamic linear model BDLM, which includes four steps: model order selection, nonlinear state transition modeling, variational Bayesian inference optimization, and abnormal data self-repair.

[0033] Model order selection: Construct a set of K candidate BDLM models with different orders {M1, M2, ..., M}. k The model score is calculated based on the Akaike Information Content Criterion (AIC) and the Root Mean Square Error (RMSE), and then fused into a normalized score index S. k S k A larger value indicates a better model order.

[0034]

[0035]

[0036]

[0037] In the formula: p k The number of parameters in the model; y is the maximum likelihood estimate of the model on the data; i These are actual observations; The model's predicted output is given by N; N is the total number of samples; exp(·) is the exponential function; min j AIC j This represents the minimum AIC value among all candidate models; ε is a very small positive number to prevent the denominator from being zero.

[0038] Multiple candidate models are weighted using Bayesian model average (BMA) to calculate the weighted prediction value:

[0039]

[0040] In the formula, This is the model's predicted output; This is the final predicted value after weighted fusion;

[0041] Nonlinear state transition modeling: To address the nonlinear characteristics of wind speed measurement errors, a nonlinear term f(X) is introduced into the state equation. t Construct a generalized state transition model:

[0042] X t =F t X t-1 +G t +f(X t-1 )+W t

[0043] in,

[0044] f(X t )=αtanh(βX t )

[0045] In the formula, α and β are the parameters to be optimized;

[0046] Variational Bayesian inference optimization: Optimizing the objective function using variational inference.

[0047]

[0048] In the formula, D KL (.) represents the KL divergence, which can be optimized by adjusting the variational distribution q(X) to improve the model convergence speed;

[0049] Anomaly self-repair: An improved BDLM is used to model and predict the normal temporal distribution of wind speed. After detecting outliers, an optimal autoregressive filter is employed to repair the anomaly data.

[0050]

[0051] In the formula, λ∈(0,1) is the dynamic adjustment factor.

[0052] The specific process of step 2 above is as follows:

[0053] Let the Eulerian coordinate system of the wind field in the bridge area be represented as:

[0054] V(x,y,z,t)=(u(x,y,z,t),v(x,y,z,t),w(x,y,z,t))

[0055] Where x, y, and z represent the three-dimensional spatial location of the wind speed measurement; t represents the time dimension of the wind speed; (u, v, w) are the three components of the wind speed; let the set of anemometer locations be P = {p1, p2, ..., p...} n}, where each sensor p i It has a set of performance parameters:

[0056] Θ i ={R i ,S i ,T i E i}

[0057] Where: Ri is the sensor measurement accuracy; S i T represents the spatial resolution of the sensor. i E represents the sensor signal transmission delay. i For sensor power consumption;

[0058] Define the optimization objective function:

[0059]

[0060]

[0061] in: Indicates the gradient intensity of the wind speed field; Information entropy; Represents signal transmission delay loss; w1, w2, and w3 are weighting coefficients; τ1 is the wind speed gradient threshold; τ2 is the lower limit of information entropy; τ3 is the maximum permissible signal transmission delay; E max This represents the upper limit of total energy consumption.

[0062] To address the aforementioned multi-objective optimization problem, a genetic algorithm is employed to determine the optimal ultrasonic anemometer array scheme.

[0063] The specific process of step 3 above is as follows:

[0064] Step 31: Perform scale matching between the high-frequency local data of the anemometer and the low-frequency large-area data of the lidar to establish a global alignment framework for the measured data of multiple devices.

[0065] Step 32: To address the radar detection blind zone problem, perform spatial interpolation of the flow field and dynamically correct the spatial interpolation results based on experimental calibration parameters.

[0066] Step 33: Combine the physical characteristics of the flow field with the measured data to reconstruct a high-precision three-dimensional flow field, and verify it with the measured data.

[0067] The specific process of step 31 above is as follows:

[0068] Scale matching and global alignment: using scale normalization transformation S(f) s ,f l To match the anemometer and lidar data on a spatiotemporal scale, a global alignment framework for multi-device measured data is constructed. The flow field distribution after scale matching and global alignment is as follows:

[0069] V′(x,y,z,t)=S(f s ,f l V(x,y,z,t)

[0070] Where V'(x,y,z,t) is the flow field coordinate system after scale matching and global alignment, V(x,y,z,t) is the Eulerian coordinate system of the wind field in the bridge area, x, y, and z represent the three-dimensional spatial position of wind speed measurement, and t represents the time dimension of wind speed.

[0071] The specific process of step 32 above is as follows:

[0072] To address the blind spots in lidar detection, flow field completion is achieved by combining radial basis functions and Kriging interpolation; the optimized flow field data after interpolation is as follows:

[0073]

[0074] Where, λ i φ(.) is the interpolation weight, φ(.) is the basis function, and h(x) is the polynomial trend term.

[0075] The specific process of step 33 above is as follows:

[0076] Based on flow field physical constraints and experimental calibration parameter C calib After dynamic correction and physical constraint correction, the high-precision flow field distribution is as follows:

[0077]

[0078] Using measured flow field data V exp (x,y,z,t) Evaluate reconstruction error:

[0079] ε=||V * (x,y,z,t)-V exp (x,y,z,t)||

[0080] This is to ensure the accuracy and physical consistency of the reconstructed flow field.

[0081] After adopting the above solution, the beneficial effects of the present invention are as follows:

[0082] (1) By networking the anemometers, the spatiotemporal consistency calibration of multi-sensor data is achieved, and compensation is performed by combining radar measurement data, which effectively reduces measurement errors and improves the accuracy and reliability of wind speed measurement.

[0083] (2) An adaptive anemometer deployment strategy based on flow field gradient characteristics and information entropy weights, combined with spatial interpolation methods to refine radar detection blind spots and improve data integrity in complex flow field areas.

[0084] (3) The scale matching technology is used to integrate the high-frequency local anemometer measurement data and the low-frequency large-area lidar data, and the physical constraint optimization interpolation algorithm is combined to achieve high-precision three-dimensional flow field reconstruction.

[0085] (3) By using the spatial matching search compensation mechanism of geometric array element coupling, the anemometer array is collaboratively networked, thereby improving the stability and adaptability of the measurement system under different environmental conditions. Attached Figure Description

[0086] Figure 1 This is a flowchart of the present invention;

[0087] Figure 2 This is a flowchart of the GAN learning normal wind speed distribution;

[0088] Figure 3 This is a flowchart of the automatic wind speed data repair process. Detailed Implementation

[0089] The technical solution and beneficial effects of the present invention will be described in detail below with reference to the accompanying drawings.

[0090] like Figure 1 As shown, this invention provides a method for fine reconstruction of wind field in radar detection blind zones using anemometer collaborative networking, comprising the following steps:

[0091] Step 1: Separate the measurement deviation caused by vibration displacement and perform automatic diagnosis and repair of anemometer data.

[0092] Step 2: Conduct optimization of the ultrasonic anemometer array layout for full-area wind field perception.

[0093] Step 3: Establish a high-resolution intelligent reconstruction model of the bridge flow field based on data-physics fusion.

[0094] In step 1, a wavelet packet-Kalman filter cascade algorithm is designed to separate the wind speed measurement deviation caused by vibration displacement in real time and eliminate vibration interference. The initial signal of the anemometer is converted into a visualized Gramian Angular Field (GAF) image. Generative Adversarial Networks (GANs) are used to learn the popular distribution of normal data, and an anomaly diagnosis method for anemometer test data is established by combining Autoencoder (AE) network and control chart theory. On this basis, a self-repair mechanism for anemometer abnormal data is established based on the improved Bayesian Dynamic Linear Model (BDLM), thereby optimizing data integrity and measurement continuity.

[0095] In step 2, the flow characteristics of the wind field on and around the bridge surface are analyzed based on the CFD (Computational Fluid Dynamics) simulation system to identify key areas and blind spots with significant wind speed changes. Taking into account various factors such as sensor performance parameters, installation location, signal transmission path, and environmental conditions in the bridge site area, anemometers are adaptively deployed based on flow field gradient characteristics and information entropy weights. Through automated time synchronization technology with simultaneous sampling of multiple anemometers, a spatial matching search compensation mechanism for geometric array coupling is established to calibrate the spatiotemporal consistency of flow field data in the anemometer array network.

[0096] In step 3, the high-frequency local data from the anemometer is scale-matched with the low-frequency large-area data from the lidar to establish a global alignment framework for multi-device measured data. To address the radar detection blind zone problem, spatial interpolation of the flow field is performed based on techniques such as radial basis function and Kriging interpolation to improve the data integrity of complex flow field regions. The spatial interpolation results are dynamically corrected in conjunction with experimental calibration parameters. By combining the physical characteristics of the flow field with the measured data, a high-precision three-dimensional flow field is reconstructed and verified using measured data.

[0097] Furthermore, the wavelet packet-Kalman filter cascade algorithm described in step 1 includes the following steps:

[0098] First, the wind speed signal S(t) is subjected to J-order wavelet packet decomposition to obtain the sub-band signal.

[0099]

[0100] in It is a wavelet packet basis function.

[0101] Translate wavelet packet signal Further decomposed into several modal components u i (t) represents the signal components in different frequency bands:

[0102]

[0103] Where H(.) denotes the Hilbert transform, ω i The center frequency for each mode.

[0104] Introducing non-Gaussian noise modeling and using Kalman filtering correction, a state-space model is constructed:

[0105] X t+1 =AX t +W t Z t =HX t +V t

[0106] Among them, X t+1 For actual wind speed conditions, Z t To observe wind speed data, W t V t It is Gaussian white noise.

[0107] The anemometer test data anomaly diagnosis method described in step 1 includes the following process:

[0108] like Figure 2 As shown, the wind speed signal S(t) is converted into a GAF image format, and a GAN is used to learn the normal wind speed data distribution, including:

[0109] Generator network G: Input noise vector z, output generated samples

[0110]

[0111] Discriminator Network D: Input is a GAF image transformed from the actual wind speed distribution or a generated sample. Output the probability of being true or false:

[0112] D(S)∈[0,1)

[0113] Through the optimal adversarial loss function:

[0114]

[0115] Train the GAN to generate samples that match the wind speed measurement distribution, thereby improving anomaly detection performance.

[0116] Construct an encoder-decoder AE network, input wind speed signal S(t), output reconstructed signal:

[0117]

[0118] Calculate reconstruction error If E(t) exceeds the control chart threshold, the wind speed is determined to be abnormal.

[0119] Furthermore, such as Figure 3 As shown, the wind speed anomaly detection and data self-repair method based on the improved BDLM described in step 1 includes four steps: model order selection, nonlinear state transition modeling, variational Bayesian inference optimization, and anomaly data self-repair. Specifically, adaptive model order selection optimizes parameter updates, nonlinear state transition modeling improves adaptability to non-Gaussian noise, and variational Bayesian inference accelerates convergence. Furthermore, through the anomaly data self-repair mechanism, the optimal autoregressive filtering strategy is used to correct anomaly wind speed data, improving the integrity and continuity of the measurement data. Detailed steps are as follows:

[0120] Model order selection: Construct a set of K candidate BDLM models with different orders {M1, M2, ..., M}. k The model score is calculated based on the Akaike information criterion (AIC) and root mean squared error (RMSE), and then fused into a normalized score index S. k S k A larger value indicates a better model order.

[0121]

[0122]

[0123]

[0124] In the formula: p k The number of parameters in the model; y is the maximum likelihood estimate of the model on the data; i These are actual observations; The model's predicted output is given by N; N is the total number of samples; exp(·) is the exponential function; min j AIC j ε represents the minimum AIC value among all candidate models; ε is a very small positive number to prevent the denominator from being zero.

[0125] Based on the selection of the optimal model order, a Bayesian Model Averaging (BMA) strategy is introduced to fuse the prediction results of multiple superior candidate models, avoiding model selection bias and improving the system's robustness and generalization ability. Multiple candidate models are weighted using BMA, and the weighted prediction value is calculated.

[0126]

[0127] In the formula, This is the model's predicted output; This is the final predicted value after weighted fusion.

[0128] Nonlinear state transition modeling: To address the nonlinear characteristics of wind speed measurement errors, a nonlinear term f(X) is introduced into the state equation. t Construct a generalized state transition model:

[0129] X t =F t X t-1 +G t +f(X t-1 )+W t

[0130] in

[0131] f(X t )=αtanh(βX t )

[0132] In the formula, α and β are parameters to be optimized to enhance the model's adaptability to nonlinear signals;

[0133] Variational Bayesian Inference Optimization: To address the lag in parameter updates in traditional Bayesian methods, variational inference is used to optimize the objective function.

[0134]

[0135] In the formula, D KL (.) represents the Kullback-Leibler (KL) divergence, which improves the model convergence speed and enhances outlier identification and data repair capabilities by optimizing the variational distribution q(X).

[0136] Anomaly data self-repair mechanism: An improved BDLM is used to model and predict the normal temporal distribution of wind speed. After detecting anomalies, an optimal autoregressive filter is used to repair the abnormal data.

[0137]

[0138] In the formula, λ∈(0,1) is the dynamic adjustment factor.

[0139] If outliers occur consecutively, the model prediction portion with higher weight is used to ensure that the repaired wind speed data achieves a balance between continuity and physical consistency.

[0140] Furthermore, the anemometer adaptive deployment method described in step 2 is characterized by the following optimization process:

[0141] Let the Eulerian coordinate system of the wind field in the bridge area be represented as:

[0142] V(x,y,z,t)=(u(x,y,z,t),v(x,y,z,t),w(x,y,z,t))

[0143] Where x, y, and z represent the three-dimensional spatial location of the wind speed measurement; t represents the time dimension of the wind speed, reflecting the change of wind speed over time; and (u, v, w) are the three components of the wind speed. Let the set of anemometer locations be P = {p1, p2, ..., p...} n}, where each sensor p i It has a set of performance parameters:

[0144] Θ i ={R i ,S i ,T i E i}

[0145] Where: Ri is the sensor measurement accuracy; S i T represents the spatial resolution of the sensor. i E represents the sensor signal transmission delay. i This refers to the sensor's power consumption.

[0146] Define the optimization objective function:

[0147]

[0148]

[0149] in: This represents the wind speed field gradient intensity, used to determine key areas; To maximize the information gain of the measurement point distribution, the information entropy is used. τ1 represents signal transmission delay loss, ensuring network synchronization after sensor deployment; w1, w2, and w3 are weighting coefficients used to balance different optimization objectives; τ1 is the wind speed gradient threshold, ensuring the anemometer deployment covers areas with significant wind field changes; τ2 is the lower limit of information entropy to avoid sensor redundancy; τ3 is the maximum permissible signal transmission delay; E max This represents the upper limit of total energy consumption.

[0150] To address the aforementioned multi-objective optimization problem, a genetic algorithm is employed to determine the optimal ultrasonic anemometer array scheme. The solution process is as follows: Anemometer deployment schemes are randomly generated within the bridge area; the merits of the current scheme are evaluated based on the objective function J(P); a new sensor deployment scheme is generated using roulette wheel selection and two-point crossover; a small-probability perturbation is introduced to adjust the positions of individual anemometers to avoid local optima; if the objective function converges, the optimal deployment scheme is output.

[0151] Furthermore, the high-precision three-dimensional flow field reconstruction achieved through multi-device data fusion and flow field interpolation optimization described in step 3 includes the following steps:

[0152] Scale matching and global alignment: using scale normalization transformation S(f) s ,f l This involves matching anemometer (high-frequency local) and lidar (low-frequency large-area) data on a spatiotemporal scale, constructing a multi-device global data alignment framework. The flow field distribution after scale matching and global alignment is as follows:

[0153] V′(x,y,z,t)=S(f s ,f l V(x,y,z,t)

[0154] Flow field interpolation optimization: For the blind zone of lidar detection, flow field completion is performed by combining radial basis functions and Kriging interpolation. The flow field data after interpolation optimization (filling in the lidar blind zone) is as follows:

[0155]

[0156] Where, λ i φ(.) is the interpolation weight, φ(.) is the basis function, and h(x) is the polynomial trend term.

[0157] High-precision flow field reconstruction: based on flow field physical constraints (mass conservation) ) and experimental calibration parameter C calib Dynamic correction is performed. The high-precision flow field distribution after physical constraint correction is as follows:

[0158]

[0159] Using measured flow field data V exp (x,y,z,t) Evaluate reconstruction error:

[0160] ε=||V * (x,y,z,t)-V exp (x,y,z,t)||

[0161] This is to ensure the accuracy and physical consistency of the reconstructed flow field.

[0162] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0163] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0164] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0165] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0166] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0167] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for fine reconstruction of wind field in radar detection blind zone using anemometer collaborative networking, characterized in that... Includes the following steps: Step 1: Preprocess the anemometer data and repair any abnormal data. Step 2: Based on the flow characteristics of the bridge surface and the surrounding wind field, design the optimal ultrasonic anemometer array scheme. Step 3: Establish a high-resolution intelligent reconstruction model of the bridge flow field based on data-physics fusion; In step 1, abnormal data is repaired based on the improved Bayesian dynamic linear model BDLM, which includes four steps: model order selection, nonlinear state transition modeling, variational Bayesian inference optimization, and abnormal data self-repair. Model order selection: Construct a set of K candidate BDLM models with different orders {M1, M2, ..., M}. k The model score is calculated based on the Akaike Information Content Criterion (AIC) and the Root Mean Square Error (RMSE), and then fused into a normalized score index S. k S k A larger value indicates a better model order. ， ， ， In the formula: p k The number of parameters in the model; y is the maximum likelihood estimate of the model on the data; i These are actual observations; is the predicted output of the model; N is the total number of samples; exp(⋅) is the exponential function; This represents the minimum AIC value among all candidate models; ε is a very small positive number to prevent the denominator from being zero. Multiple candidate models are weighted using Bayesian model average (BMA) to calculate the weighted prediction value: ， In the formula, This is the model's predicted output; This is the final predicted value after weighted fusion; Nonlinear state transition modeling: To address the nonlinear characteristics of wind speed measurement errors, a nonlinear term f(X) is introduced into the state equation. t Construct a generalized state transition model: ， in, ， In the formula, α and β are the parameters to be optimized; Variational Bayesian inference optimization: Optimizing the objective function using variational inference. ， In the formula, D KL (.) represents the KL divergence, which can be optimized by adjusting the variational distribution q(X) to improve the model convergence speed; Anomaly self-repair: An improved BDLM is used to model and predict the normal temporal distribution of wind speed. After detecting outliers, an optimal autoregressive filter is employed to repair the anomaly data. ， In the formula, λ∈(0,1) is the dynamic adjustment factor; The specific process of step 2 is as follows: Let the Eulerian coordinate system of the wind field in the bridge area be represented as: ， Where x, y, and z represent the three-dimensional spatial location of the wind speed measurement; t represents the time dimension of the wind speed; (u, v, w) are the three components of the wind speed; let the set of anemometer locations be P = {p1, p2, ..., p...} n }, where each sensor p i It has a set of performance parameters: ， Where: Ri is the sensor measurement accuracy; S i T represents the spatial resolution of the sensor. i E represents the sensor signal transmission delay. i For sensor power consumption; Define the optimization objective function: ， ， in: Indicates the gradient intensity of the wind speed field; Information entropy; Represents signal transmission delay loss; w1, w2, and w3 are weighting coefficients; τ1 is the wind speed gradient threshold; τ2 is the lower limit of information entropy; τ3 is the maximum permissible signal transmission delay; E max This represents the upper limit of total energy consumption. For the multi-objective optimization problem, a genetic algorithm is used to determine the optimal ultrasonic anemometer array scheme. The specific process of step 3 is as follows: Step 31: Perform scale matching between the high-frequency local data of the anemometer and the low-frequency large-area data of the lidar to establish a global alignment framework for the measured data of multiple devices. Step 32: To address the radar detection blind zone problem, perform spatial interpolation of the flow field and dynamically correct the spatial interpolation results based on experimental calibration parameters. Step 33: Combine the physical characteristics of the flow field with the measured data to reconstruct a high-precision three-dimensional flow field, and verify it with the measured data; The specific process of step 31 is as follows: Scale matching and global alignment: using scale normalization transformation S(f) s , f l To match the anemometer and lidar data on a spatiotemporal scale, a global alignment framework for multi-device measured data is constructed. The flow field distribution after scale matching and global alignment is as follows: ， Where V'(x,y,z,t) is the flow field coordinate system after scale matching and global alignment, V(x,y,z,t) is the Eulerian coordinate system of the wind field in the bridge area, x, y, and z represent the three-dimensional spatial position of wind speed measurement, and t represents the time dimension of wind speed. The specific process of step 32 is as follows: To address the blind spots in lidar detection, flow field completion is achieved by combining radial basis functions and Kriging interpolation; the optimized flow field data after interpolation is as follows: ， in, For interpolation weights, As basis functions, For polynomial trend terms; The specific process of step 33 is as follows: Based on flow field physical constraints and experimental calibration parameter C calib After dynamic correction and physical constraint correction, the high-precision flow field distribution is as follows: ， Using measured flow field data Assess reconstruction error: ， This is to ensure the accuracy and physical consistency of the reconstructed flow field.

2. The method as described in claim 1, characterized in that: In step 1, a wavelet packet-Kalman filter cascade algorithm is used to separate the wind speed measurement deviation caused by vibration displacement. First, the wind speed signal S(t) is subjected to J-level wavelet packet decomposition to obtain the sub-band signal. : ， in These are wavelet packet basis functions; Translate wavelet packet signal Further decomposed into several modal components These represent signal components at different frequency bands: ， Where H(.) denotes the Hilbert transform, The center frequency of each mode; Introducing non-Gaussian noise modeling and using Kalman filtering correction, a state-space model is constructed: ， in, This represents the actual wind speed. To observe wind speed data, , It is Gaussian white noise.

3. The method as described in claim 1, characterized in that: In step 1, abnormal data is identified according to the following method: The wind speed signal S(t) is converted into a Grammy angle field image, and a generative adversarial network is used to learn the normal wind speed data distribution, including: Generator network G: Input noise vector z, output generated samples , ， Discriminator Network D: Input is a Grammy angle field image transformed from the real wind speed distribution or a generated sample. Output the probability of being true or false: ， Through the optimal adversarial loss function: ， Train a generative adversarial network to generate samples that conform to the wind speed measurement distribution; Construct an encoder-decoder network, input wind speed signal S(t), output reconstructed signal: ， Calculate reconstruction error If E(t) exceeds the control chart threshold, the wind speed is determined to be abnormal.

Images