Abnormal wind data self-adaptive completion system for long-span bridge based on pigan
By constructing an adaptive completion system for abnormal wind data of long-span bridges based on PIGAN, and using predefined prior criteria and generative adversarial networks to perform adaptive and accurate completion of abnormal wind data, the shortcomings of traditional methods in wind field characteristic characterization are solved, and the reliability and timeliness of bridge wind vibration analysis are improved.
Patent Information
- Application Number
- CN202610680267.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-18
- Publication Date
- 2026-08-25
AI Technical Summary
Traditional data restoration methods are difficult to accurately characterize the strong nonlinear and non-stationary characteristics of wind fields around long-span bridges. Deep learning-based completion models face the problem of physical law distortion under small sample conditions, resulting in insufficient timeliness and reliability of bridge wind-induced vibration early warning.
An adaptive completion system for abnormal wind data of long-span bridges based on PIGAN is constructed, including an abnormal wind data discrimination module, an interpolation module and an evaluation module. The system uses a set prior criteria and a generative adversarial network to perform adaptive and accurate completion of abnormal wind data, combining the physical constraints of wind data and a dynamic adversarial generation mechanism.
It achieves adaptive and accurate completion of abnormal wind data for long-span bridges, providing reliable wind field input for wind-induced vibration analysis of bridge structures and improving the timeliness and reliability of wind-induced vibration early warning.
Smart Images
Figure CN122634700A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an adaptive data completion system for abnormal wind data of long-span bridges based on PIGAN, belonging to the field of fault diagnosis technology for structural health monitoring systems of long-span bridges. Background Technology
[0002] With the rapid development of long-span bridge construction, structural health monitoring systems have become a core technological means to ensure the safe operation of bridges. Wind load, as one of the main environmental loads on long-span bridges, directly affects the bridge's flutter stability, vortex-induced vibration response, and fatigue life assessment. However, in actual monitoring, sensor failures, extreme weather interference, or abnormal data transmission often lead to localized missing or abrupt changes in wind data, severely restricting the timeliness and reliability of wind-induced vibration early warning for bridges. Traditional data restoration methods struggle to accurately characterize the strong nonlinear and non-stationary characteristics of the wind field surrounding long-span bridges, while deep learning-based completion models face the problem of distorted physical laws under small sample conditions. Therefore, there is an urgent need to explore new data restoration technologies that integrate domain knowledge.
[0003] In recent years, Physical Information Neural Networks (PINs) have demonstrated unique advantages in fluid dynamics modeling by embedding constraints from governing equations. However, standard PINs suffer from drawbacks such as slow convergence and insufficient multi-physics coupling modeling capabilities when processing high-dimensional spatiotemporal data. Meanwhile, the combination of Generative Adversarial Networks' strong feature extraction capabilities with physical constraints offers a new research path for anomalous data repair.
[0004] Against this backdrop, there is an urgent need to construct an intelligent data completion system for the wind field characteristics of long-span bridges. By deeply integrating prior knowledge of boundary layer wind field physics with a dynamic adversarial generation mechanism, the system can achieve adaptive and accurate completion of abnormal wind data for long-span bridges, providing reliable wind field input for wind vibration analysis of bridge structures. Summary of the Invention
[0005] Technical Problem: The purpose of this invention is to provide an adaptive wind data completion system for long-span bridges based on PIGAN, including a long-span bridge anomalous wind data discrimination module, a long-span bridge anomalous wind data interpolation module, and a long-span bridge interpolated wind data evaluation module. It can achieve adaptive and accurate completion of anomalous wind data for long-span bridges, providing reliable wind field input for wind vibration analysis of bridge structures.
[0006] Technical solution: To solve the above technical problems, the present invention provides an adaptive completion system for abnormal wind data of long-span bridges based on PIGAN, including a long-span bridge abnormal wind data discrimination module, a long-span bridge abnormal wind data interpolation module, and a long-span bridge interpolated wind data evaluation module connected in sequence. The abnormal wind data discrimination module for long-span bridges uses a set prior criterion to make real-time judgments on the abnormal state of measured wind data, and completes the marking and classification of abnormal data. The abnormal wind data interpolation module for long-span bridges interpolates measured wind data of different abnormal types based on the abnormal data classification results of the abnormal wind data discrimination module for long-span bridges. If the abnormal wind data belongs to the accidental abnormal category, the weighted K-nearest neighbor (WKNN) algorithm is used; if the abnormal wind data belongs to the continuous abnormal category, the physical-data collaboratively driven generative adversarial network (PIGAN) is used. The large-span bridge interpolation wind data evaluation module is used to evaluate whether the interpolated wind data meets the preset quality requirements. If the preset quality requirements are not met, the system returns to the large-span bridge abnormal wind data interpolation module; if the preset quality requirements are met, the large-span bridge abnormal wind data adaptive interpolation is completed.
[0007] The abnormal wind data discrimination module for long-span bridges includes a wind data recording submodule, a data anomaly discrimination submodule, and a data anomaly classification submodule. The wind data recording submodule records the measured wind data X in real time. i Corresponding time t i and the data storage location is Z i ; The data anomaly detection submodule is based on the Rheinland criterion to detect data anomalies, as shown in equations (1a) and (1b): , In the formula: and Let represent the mean and standard deviation of the previous i measured wind data, respectively.
[0008] The data anomaly detection submodule defines a time window T. If the data anomaly detection submodule runs for t=T, and anomaly data location marking is performed within time period T, then the measured wind data X, the corresponding time t, and the data storage location are set to Z. i Transmit to the data anomaly classification submodule: If t=T, and no anomaly data location marking operation is performed within the time period T, then return to the wind data recording submodule; If t<T, then return to the wind data recording submodule. The data anomaly classification submodule divides the measured wind data sequence with a time length of T into a normal segment T1 and an abnormal segment T2, and calculates the autocorrelation coefficient ACF(T1) of the normal segment T1 and the autocorrelation coefficient ACF(T2) of the abnormal segment T2.
[0009] The data anomaly classification submodule uses the confidence interval method to compare ACF(T2) with ACF(T1). If ACF(T2) exceeds the confidence interval of ACF(T1), then the anomaly segment T2 belongs to continuous anomaly; if ACF(T2) is within the confidence interval of ACF(T1), then the anomaly segment T2 belongs to random anomaly.
[0010] The aforementioned abnormal wind data interpolation module for long-span bridges includes an occasional abnormal data interpolation submodule and a continuous abnormal data interpolation submodule. The random anomaly data interpolation submodule is based on the WKNN algorithm. It selects the K nearest neighboring points around the anomaly point, calculates the interpolation value of the random anomaly point by defining distance metric and weight allocation coefficient, and transmits it to the long-span bridge interpolation wind data evaluation module. The continuous anomalous data interpolation submodule generates the objective function of the adversarial network based on the following formula (2), integrates multi-source wind effect data and structural dynamic physical constraints, and realizes the interpolation of continuous anomalous data by constructing PIGAN, and transmits it to the long-span bridge interpolation wind data evaluation module. , In the formula: G and D represent the generator model and discriminator model of GAN, respectively; Let it be the expected function; Distribution of noise data; Let x be the target data distribution; let z be the target random variable; and z be the noise random variable.
[0011] The aforementioned long-span bridge interpolation wind data evaluation module first performs interpolation values based on the Rheinda criterion in equations (1a, 1b). The distribution rationality assessment is based on the turbulence intensity I in equation (3). u I v and I w The power spectral density S in equation (4) u S v and S w As the interpolation value Physical rationality assessment; , , , In the formula, I u I v and I w These represent the forward, lateral, and vertical turbulence intensities, respectively. , and These represent the longitudinal, lateral, and vertical pulsating wind speeds, respectively; σ uσ v and σ w These represent the standard deviations of the longitudinal, lateral, and vertical pulsating wind speeds, respectively. Here are the Monin coordinates; n is the natural frequency; h is the altitude; and U is the average wind speed. For shear wind speed, ε = u, v, w; , , and All are dimensionless coefficients to be determined.
[0012] If interpolation value The following three conditions must be met simultaneously: First... Second, I u : I v : I w =1 :0.88 : 0.5; Third, If the preset quality requirements are not met, the continuous abnormal data interpolation is completed. If the interpolated samples cannot meet the requirements simultaneously, the process returns to the continuous abnormal interpolation submodule.
[0013] Beneficial effects: The adaptive wind data completion system for long-span bridges of the present invention consists of a wind data discrimination module, a wind data interpolation module, and an interpolated wind data evaluation module. The wind data discrimination module can use predefined prior criteria to identify abnormal states in measured wind data in real time, and complete the marking and classification of abnormal data. The wind data interpolation module can adaptively interpolate measured wind data according to the type of abnormal data. The interpolated wind data evaluation module can evaluate whether the interpolated wind data achieves adaptive and accurate completion of wind data for long-span bridges, thereby providing reliable input for wind vibration analysis of bridge structures. Attached Figure Description
[0014] Figure 1 This is a schematic diagram of an adaptive data completion system for abnormal wind data of a long-span bridge based on PIGAN, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the abnormal wind data discrimination module for long-span bridges according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the abnormal wind data interpolation module for long-span bridges according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the long-span bridge interpolation wind data evaluation module according to an embodiment of the present invention.
[0015] The diagram includes: a large-span bridge abnormal wind data discrimination module 101, a large-span bridge abnormal wind data interpolation module 102, and a large-span bridge interpolated wind data evaluation module 103. Detailed Implementation
[0016] Figure 1 This is a schematic diagram of an adaptive data completion system for abnormal winds in long-span bridges based on PIGAN, according to an embodiment of the present invention. Figure 1 The adaptive completion system for abnormal wind data of long-span bridges shown in this embodiment may include: a long-span bridge abnormal wind data discrimination module 101, a long-span bridge abnormal wind data interpolation module 102, and a long-span bridge interpolated wind data evaluation module 103. The abnormal wind data discrimination module 101 for long-span bridges uses a set prior criterion to make real-time judgments on the abnormal state of measured wind data, and completes the marking and classification of abnormal data. The long-span bridge abnormal wind data interpolation module 102 adaptively interpolates the measured wind data according to the type of abnormal data. If the abnormal data belongs to the accidental abnormal category, the weighted K-nearest neighbor (WKNN) algorithm is used for adaptive interpolation; if the abnormal data belongs to the continuous abnormal category, the physical-data collaboratively driven generative adversarial network (PIGAN) is used for adaptive interpolation. The large-span bridge interpolation wind data evaluation module 103 is used to evaluate whether the interpolated wind data meets the preset quality requirements. If the preset quality requirements are not met, the system returns to the large-span bridge abnormal wind data interpolation module. If the preset quality requirements are met, the system completes the large-span bridge abnormal wind data adaptive interpolation.
[0017] Figure 2 This is a schematic diagram of the abnormal wind data discrimination module for long-span bridges according to an embodiment of the present invention, as shown below. Figure 2 The abnormal wind data discrimination module for long-span bridges shown can include: a wind data recording submodule, a data anomaly discrimination submodule, and a data anomaly classification submodule. Through the coordinated operation of the three submodules, the abnormal state can be judged in real time, and the abnormal data can be marked and classified.
[0018] Specifically, the wind data recording submodule records the measured data X in real time. i Corresponding time t i and the data storage location is Z i The data anomaly detection submodule determines data anomaly states based on the Rheinda criterion in equations (1a, 1b). ,
[0019] In the formula: and Let represent the mean and standard deviation of the previous i measured wind data, respectively.
[0020] The data anomaly classification submodule sets a time window T, divides the time series into a normal segment T1 and an anomaly segment T2, and calculates the autocorrelation coefficient ACF(T1) for the normal segment T1 and the autocorrelation coefficient ACF(T2) for the anomaly segment T2: , In the formula, k represents the lag order, and in this example of the invention, k takes the values 1, 2, or 3; express and Covariance between them; express The variance; The data anomaly classification submodule defines the confidence interval of ACF(T1). Where: the critical value z represents the quantile corresponding to α / 2 in the standard normal distribution; α represents the significance level.
[0021] Table 1 shows the significance level α and its corresponding z-value. In this invention example, α = 0.05, and its corresponding z-value is... =1.96.
[0022] Table 1. Significance level α and its corresponding z-value , The standard error of the normal segment T1 is calculated according to equation (3): , In the formula, t s This indicates the wind data sampling frequency of the structural health monitoring system.
[0023] If ACF(T2) If the anomaly segment T2 is a continuous anomaly, then ACF(T2) is a continuous anomaly. If so, then the abnormal segment T2 is an occasional abnormality.
[0024] If the data anomaly detection submodule runs for t=T and performs an anomaly data location marking operation within the time period T, then the measured wind data X, the corresponding time t, and the data storage location Z are transmitted to the data anomaly classification submodule; if t=T and no anomaly data location marking operation is performed within the time period T, then the data is returned to the wind data recording submodule; if t<T, then the data is returned to the wind data recording submodule. Figure 3 This is a schematic diagram of the abnormal wind data interpolation module for long-span bridges according to an embodiment of the present invention, as shown below. Figure 3 The abnormal wind data interpolation module for long-span bridges shown may include an accidental abnormal data interpolation submodule and a continuous abnormal data interpolation submodule. Specifically, when the abnormal wind data discrimination module for long-span bridges identifies abnormal data as random outliers, the random outlier data interpolation submodule is run, and its operation steps are as follows: (1) Obtain measured data X and abnormal data location Z from the abnormal wind data discrimination module of long-span bridge, locate the abnormal data point, and determine the selection of K value. In this example, K=3 is determined; (2) Based on the three types of distance metrics in equation (3), calculate the distance d between the outlier and other points, and determine the nearest K data points under the three types of metrics in ascending order of distance; European distance: (4a), In the formula, x i and y i Let x1, x2, and y2 represent the coordinates of the anomaly point and other points in n-dimensional space, respectively. In this invention, n=2, (x1, x2) represent the wind speed and wind direction values of the anomaly point, and (y1, y2) represent the wind speed and wind direction values of other points.
[0025] Manhattan distance: (4b), Chebyshev distance: (4c), (3) Based on the three weighting algorithms in equation (5), assign corresponding weights w to all selected neighboring points under the three types of index conditions: Inverse distance weighting: (5a), In the formula, This represents an outlier point X and its neighboring points X. i The distance is calculated according to formula (3); p represents the power parameter, which is usually 1 or 2.
[0026] Gaussian kernel weights: (5b),
[0027] In the formula, σ represents the standard deviation of the Gaussian kernel.
[0028] Locally weighted linear regression: (5c), In the formula, Represents X and X in vector space i The square of the distance between them; This is the bandwidth parameter.
[0029] (4) Calculate the interpolation value of the outlier point X based on equation (6). ; , Specifically, when the abnormal wind data discrimination module for long-span bridges identifies abnormal data as continuous outliers, the continuous outlier data interpolation submodule is run, and its operation steps are as follows: (1) Obtain measured data X and abnormal data location Z from the abnormal wind data discrimination module of long-span bridge to form the initial wind speed dataset X; (2) Index the abnormal data position Z, delete the abnormal data at the corresponding position in X, and construct the target sample required for training the adversarial network. ; (3) Construct the generator model G and discriminator model D of the generative adversarial network; Table 2 shows an example of the network structure of the generator model G. The network structure of the generator model G contains 5 processing layers. The first 4 layers include: a fully connected Dense layer, a normalized Batchnorm layer, and a nonlinear LeakyReLU layer. The 5th layer contains a fully connected Dense layer.
[0030] The generator G takes random noise Z as input and outputs interpolated wind data samples. .
[0031] Table 2. Examples of network structures for generator model G ,
[0032] Table 3 shows an example of the network structure of the discriminator model D. The network structure of the generator model D contains 5 processing layers. The first 4 layers include a fully connected Dense layer, a normalized Batchnorm layer, and a nonlinear LeakyReLU layer. The 5th layer includes a fully connected Dense layer and a Sigmoid layer.
[0033] The input to the discriminator is the target sample. and interpolated samples The output is a mapping value of the current wind data information.
[0034] Table 3. Examples of network structures for generator model D , (4) Integrating multi-source wind effect data and structural dynamic physical constraints, the objective function L of G is set. G As shown in equation (7): (7a), (7b), (7c), In the formula, L a and Lp λ represents adversarial loss and physical loss, respectively; a and λ p These represent the correction coefficients for the corresponding loss terms; Let it be the expected function; Distribution of noise data; and These represent the predicted and measured accelerations of long-span bridges, respectively. N is the number of time discrete terms, where The abnormal wind data interpolation module for long-span bridges is input from the bridge structural health monitoring system. Calculate according to equation (7): (8a), (8b), In the formula: M, C and K are the mass matrix, damping matrix and stiffness matrix of the bridge, respectively; , and These represent the bridge under wind loads. The acceleration, velocity, and displacement of the following values, among which That is A represents the windward area of the bridge. U is the air density; C is the aerodynamic coefficient of the bridge; U and These represent the average wind speed term and the fluctuating wind speed term, respectively.
[0035] (5) Set the objective function of D as shown in equation (9): , In the formula: For the target data distribution, (6) The objective functions of both the generator model G and the adversary model D are the Adam optimizer function. The model parameters of G and D are iteratively optimized based on the gradient descent algorithm to obtain the optimal value of the objective function. The iteration stops when the objective function reaches the optimal value. (7) After the iterative training is completed, the interpolated wind data is generated and transmitted to the long-span bridge interpolated wind data evaluation module 103.
[0036] Figure 4 This is a schematic diagram of the long-span bridge interpolation wind data evaluation module according to an embodiment of the present invention, as shown below. Figure 4 The large-span bridge interpolation wind data evaluation module shown is used to evaluate whether the interpolated wind data meets the preset quality requirements. Specifically, in the large-span bridge abnormal wind data interpolation module 102, the interpolated wind speed sample is input into 103. 103 first evaluates the reasonableness of the distribution of interpolated values based on the Rhine criterion in equation (1), and then evaluates the turbulence intensity I based on equation (10). u I v and I w and the power spectral density S in equation (11) u S v and S w As an assessment of the physical rationality of performing interpolation values.
[0037] , , In the formula, I u I v and I w These represent the forward, lateral, and vertical turbulence intensities, respectively. , and These represent the longitudinal, lateral, and vertical pulsating wind speeds, respectively; σ u σ u and σ u S represents the standard deviation of the longitudinal, lateral, and vertical fluctuating wind speeds, respectively; u S v and S w These represent the power spectral densities in the forward, transverse, and vertical directions, respectively. Here are the Monin coordinates; n is the natural frequency; h is the altitude; and U is the average wind speed. Shear wind speed; , , and is a dimensionless coefficient to be determined.
[0038] Power spectral density , and The Welch method can be used for calculation. The specific steps are as follows: Divide the longitudinal, transverse, and vertical fluctuating wind speeds at T=10 min into 8 segments with an overlap factor of 0.5. Calculate the measured wind spectrum using the Welch method. Based on the measured spectrum results, fit the dimensionless parameters using the least squares method. , , and And set the objective function value change to be less than 10. -7 Terminate the iteration.
[0039] If interpolation air data The following three conditions must be met simultaneously: 1) ;2)I u : Iv : I w =1 : 0.88 : 0.5; 3) If the preset quality requirements are not met, the continuous abnormal data interpolation is completed. If the interpolated samples cannot meet the requirements simultaneously, the process returns to the continuous abnormal interpolation evaluation submodule.
Claims
1. An adaptive data completion system for anomalous wind data of long-span bridges based on PIGAN, characterized in that, It includes a long-span bridge abnormal wind data discrimination module (101), a long-span bridge abnormal wind data interpolation module (102), and a long-span bridge interpolated wind data evaluation module (103) connected in sequence. The abnormal wind data discrimination module (101) for long-span bridges uses a set prior criterion to make real-time judgments on the abnormal state of measured wind data, and completes the marking and classification of abnormal data. The long-span bridge abnormal wind data interpolation module (102) interpolates measured wind data of different abnormal types based on the abnormal data classification results of the long-span bridge abnormal wind data discrimination module (101). If the abnormal wind data belongs to the accidental abnormal category, the weighted K-nearest neighbor WKNN algorithm is used; if the abnormal wind data belongs to the continuous abnormal category, the physical-data collaboratively driven generative adversarial network PIGAN is used. The large-span bridge interpolation wind data evaluation module (103) is used to evaluate whether the interpolated wind data meets the preset quality requirements. If the preset quality requirements are not met, the system returns to the large-span bridge abnormal wind data interpolation module. If the preset quality requirements are met, the system completes the large-span bridge abnormal wind data adaptive interpolation.
2. The adaptive data completion system for abnormal winds in long-span bridges based on PIGAN as described in claim 1, characterized in that: The abnormal wind data discrimination module (101) for long-span bridges includes a wind data recording submodule, a data anomaly discrimination submodule, and a data anomaly classification submodule; The wind data recording submodule records the measured wind data X in real time. i Corresponding time t i and the data storage location is Z i ; The data anomaly detection submodule is based on the Rheinland criterion to detect data anomalies, as shown in equations (1a) and (1b): ， In the formula: and Let represent the mean and standard deviation of the first i measured wind data, respectively. The data anomaly detection submodule defines a time window T. If the data anomaly detection submodule runs for t=T, and anomaly data location marking is performed within time period T, then the measured wind data X, the corresponding time t, and the data storage location are set to Z. i Transmit to the data anomaly classification submodule: If t=T, and no anomaly data location marking operation is performed within the time period T, then return to the wind data recording submodule; If t<T, then return to the wind data recording submodule. The data anomaly classification submodule divides the measured wind data sequence with a time length of T into a normal segment T1 and an abnormal segment T2, and calculates the autocorrelation coefficient ACF(T1) of the normal segment T1 and the autocorrelation coefficient ACF(T2) of the abnormal segment T2.
3. The adaptive data completion system for abnormal winds in long-span bridges based on PIGAN as described in claim 2, characterized in that: The data anomaly classification submodule uses the confidence interval method to compare ACF(T2) with ACF(T1). If ACF(T2) exceeds the confidence interval of ACF(T1), then the anomaly segment T2 belongs to continuous anomaly; if ACF(T2) is within the confidence interval of ACF(T1), then the anomaly segment T2 belongs to random anomaly.
4. The adaptive data completion system for abnormal winds in long-span bridges based on PIGAN as described in claim 3, characterized in that: The long-span bridge abnormal wind data interpolation module (102) includes an occasional abnormal data interpolation submodule and a continuous abnormal data interpolation submodule; The random anomaly data interpolation submodule is based on the WKNN algorithm. It selects the K nearest neighboring points around the anomaly point, calculates the interpolation value of the random anomaly point by defining distance metric and weight allocation coefficient, and transmits it to the long-span bridge interpolation wind data evaluation module (103). The continuous anomalous data interpolation submodule generates the objective function of the adversarial network based on the following formula (2), integrates multi-source wind effect data and structural dynamic physical constraints, and realizes the interpolation of continuous anomalous data by constructing PIGAN, and transmits it to the long-span bridge interpolation wind data evaluation module (103). ， In the formula: G and D represent the generator model and discriminator model of GAN, respectively; Let it be the expected function; Distribution of noise data; Let x be the target data distribution; let z be the target random variable; and z be the noise random variable.
5. The adaptive data completion system for abnormal winds in long-span bridges based on PIGAN as described in claim 4, characterized in that: The long-span bridge interpolation wind data evaluation module (103) first performs interpolation values based on the Rheinda criterion in equations (1a) and (1b). The distribution rationality assessment is based on the turbulence intensity I in equation (3). u I v and I w The power spectral density S in equation (4) u S v and S w As the interpolation value Physical rationality assessment; ， ， ， In the formula, I u I v and I w These represent the forward, lateral, and vertical turbulence intensities, respectively. , and These represent the longitudinal, lateral, and vertical pulsating wind speeds, respectively; σ u σ v and σ w These represent the standard deviations of the longitudinal, lateral, and vertical pulsating wind speeds, respectively. Here are the Monin coordinates; n is the natural frequency; h is the altitude; and U is the average wind speed. For shear wind speed, ε = u, v, w; , , and All are dimensionless coefficients to be determined.
6. The adaptive data completion system for abnormal winds in long-span bridges based on PIGAN as described in claim 5, characterized in that: If interpolation value The following three conditions must be met simultaneously: First... Second, I u : I v : I w =1 :0.88 : 0.5; Third, If the preset quality requirements are not met, the continuous abnormal data interpolation is completed. If the interpolated samples cannot meet the requirements simultaneously, the process returns to the continuous abnormal interpolation submodule.