High-rise structure wind load spectrum identification method and device based on migration deep learning

Abstract

The invention provides a high-rise structure wind load spectrum identification method and device based on migration deep learning. The method comprises the following steps: calculating according to dynamic response time history data of a target structure under the action of a target wind load to obtain a corresponding observation response power spectrum matrix; performing discrete cosine transform decomposition on the wind speed spectrum of the target wind load to obtain a dimension-reduced DCT coefficient; constructing a deep learning model based on a physical guidance mechanism, taking the observation response power spectrum matrix as an input of the deep learning model, taking the DCT coefficient as an output, and enabling the deep learning model to learn a mapping relation between the observation response power spectrum matrix and the DCT coefficient through training; and outputting a corresponding DCT (Discrete Cosine Transform) coefficient by adopting the trained deep learning model based on the input observation response power spectrum matrix, and reconstructing a power spectrum matrix of the wind load according to the DCT coefficient. According to the technical scheme provided by the embodiment of the invention, the incidence relation between the response power spectrum and the load power spectrum can be effectively established, and the accuracy of wind load identification is ensured.

Description

Technical Field

[0001] The present disclosure relates to the technical field of data processing, and particularly to a method and device for identifying high-rise structural wind load spectra based on transfer deep learning. Background Art

[0002] In the field of civil engineering, the dynamic response analysis of large-scale structures such as high-rise buildings and bridges under natural wind loads is of great significance for structural health monitoring, safety assessment, and optimal design. Due to the significant randomness and spatial distribution characteristics of wind loads, how to accurately identify their power spectral matrix has become a key challenge.

[0003] In current technical solutions, most of the solutions to the problem of wind load identification rely on inverse problem solving, that is, inferring the load from the structural response. However, this process is easily interfered by noise, resulting in unstable solutions, and it is necessary to rely on regularization techniques to increase the computational complexity. Moreover, the identification of the full-degree-of-freedom load spectrum requires high-dimensional observation data, which is difficult to achieve in actual engineering and has insufficient noise resistance. Therefore, how to effectively establish the correlation between the response power spectrum and the load power spectrum to ensure the accuracy of wind load identification has become an urgent technical problem to be solved. Summary of the Invention

[0004] The present disclosure provides a method and device for identifying high-rise structural wind load spectra based on transfer deep learning, which can, at least to a certain extent, effectively establish the correlation between the response power spectrum and the load power spectrum and ensure the accuracy of wind load identification.

[0005] According to one aspect of the present disclosure, there is provided a method for identifying high-rise structural wind load spectra based on transfer deep learning, including: Calculating based on the dynamic response time history data of the target structure under the target wind load to obtain the corresponding observed response power spectral matrix; Performing discrete cosine transform decomposition on the wind speed spectrum of the target wind load to obtain the dimension-reduced DCT coefficients; Constructing a deep learning model for wind load identification based on a physical guidance mechanism, using the observed response power spectral matrix as the input of the deep learning model and the DCT coefficients as the output, and training to enable the deep learning model to learn the mapping relationship between the two. Among them, the physical guidance mechanism embeds the frequency-domain transfer equation of the wind load and the structural response into the loss function to constrain the model training process; When performing wind load identification, using the trained deep learning model based on the input observed response power spectral matrix to output the corresponding DCT coefficients, and reconstructing the power spectral matrix of the wind load according to the DCT coefficients; After the deep learning model is trained, the method further includes: Based on the transfer learning strategy, for the wind loads at different discrete frequency points, the model parameters of the target frequency are initialized using the model parameters trained with the wind loads based on the source frequency, and the model parameters are fine-tuned with the target frequency data to obtain a deep learning model corresponding to the wind loads of the target frequency.

[0006] According to one aspect of the present disclosure, there is provided a device for identifying the wind load spectrum of a high-rise structure based on transfer deep learning, including: A calculation module, configured to calculate based on the dynamic response time history data of the target structure under the action of the target wind load to obtain a corresponding observed response power spectrum matrix; A decomposition module, configured to perform discrete cosine transform decomposition on the wind speed spectrum of the target wind load to obtain the dimension-reduced DCT coefficients; A training module, configured to construct a deep learning model for wind load identification based on a physical guidance mechanism, use the observed response power spectrum matrix as the input of the deep learning model, and the DCT coefficients as the output, and through training, enable the deep learning model to learn the mapping relationship between the two, wherein the physical guidance mechanism embeds the frequency domain transfer equation of the wind load and the structural response into the loss function to constrain the model training process; A processing module, configured to, when performing wind load identification, use the trained deep learning model based on the input observed response power spectrum matrix to output the corresponding DCT coefficients, and reconstruct the power spectrum matrix of the wind load according to the DCT coefficients; After the deep learning model is trained, the processing module is further configured to: Based on the transfer learning strategy, for the wind loads at different discrete frequency points, the model parameters of the target frequency are initialized using the model parameters trained with the wind loads based on the source frequency, and the model parameters are fine-tuned with the target frequency data to obtain a deep learning model corresponding to the wind loads of the target frequency.

[0007] According to another aspect of the present disclosure, there is provided an electronic device, including: a memory storing execution instructions; and a processor, the processor executing the execution instructions stored in the memory, such that the processor executes the method for identifying the wind load spectrum of a high-rise structure based on transfer deep learning according to any one of the embodiments of the present disclosure.

[0008] According to still another aspect of the present disclosure, there is provided a readable storage medium storing execution instructions, and when the execution instructions are executed by a processor, they are used to implement the method for identifying the wind load spectrum of a high-rise structure based on transfer deep learning according to any one of the embodiments of the present disclosure.

[0009] According to another aspect of the present disclosure, there is provided a computer program product including a computer program which, when executed by a processor, implements the method for identifying high-rise structural wind load spectra based on transfer deep learning according to any one of the embodiments of the present disclosure.

[0010] In the technical solutions provided in some embodiments of the present application, calculations are performed based on the dynamic response time history data of the target structure under the action of the target wind load to obtain the corresponding observed response power spectrum matrix, and the wind speed spectrum of the target wind load is decomposed by discrete cosine transform to obtain the dimension-reduced DCT coefficients. Then, a deep learning model for wind load identification based on a physical guidance mechanism is constructed. The observed response power spectrum matrix is used as the input of the deep learning model, and the DCT coefficients are used as its output. Through training, the wind load model can learn the mapping relationship between the two. Among them, the physical guidance mechanism constrains the model training process by embedding the frequency-domain transfer method of wind load and structural response into the loss function. When performing wind load identification, the trained deep learning model is used to output the corresponding DCT coefficients based on the input observed response power spectrum matrix, and the power spectrum matrix of the wind load is reconstructed according to the output DCT coefficients to achieve wind load identification.

[0011] In this way, by establishing a deep learning model, the connection between the response power spectrum and the load power spectrum is directly found, avoiding the ill-posed problem of the inverse process of traditional methods under noise and avoiding the use of regularization techniques, making the identification process more stable and accurate and improving its anti-noise performance.

[0012] Moreover, by introducing discrete cosine transform for dimension reduction, it is possible to identify the wind load spectra of complete degrees of freedom using partial structural response observations. Furthermore, by embedding a physical guidance mechanism in the deep learning model, the amount of data required for model training can be greatly reduced, while ensuring the accuracy of the identification results of the deep learning model and improving its training efficiency.

[0013] In addition, through the transfer learning strategy, the physical guidance feature extraction ability of the source frequency model is reused, and combined with fine-tuning of a small amount of data in the target frequency, efficient and high-precision identification of cross-frequency wind loads is achieved. This method significantly reduces data requirements and computational costs, providing an extensible technical path for full-band load identification in complex wind environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The drawings illustrate exemplary embodiments of the present disclosure and, together with the description thereof, are used to explain the principles of the present disclosure. These drawings are included to provide a further understanding of the present disclosure and are included in this specification and form a part of this specification.

[0015] Figure 1Shows a schematic flow chart of a high-rise structure wind load spectrum identification method based on transfer deep learning according to an embodiment of the present application; Figure 2 Shows a schematic flow chart of a high-rise structure wind load spectrum identification method based on transfer deep learning according to another embodiment of the present application; Figure 3 And Figure 4 Shows the network identification results for a stationary and uniform wind speed spectrum at different frequencies according to an embodiment of the present application; Figure 5 Shows a comparison schematic diagram of the loss function when there is and is not a physical guidance mechanism to constrain the network training process according to an embodiment of the present application; Figure 6 And Figure 7 Shows the network identification results for a stationary and non-uniform wind speed spectrum at different frequencies according to an embodiment of the present application; Figure 8 Shows a schematic block diagram of the structure of a high-rise structure wind load spectrum identification device based on transfer deep learning according to an embodiment of the present application; Figure 9 Shows a schematic diagram of the structure of a computer system of an electronic device suitable for implementing the embodiments of the present application. Detailed implementation manners

[0016] The present disclosure will be further described in detail below with reference to the accompanying drawings and examples. It can be understood that the specific examples described herein are only used to explain the relevant content and do not limit the present disclosure. Additionally, it should be noted that for the sake of convenience of description, only parts related to the present disclosure are shown in the drawings.

[0017] It should be noted that, without conflict, the embodiments in the present disclosure and the features in the embodiments may be combined with each other. The technical solutions of the present disclosure will be described in detail below with reference to the drawings and in combination with the embodiments.

[0018] In actual engineering, structures are constantly subjected to various forms of dynamic loads in nature. Therefore, the dynamic loads acting on structures play an important role in structural health monitoring, safety and reliability analysis, and dynamic optimization design. The loads in actual engineering are often affected by various factors and exhibit a certain degree of randomness, making it difficult to identify an explicit expression, and the identification difficulty is even greatly increased.

[0019] For dynamic loads, according to the nature of the loads, they can be divided into concentrated dynamic loads and distributed dynamic loads. Since distributed loads are not only dynamic in time, but also the time history and amplitude of the loads will change with the spatial position, their identification is more challenging. The most common distributed load in nature is the wind load. In recent years, some scholars have proposed some methods. For example, some related technologies use continuous and discrete Kalman filters, Taylor polynomial expansion to identify the pulsating wind load, but the pulsating wind needs to be regarded as a random walk process or zero-mean white noise. Some other related technologies construct a stable input estimator based on the Kalman filter to estimate the structural state and unknown input, and use the measured data of the Canton Tower for verification and extend it to complex structures. However, the above methods all identify the deterministic wind load.

[0020] Regarding random distributed loads, some scholars have also conducted research. The more common ones include frequency domain method, probability estimation, interval estimation, etc. For example, some related technologies consider the frequency domain identification model of frequency response function error and response measurement error, and analyze the error propagation and amplification effect of load identification, and find and control the errors in principle. Another part of scholars use B-spline functions to fit the spatial distribution function, establish the transfer relationship between excitation and response using the classical frequency response function method in the frequency domain, and this method can achieve good results for annular distributed dynamic loads and can be used for the dynamic load identification of the star-rocket connection interface. There is also a part of scholars who study methods such as Monte Carlo method, K-L expansion, etc. to identify the statistical characteristics such as the mean and variance of the load. Or use the interval estimation method, combined with the genetic algorithm of Latin hypercube sampling and the improved L-curve method to identify the upper and lower bounds of the distributed load. However, the above methods do not consider the spatial correlation of the distributed load, and many distributed dynamic loads in nature have spatial correlation, such as wind load, etc.

[0021] Although some methods consider the spatial correlation of the load on the basis of most existing methods, that is, the load power spectrum can follow an expression that changes with distance, fit this expression with orthogonal polynomials, establish the relationship between the polynomial coefficients and the response power spectrum, and reconstruct the load power spectrum by observing and identifying the coefficients. However, when calculating the load power spectrum from the response power spectrum by the above method, this inverse process often has an ill-posed problem under the influence of noise, and generally needs to introduce regularization technology, making the calculation more complex.

[0022] Therefore, this application provides a method for identifying the wind load spectrum of high-rise structures based on transfer deep learning to at least partially solve the above existing technical problems.

[0023] For the sake of convenience in description and to make the technical solutions of the specific embodiments of the present disclosure easier to understand, before describing the method for identifying the high-rise structural wind load spectrum based on transfer deep learning implemented in the present disclosure, the physical mechanisms involved in the specific embodiments of the present disclosure are explained as follows: For a high-rise building, the wind load acts on it in the form of a distributed load. When the aspect ratio of the structure is relatively large, the influence in the width direction can be ignored, and thus it can be regarded as a one-dimensional distributed dynamic load along the height direction. Considering the complexity of the actual structure, the finite element method is adopted to N discretize the high-rise building structure along the height; similarly, the wind load is also discretized. Let the number of discretized wind loads along the height direction be N .

[0024] When the wind load is a random load, its relationship can be described by the wind load power spectrum and the structural response power spectrum in the frequency domain: (1) Among them, ω is the circular frequency after the structural frequency conversion, represents the observed response power spectrum matrix of the structure; represents the power spectrum matrix of the wind load, is the structural frequency response function, represents the conjugate transpose of the structural frequency response function, A is the observation matrix, B is the wind load influence matrix, and the superscript T represents the transpose.

[0025] In an example, the structural frequency response function includes the frequency response function of the structural displacement response and the frequency response function of the structural acceleration response. It should be noted that the above structural frequency response function can be determined based on the dynamic response time history data of the selected structure. For example, if the dynamic response time history data of the selected structure is the displacement response, the corresponding structural frequency response function is the frequency response function of the structural displacement response; or, if the dynamic response time history data of the selected structure is the acceleration response, the corresponding structural frequency response function is the frequency response function of the structural acceleration response.

[0026] Among them, the frequency response function of the structural displacement response is as follows: , where ω is the circular frequency after the structural frequency conversion; M is the mass matrix of the structure; Q is the damping matrix of the structure; i is the imaginary unit with a value of -1; K is the stiffness matrix of the structure.

[0027] The frequency response function of the structural acceleration response is as follows: 。

[0028] Next, the wind load and wind speed have the following relationship in the time domain along the height direction: (2) where, fi ( t ) represents the wind load value at the t th discrete point along the height direction of the structure at i time, represents the air density, represents the height at the i th point, is the aerodynamic coefficient of , is the wind action area of ; represents the wind speed at the height of the i th point.

[0029] The wind speed is composed of the mean wind speed and the fluctuating wind speed : , (3) where, represents the mean wind speed at 10m, represents the height of 10m, is a constant between 0.15 - 0.5. Substituting formula (3) into formula (2) gives: (4) It should be noted that this application only considers the fluctuating wind load and assumes that the turbulence intensity is small, that is, the quadratic term of the fluctuating wind is ignored. Combining formula (3), formula (4) can be rewritten as: (5) where 。

[0030] The relationship between the power spectrum of the fluctuating wind load and the power spectrum of the wind speed is expressed as: (6) where, C is the relationship matrix between the fluctuating wind load spectrum and the fluctuating wind speed spectrum, and is a positive definite matrix; represents the power spectrum matrix of the fluctuating wind speed, and can be specifically expressed as: (7) where the diagonal elements are the auto-spectrum elements of the wind speed spectrum, and the rest are the cross-spectrum elements of the wind speed spectra at two different heights.

[0031] Substituting Equation (6) into Equation (1) gives: (8) This determines the connection between the observed response spectrum of the structure and the wind load spectrum.

[0032] However, different treatment methods are adopted for stationary uniform wind loads and non-uniform wind loads, which are specifically described as follows: I. Identification of uniform wind load spectrum: Since the uniform fluctuating wind speed spectrum is only related to the height difference between two points, therefore , there are only N independent auto-spectrum and cross-spectrum elements in the wind speed spectrum matrix. Therefore, the independent elements in the fluctuating wind speed spectrum matrix can be extracted to form a uniform fluctuating wind speed spectrum vector : (9) The embodiment of the present application uses Shitoni's coherence function for elaboration: (10) Wherein, and represent the heights of the i th point and the jth point. This coherence function is only related to the height difference. Take L z as a fixed value of 60, and exp(•) is the exponential function with e as the base.

[0033] A remarkable characteristic of the discrete cosine transform (DCT) is that it can concentrate most of the energy of the signal on a few transform coefficients. By converting the signal into the frequency domain, it can effectively reflect this energy concentration.

[0034] Then the DCT decomposition process of the elements of the uniform fluctuating wind speed spectrum vector is as follows: (11) Wherein, n represents the serial number of the element of the fluctuating wind speed spectrum vector, with a total of N ; k is the DCT order, represents the k th DCT coefficient; is the Kronecker delta function.

[0035] The reconstruction of the elements of the uniform fluctuating wind speed spectrum vector can be expressed in the following form: (12) Abbreviation: (13) Then the power spectral matrix of the pulsating wind speed (Equation (7)) can be expressed as: (14) According to the characteristic of DCT energy concentration, this application only considers the first p DCT coefficients, and the approximate value of the original wind load spectrum signal can be reconstructed. The truncated DCT coefficient vector is . From Equations (14) and (8), we can get: (15) Where represents the basis function matrix after truncation processing, that is, the matrix of the first p cosine basis function vectors.

[0036] Thus, the mapping relationship between the power spectral matrix of the observed response of the structure and the DCT coefficients after truncation of the pulsating stationary uniform wind speed spectrum decomposition is found.

[0037] II. Identification of non-uniform wind load spectrum The embodiment of this application adopts Davenport's exponential coherence function: (16) (17) Where and represent the average wind speeds at the i th point and the j th point respectively, and is the attenuation coefficient. This coherence function is related to frequency.

[0038] The two-dimensional DCT decomposition of the non-uniform pulsating wind speed spectrum matrix elements is as follows: (18) Where k , l represent the orders of the rows and columns of the DCT coefficient matrix respectively, and and have the same expression form .

[0039] Its corresponding reconstruction formula is as follows: (19) It can be expressed in matrix form: (20) Similarly, the complete wind load spectrum can be reconstructed by using the front DCT coefficients in the partial concentration, which can be expressed as: (21) Wherein, represents the fluctuating wind speed spectrum matrix reconstructed by the first coefficients, represents the first coefficients of the rows and columns of the matrix, D represents the first q rows of the orthogonal matrix represents the first D rows of the orthogonal matrix p rows.

[0040] Substituting formula (21) into formula (7) gives: (22) Since both the observed response power spectrum matrix and the DCT coefficient matrix are symmetric matrices, when the number of independent elements of the observed response power spectrum matrix is greater than the number of independent DCT coefficients, that is, , the coefficient matrix can be solved.

[0041] Thus, the mapping relationship between the response power spectrum of the structure and the DCT coefficients after decomposition and truncation of the fluctuating non-stationary uniform wind speed spectrum is found.

[0042] Based on the above description, Figure 1 FIG. shows a schematic flow chart of a method for identifying a wind load spectrum of a high-rise structure based on transfer deep learning according to an embodiment of the present application.

[0043] It should be noted that this method can be applied to a terminal device or a server. Among them, the terminal device may include, but is not limited to, one or more of a smart phone, a tablet computer, a portable computer, and a desktop computer; the server may be a physical server or a cloud server.

[0044] The following takes the application of this method to a terminal device as an example for description. As Figure 1 shown, the method for identifying a wind load spectrum of a high-rise structure based on transfer deep learning at least includes steps S110 to S140, which are introduced in detail as follows.

[0045] In step S110, according to the dynamic response time history data of the target structure under the action of the target wind load, the corresponding observed response power spectrum matrix is calculated.

[0046] Among them, the dynamic response time history data can be parameter information of the structure that changes with time under the action of wind load, which can include but is not limited to acceleration, velocity, displacement response, etc. In one example, the above dynamic response time history data can be obtained through on-site measurement by sensors or finite element simulation, such as an acceleration signal sampled 100 times per second, etc.

[0047] The observed response power spectral matrix can be a matrix that characterizes the energy distribution of the structural dynamic response in the frequency domain. The diagonal elements thereof are the response power spectral densities of each degree of freedom, and the non-diagonal elements are the cross-power spectral densities between different degrees of freedom, reflecting the correlation of the response in the frequency domain.

[0048] In this embodiment, the terminal can obtain the dynamic response time history data of the target structure under the action of the target wind load through on-site measurement by sensors or finite element simulation, and perform calculations based on this to obtain the corresponding observed response power spectral matrix.

[0049] In one example, after obtaining the dynamic response time history data, the terminal can first perform preprocessing on it, and the preprocessing can include but is not limited to denoising and filtering, segmentation and windowing. Specifically, the terminal can perform low-pass filtering on the original dynamic response time history data to eliminate high-frequency noise and interference signals. Then, the long-time data is segmented into several sub-segments, such as 1024 sampling points per segment, etc., and a Hanning window is applied to each segment to reduce spectral leakage.

[0050] Then, the terminal can perform a fast Fourier transform on each segment of the time history data to convert the time-domain signal into a frequency-domain complex spectrum, and then average the results of multiple sub-segments to reduce the influence of random noise. After the above processing, the terminal can calculate the auto-power spectral density according to the frequency-domain signal of each degree of freedom, and calculate the cross-power spectral density between the two according to the frequency-domain signals of different degrees of freedom. Then, the auto-power spectral densities of each degree of freedom and the non-diagonal cross-power spectral densities are arranged according to the degrees of freedom to form a symmetric observed response power spectral matrix.

[0051] In step S120, the wind speed spectrum of the target wind load is decomposed by discrete cosine transform to obtain the dimension-reduced DCT coefficients.

[0052] Among them, the wind speed spectrum can be the statistical characteristics describing the change of wind speed with frequency, representing the distribution of the pulsating energy of the wind speed at different frequencies. It is obtained by Fourier transform to convert the time history of the wind speed into the frequency domain, and is usually used to analyze the turbulence characteristics of the wind.

[0053] The discrete cosine transform (DCT) is a signal processing technology that realizes energy concentration by decomposing a signal into cosine function components of different frequencies, and is suitable for the compression and feature extraction of high-dimensional data.

[0054] The DCT coefficients are the coefficients obtained after DCT transformation, which are used to characterize the projection weights of the original signal on different frequency cosine basis functions. The front coefficients usually contain the main energy of the signal.

[0055] In this embodiment, the terminal can obtain the wind speed spectrum (vector or matrix) corresponding to the target wind load, and then decompose it through discrete cosine transform to obtain the DCT coefficients after dimensionality reduction. In this way, the wind speed spectrum of the high-dimensional wind load is compressed into low-dimensional coefficients through DCT transformation, significantly reducing the data complexity while retaining the core physical information, and providing efficient and noise-resistant input features for the machine learning model.

[0056] In some embodiments of the present application, decomposing the wind speed spectrum of the target wind load through discrete cosine transform to obtain the DCT coefficients after dimensionality reduction includes: For the wind speed spectrum vector of the uniform wind load, use one-dimensional DCT coefficients to decompose it, and truncate and retain the first p principal coefficients; For the wind speed spectrum matrix of the non-uniform wind load, use two-dimensional DCT coefficients to decompose it, and truncate and retain the first q × p principal coefficients, where q and p are the truncation orders of the rows and columns respectively.

[0057] In this embodiment, for the wind speed spectrum of the uniform wind load, since the elements in its vector are only related to the position height difference, one-dimensional DCT coefficients can be used to decompose it, and the first p principal coefficients can be truncated and retained. For the wind speed spectrum matrix of the non-uniform wind load, since the elements in its matrix are related to both frequency and spatial position, two-dimensional DCT coefficients can be used to decompose it, and the first q × p principal coefficients can be truncated and retained, where q and p are the truncation orders of the rows and columns respectively. It should be understood that the energy concentration characteristic of the DCT coefficients makes its main information concentrated in the front coefficients and the noise dispersed in the high-frequency coefficients. Therefore, after truncating it, the influence of noise can be effectively suppressed.

[0058] Please continue to refer to Figure 1 In step S130, a deep learning model for wind load identification based on a physics-guided mechanism is constructed. The observed response power spectrum matrix is used as the input of the deep learning model, and the DCT coefficients are used as the output. Through training, the deep learning model is enabled to learn the mapping relationship between the two, where the physics-guided mechanism embeds the frequency-domain transfer equation of the wind load and the structural response into the loss function to constrain the model training process.

[0059] Among them, the physical guidance mechanism embeds physical laws (such as the frequency-domain transfer equation) into the machine learning model training process, and enhances its physical rationality and generalization ability by constraining the model output.

[0060] It should be noted that the frequency-domain transfer equation is used to describe the relationship between the structural response power spectrum and the wind load power spectrum.

[0061] The deep learning model can be a model pre-constructed by those skilled in the art, which is used to output the corresponding DCT coefficients according to the input observed response power spectrum matrix, so as to realize wind load identification.

[0062] In this embodiment, the terminal can train the pre-constructed deep learning model for wind load identification according to the observed response power spectrum matrix and its corresponding DCT coefficients obtained in the foregoing steps. Specifically, the observed response power spectrum matrix can be used as the input of the deep learning model, and through training, it is made to output the corresponding DCT coefficients. The parameters of the deep learning model can be optimized by minimizing the loss function, thereby improving the accuracy of the output results of the deep learning model.

[0063] In an example, the loss function of the deep learning model can include a data error term (such as the difference between the predicted DCT coefficient and the true value) and a physical constraint term (such as the residual of the frequency-domain transfer equation).

[0064] In this way, by forcing the output of the model to conform to the frequency-domain transfer equation through the physical constraint term, the accuracy of the model output can be effectively improved. Moreover, based on the physical guidance mechanism, the dependence of the deep learning model on large-scale data can be reduced, and the amount of training data required can be reduced while ensuring the output accuracy.

[0065] Please continue to refer to Figure 1 , in step S140, when performing wind load identification, the trained deep learning model is used to output the corresponding DCT coefficients based on the input observed response power spectrum matrix, and the power spectrum matrix of the wind load is reconstructed according to the DCT coefficients.

[0066] In this embodiment, after the deep learning model is trained, the terminal can perform wind load identification based on it. Specifically, the terminal can input the observed response power spectrum matrix obtained by actual measurement or simulation into the deep learning model, and the deep learning model can perform inference based on it to output the corresponding DCT coefficients. Then, the terminal can reconstruct the output DCT coefficients to obtain the corresponding power spectrum matrix of the wind load, so as to realize wind load identification.

[0067] In some embodiments of the present application, reconstructing the power spectrum matrix of the wind load according to the DCT coefficients includes: Substitute the DCT coefficients output by the deep learning model into the inverse DCT formula to restore the complete power spectral matrix of uniform or non-uniform wind loads.

[0068] In this embodiment, the terminal can substitute the DCT coefficients output by the deep learning model into the corresponding inverse DCT formula, so as to restore them to the original signal or matrix through inverse operation, and then calculate the corresponding power spectral matrix of the wind load.

[0069] In this way, the trained model realizes the end-to-end recognition from the response power spectrum to the wind load spectrum. Combining the inverse DCT transformation and error control, it significantly improves the calculation efficiency while ensuring high accuracy.

[0070] In some embodiments of the present application, after the deep learning model is trained, the method further includes: Based on the transfer learning strategy, for the wind loads at different frequency discrete points, use the model parameters trained by the wind loads based on the source frequency to initialize the model of the target frequency, and fine-tune the model parameters through the target frequency data to obtain the deep learning model corresponding to the wind loads of the target frequency.

[0071] In this embodiment, the transfer learning strategy can be to transfer the model parameters trained in the source domain (such as the wind load at a certain frequency) to the target domain (such as the wind load at another frequency), and use the existing knowledge to accelerate the learning process of the new task. That is to say, the terminal can fully train the deep learning model according to the training data set obtained from the wind loads based on the source frequency. It should be understood that this training data set includes several groups of observed response power spectral matrices and their corresponding DCT coefficients.

[0072] When it is necessary to identify the wind loads of the target frequency (different from the source frequency), the terminal can reuse the model parameters trained for the wind loads of the source frequency, and fine-tune the model parameters according to a small amount of the training data set corresponding to the wind loads of the target frequency, so as to obtain the deep learning model corresponding to the wind loads of the target frequency.

[0073] In this way, the embodiment of the present application realizes the efficient and high-precision identification of cross-frequency wind loads through the transfer learning strategy, reusing the physical-guided feature extraction ability of the source frequency model and combining with the fine-tuning of a small amount of data of the target frequency. This method significantly reduces the data requirements and calculation costs, and provides an extensible technical path for the full-band load identification in complex wind environments.

[0074] Based on the technical solutions of the above embodiments, a specific application scenario of the embodiments of the present application is introduced below: Figure 2 Fig. shows a schematic flow chart of a method for identifying the wind load spectrum of a high-rise structure based on transfer deep learning according to another embodiment of the present application.

[0075] The following is an explanation of the technical terms in the attached drawings first: Wind load spectra: Represents the distribution of the force exerted by the wind on the structure in the frequency domain, and describes the variation of the pulsating energy of the wind load with frequency.

[0076] Wind velocity spectra: Describes the pulsating characteristics of the wind speed, that is, the distribution of the pulsating energy of the wind speed at different frequencies.

[0077] Wind load: That is, the pressure or suction force generated by the wind on the building or structure.

[0078] Finite element method (FEM): Uses the finite element method to simulate the response of the structure under the action of wind load and obtains the time history data of the structural response.

[0079] Structural response: The time history data of the response of the structure under the action of wind load obtained by simulation using the finite element method, such as displacement, acceleration, etc.

[0080] Discrete cosine transform (DCT): Performs DCT decomposition on the wind speed spectra or wind load spectra to extract independent DCT coefficients. DCT can concentrate most of the energy of the signal on a few coefficients, thus achieving dimensionality reduction.

[0081] Observed response spectra: The representation in the frequency domain of the response data of the structure under the action of wind load obtained through actual measurement or simulation. These response data usually include the acceleration, velocity, displacement, etc. of the structure.

[0082] Training / Testing: Uses the observed response spectra as input and the DCT coefficients as output to train a physics-guided convolutional neural network (CNN). A physical loss function is embedded during the training process to reduce the number of training sets and improve the training efficiency.

[0083] Data Loss: Represents the difference between the predicted wind load spectra and the true values, and is used to evaluate the prediction accuracy of the model.

[0084] Physical guided CNN: A CNN network embedded with a physical model, which guides the network training through a physical loss function to improve the accuracy and generalization ability of the model.

[0085] Physical Loss: Embed a physical model (such as the relationship between wind load and structural response) into the loss function, enabling the network to consider physical laws during training and improving the stability and accuracy of the model.

[0086] Uniform spectrum: A frequency spectrum in which the wind speed spectrum or wind load spectrum has a uniform distribution characteristic in space.

[0087] Non-uniform spectrum: That is, the distribution of the wind speed spectrum or wind load spectrum in space is non-uniform, meaning that the statistical characteristics of the wind speed or wind load vary with position.

[0088] In one example, for a uniform spectrum, the corresponding Physical Loss is: .

[0089] Where β is the weight coefficient; represents the observed response power spectral matrix of the structure at frequency ω ; W ( ω ) is the coefficient vector at the corresponding frequency, is the generalized orthogonal basis function; For a non-uniform spectrum, the corresponding Physical Loss is: .

[0090] It should be understood that at this time is in matrix form, , respectively represent the first p rows and the first q rows of the orthogonal basis function matrix.

[0091] Transfer Learning: Utilize the model trained at a certain frequency and transfer it to other frequencies for recognition. Through transfer learning, only a small amount of training sets need to be prepared at other frequencies to complete the wind load spectrum recognition across all frequencies, greatly improving the training efficiency.

[0092] Such as Figure 2As shown, in this method for identifying the high-rise structural wind load spectrum based on transfer deep learning, first, starting from the wind speed spectrum, the finite element method (FEM) is used to simulate the structural response under the action of wind load to obtain the structural response time history data. The frequency domain analysis is performed on the structural response data to obtain the observed response spectrum. At the same time, the discrete cosine transform (DCT) is performed on the wind speed spectrum to extract the DCT coefficients for dimensionality reduction processing.

[0093] Then, taking the obtained observed response spectrum as the input and the DCT coefficients as the output, a physics-informed convolutional neural network (CNN) is used for training. During the training process, a physical loss function is introduced, which is based on the physical relationship between the wind load and the structural response, thereby reducing the number of required training sets and improving the stability and accuracy of the model.

[0094] When the training is completed, for the new observed response spectrum data, they can be input into the trained network, and the network will output the corresponding DCT coefficients. By reconstructing these coefficients, an estimated value of the wind load spectrum can be obtained, thereby realizing the identification of the wind load.

[0095] Taking a height of 10 m as an example, by changing the wind speed at different 10 m positions, the fluctuating wind speed spectrum at a certain frequency is generated, and then the observed response spectrum of the structure is calculated. The corresponding fluctuating wind speed spectrum vector (matrix) is decomposed and truncated by one-dimensional (two-dimensional) DCT to obtain the DCT coefficient vector (matrix). Then, given the generalized cosine basis function matrix, a physics-informed neural network for identifying uniform (non-uniform) random wind speed spectra is trained by embedding the physical loss function. The network input and output are the observed response spectrum of the structure at a certain frequency and the DCT coefficient vector (matrix) in the previous steps, respectively.

[0096] Next, observe the structural response time history under the wind load generated by other forms of fluctuating wind speed spectra at a certain wind speed at 10 m, and calculate the corresponding observed response power spectrum at this frequency to form a test set. The calculated observed structural response power spectrum at this frequency is input to the trained network for testing, so that it outputs the DCT coefficient vector (matrix). Then, a uniform (non-uniform) fluctuating wind speed spectrum at a certain frequency is reconstructed according to the DCT coefficient vector (matrix).

[0097] When it is necessary to extend to other frequencies, the transfer learning method is used. The parameters of the network training in the above process are used as the initial values, a small sample data set is made at other frequencies, and the previous steps are repeated until all frequency discrete points are traversed to complete the identification of the full-band random stationary uniform (non-uniform) wind speed spectrum.

[0098] That is, as in Figure 2As shown in the lower part, after the model is fully trained at a specific frequency, the model can be transferred to other frequencies for recognition. At other frequencies, only a small amount of training set needs to be prepared for fine-tuning to complete the wind load spectrum recognition within the full frequency range, greatly improving the training efficiency and saving a large amount of time and computing resources. Specifically, after full training at a certain frequency (such as ω1) to obtain a trained model, the model is transferred to other frequencies (such as ω2, ω3, etc.), and only a small amount of training set needs to be prepared for fine-tuning at these frequencies. Thus, through transfer learning, the wind load spectrum recognition within the full frequency range can be efficiently completed, saving a large amount of training time and computing resources.

[0099] Based on the method provided in the foregoing embodiments, in order to verify the effectiveness and accuracy of the method, the 76-benchmark model is used in the numerical example. The 76-story wind-induced vibration benchmark building established by the International Association for Structural Control and the American Society of Civil Engineers (IASC-ASCE) will not be elaborated here.

[0100] For the recognition network for the uniform wind speed spectrum, it is trained using the Davenport spectrum and tested using the Harris spectrum. There are 5 groups in the test set, and the L2 norm is used as the recognition accuracy index. First, the network is trained with the training set at 1 Hz (the recognition results are as Figure 3 shown, where Identification Value is the estimated value of the wind load spectrum and True Value is the true value of the wind load spectrum; the abscissa in the figure represents the auto-spectrum or cross-spectrum between any two points on the structure. For example, S 1−1 represents the auto-spectrum between the first point and the first point on the structure, and S 1−10 represents the cross-spectrum between the first point and the 10th point on the structure, and so on. The ordinate is the amplitude of the wind load spectrum, with the unit of m 2 / Hz, which reflects the energy distribution of the wind load at the corresponding frequency), and then transferred to 4 Hz (the recognition results are as Figure 4 shown). The average recognition accuracy is above 90%. The recognition accuracy of each group of test sets is shown in Table 1 below.

[0101] Table 1: Recognition accuracy of stationary uniform wind speed spectra at different frequencies To further verify the superiority of physically-guided embedding of the objective function into the loss function, two training sets can be used for training respectively. The number of samples in Training Set 1 is 20,000, and the physical loss function is not embedded; the number of samples in Training Set 2 is 5,000, and the physical loss function is embedded. The most intuitive difference lies in the descent rate of the loss function and the training duration. Among them, the training time for 100 rounds in Training Set 1 is 737.91 seconds, while the training time for 100 rounds in Training Set 2 is 226.13 seconds. The changing trend of the loss function during the training process is as Figure 5 (where Training dataset 1 is Training Set 1 and Training dataset 2 is Training Set 2. The abscissa is the number of training rounds (i.e., Training epochs), and the ordinate is the value of the loss function (i.e., Value)). The blue dashed line is the loss function curve of Training Set 1, while the red solid line is the loss function curve of Training Set 2. It can be seen that the descent rate of the loss function curve of the network with the embedded loss function is significant, and the correct gradient can be found relatively quickly, and the value of the loss function is smaller in the first 100 rounds of training.

[0102] For the identification network for non-uniform wind speed spectra, it is trained using the Von-Karman spectrum and tested using the Kaimal spectrum. The identification results of non-uniform wind speed spectra at different frequencies are as Figure 6 and Figure 7 shown, where Figure 6 is the identification result at a frequency of 1 Hz (where Identification is the identification value, True Value is the true value, and Relative Error is the relative error), Figure 7 is the identification result at a frequency of 4 Hz. It can be seen that the error between the identification value and the true value is relatively small.

[0103] To further verify the advantages of transfer learning in the application of this method, Table 2 compares the original pre-training time with the transfer learning time at other discrete frequency points. The pre-training set uses 5,000 groups of samples, the number of samples in the transfer learning training set is 500, and the number of training rounds is 1,000. It can be seen that after the pre-training process is completed, the time required for transfer training is significantly reduced, the training efficiency is greatly improved, and a large amount of computing resources and time are saved.

[0104] Table 2: Transfer learning time of stationary non-uniform wind speed spectra at different frequencies Based on any of the above embodiments, the present disclosure also provides a device for identifying high-rise structural wind load spectra based on transfer deep learning.

[0105] Figure 8It is a structural schematic block diagram of a high-rise structure wind load spectrum recognition device based on transfer deep learning according to an embodiment of the present disclosure.

[0106] As Figure 8 shown, the high-rise structure wind load spectrum recognition device based on transfer deep learning includes: A calculation module, configured to calculate based on the dynamic response time history data of the target structure under the action of the target wind load to obtain a corresponding observed response power spectrum matrix; A decomposition module, configured to perform discrete cosine transform decomposition on the wind speed spectrum of the target wind load to obtain the dimension-reduced DCT coefficients; A training module, configured to construct a deep learning model for wind load recognition based on a physical guidance mechanism, use the observed response power spectrum matrix as the input of the deep learning model, and the DCT coefficients as the output, and through training, enable the deep learning model to learn the mapping relationship between the two, wherein the physical guidance mechanism embeds the frequency-domain transfer equation of the wind load and the structural response into the loss function to constrain the model training process; A processing module, configured to, when performing wind load recognition, use the trained deep learning model based on the input observed response power spectrum matrix to output the corresponding DCT coefficients, and reconstruct the power spectrum matrix of the wind load according to the DCT coefficients.

[0107] It should be noted that the above high-rise structure wind load spectrum recognition device based on transfer deep learning may be in the form of computer software, and each module of the above high-rise structure wind load spectrum recognition device may be implemented by computer software modules.

[0108] In some embodiments of the present disclosure, performing discrete cosine transform decomposition on the wind speed spectrum of the target wind load to obtain the dimension-reduced DCT coefficients includes: For the wind speed spectrum vector of the uniform wind load, perform one-dimensional DCT decomposition on it and truncate and retain the first p main coefficients; For the wind speed spectrum matrix of the non-uniform wind load, perform two-dimensional DCT decomposition on it and truncate and retain the first q×p main coefficients, where q and p are the truncation orders of the rows and columns respectively.

[0109] In some embodiments of the present disclosure, after the deep learning model is trained, the processing module is further configured to: Based on the transfer learning strategy, for the wind loads at different discrete frequencies, the model parameters of the target frequency are initialized using the model parameters trained with the wind loads at the source frequency, and the model parameters are fine-tuned with the target frequency data to obtain the deep learning model corresponding to the wind loads at the target frequency.

[0110] In some embodiments of the present disclosure, the frequency-domain transfer equation is as follows: Wherein, is the power spectral matrix of the observed response of the structure; represents the power spectral matrix of the wind load, is the frequency response function of the structure, represents the conjugate transpose of the frequency response function of the structure, A is the observation matrix, B is the wind load influence matrix, and the superscript T represents the transpose.

[0111] In some embodiments of the present disclosure, reconstructing the power spectral matrix of the wind load according to the DCT coefficients includes: Substituting the DCT coefficients output by the deep learning model into the inverse DCT formula to restore the complete power spectral matrix of the uniform or non-uniform wind load.

[0112] In some embodiments of the present disclosure, the dynamic response time history data includes acceleration, velocity, or displacement responses.

[0113] The implementation processes of the functions and roles of each module in the above device are specifically detailed in the implementation processes of the corresponding steps in the above method, and will not be elaborated here.

[0114] Figure 9 The structural schematic diagram of the computer system of the electronic device suitable for implementing the embodiments of the present application is shown.

[0115] It should be noted that Figure 9 The computer system of the electronic device shown is only an example and should not bring any limitations to the functions and usage scopes of the embodiments of the present application.

[0116] As Figure 9As shown, the computer system includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes according to the program stored in the Read-Only Memory (ROM) 302 or the program loaded from the storage section 308 into the Random Access Memory (RAM) 303, such as executing the methods described in the above embodiments. In the RAM 303, various programs and data required for system operation are also stored. The CPU 301, ROM 302, and RAM 303 are connected to each other via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0117] The following components are connected to the I / O interface 305: an input section 306 including a keyboard, a mouse, etc.; an output section 307 including, for example, a Cathode Ray Tube (CRT), a Liquid Crystal Display (LCD), etc., and a speaker, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the I / O interface 305 as needed. A removable medium 311, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 310 as needed so that a computer program read from it can be installed into the storage section 308 as needed.

[0118] Specifically, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network via the communication section 309, and / or installed from the removable medium 311. When the computer program is executed by the Central Processing Unit (CPU) 301, various functions defined in the system of the present application are executed.

[0119] It should be noted that the computer-readable medium shown in the embodiments of the present application can be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of a computer-readable storage medium can include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, apparatus, or device. In the present application, a computer-readable signal medium can include a data signal propagated in a baseband or as part of a carrier wave, which carries a computer-readable computer program. Such a propagated data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, and this computer-readable medium can send, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device. The computer program contained on a computer-readable medium can be transmitted by any suitable medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

[0120] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present application. Among them, each block in the flowchart or block diagram can represent a module, a program segment, or a part of code, and the above module, program segment, or part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order from that marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram or flowchart, and the combination of blocks in the block diagram or flowchart, can be implemented by a dedicated hardware-based system for performing the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0121] The units involved in the embodiments of the present application can be implemented in software or in hardware, and the described units can also be provided in a processor. In some cases, the names of these units do not constitute a limitation on the units themselves.

[0122] On the other hand, the present application also provides a computer-readable medium, which may be included in the electronic device described in the above embodiments, or may exist alone without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by an electronic device, the electronic device implements the method described in the above embodiments.

[0123] The present disclosure also provides a computer program product. The method of the present disclosure can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed, the processes or functions of the present disclosure are executed in whole or in part.

[0124] The computer program or instructions can be stored in a readable storage medium or transmitted from one readable storage medium to another. For example, the computer program or instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired or wireless manner. The readable storage medium can be any available medium that can be accessed or a data storage device such as a server or data center integrating one or more available media. The available medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; it can also be an optical medium, such as a digital video disc; or it can be a semiconductor medium, such as a solid-state drive. The computer-readable storage medium can be a volatile or non-volatile storage medium, or can include both volatile and non-volatile types of storage media.

[0125] Those skilled in the art should understand that the embodiments of the present disclosure can be provided as a method, a system, or a computer program product. Therefore, the present disclosure can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present disclosure can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0126] The present disclosure is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the present disclosure. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or in multiple blocks.

[0127] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing devices to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufacture including an instruction means, and the instruction means implements the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or in multiple blocks.

[0128] These computer program instructions can also be loaded onto a computer or other programmable data processing devices, such that a series of operation steps are executed on the computer or other programmable devices to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable devices provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or in multiple blocks.

[0129] In the description of this specification, the descriptions with reference to the terms "one embodiment / way", "some embodiments / ways", "example", "specific example", or "some examples", etc. mean that the specific features, structures, or characteristics described in connection with the embodiment / way or example are included in at least one embodiment / way or example of the present disclosure. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment / way or example. Moreover, the specific features, structures, or characteristics described can be combined in a suitable manner in any one or more embodiments / ways or examples. In addition, without conflict, those skilled in the art can combine and combine the different embodiments / ways or examples described in this specification and the features of different embodiments / ways or examples.

[0130] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In the description of the present disclosure, "a plurality of" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0131] Those skilled in the art should understand that the above-described embodiments are merely for clearly illustrating the present disclosure and are not intended to limit the scope of the present disclosure. For those skilled in the art, other changes or variations can be made based on the above disclosure, and these changes or variations are still within the scope of the present disclosure.

Claims

1. A method for identifying the wind load spectrum of high-rise structures based on transfer deep learning, characterized in that, Including: Calculating based on the dynamic response time history data of the target structure under the action of the target wind load to obtain the corresponding observed response power spectral matrix; Performing discrete cosine transform decomposition on the wind speed spectrum of the target wind load to obtain the reduced-dimensional DCT coefficients; Constructing a deep learning model for wind load identification based on a physics-guided mechanism, using the observed response power spectral matrix as the input of the deep learning model and the DCT coefficients as the output, and training to enable the deep learning model to learn the mapping relationship between the two. Among them, the physics-guided mechanism embeds the frequency-domain transfer equation of the wind load and the structural response into the loss function to constrain the model training process; When performing wind load identification, using the trained deep learning model to output the corresponding DCT coefficients based on the input observed response power spectral matrix, and reconstructing the power spectral matrix of the wind load according to the DCT coefficients; After the deep learning model is trained, the method further includes: Based on the transfer learning strategy, for wind loads at different frequency discrete points, initializing the model at the target frequency with the model parameters trained by the wind load based on the source frequency, and fine-tuning the model parameters with the target frequency data to obtain the deep learning model corresponding to the wind load at the target frequency.

2. The method according to claim 1, characterized in that, Performing discrete cosine transform decomposition on the wind speed spectrum of the target wind load to obtain the reduced-dimensional DCT coefficients, including: For the wind speed spectrum vector of the uniform wind load, perform one-dimensional DCT decomposition on it and truncate and retain the first p principal coefficients; For the wind speed spectrum matrix of non-uniform wind loads, two-dimensional DCT is used to decompose it, and the first q×p principal coefficients are truncated and retained, where q and p are the truncation orders of rows and columns respectively.

3. The method according to claim 1, wherein The frequency-domain transfer equation is as follows: Among them, is the observed response power spectrum matrix of the structure; represents the power spectrum matrix of the wind load, is the frequency response function of the structure, represents the conjugate transpose of the frequency response function of the structure, A is the observation matrix, B is the wind load influence matrix, and the superscript T represents the transpose.

4. The method according to claim 1, wherein Reconstructing the power spectral matrix of the wind load according to the DCT coefficients, including: Substituting the DCT coefficients output by the deep learning model into the inverse DCT formula to restore the complete power spectral matrix of the uniform or non-uniform wind load.

5. The method according to claim 1, wherein The dynamic response time history data includes acceleration, velocity or displacement responses.

6. An apparatus for identifying the wind load spectrum of a high-rise structure based on transfer deep learning, characterized in that, Including: A calculation module for calculating based on the dynamic response time history data of the target structure under the action of the target wind load to obtain the corresponding observed response power spectral matrix; A decomposition module for performing discrete cosine transform decomposition on the wind speed spectrum of the target wind load to obtain the reduced-dimensional DCT coefficients; A training module for constructing a deep learning model for wind load identification based on a physics-guided mechanism, using the observed response power spectral matrix as the input of the deep learning model and the DCT coefficients as the output, and training to enable the deep learning model to learn the mapping relationship between the two. Among them, the physics-guided mechanism embeds the frequency-domain transfer equation of the wind load and the structural response into the loss function to constrain the model training process; A processing module for, when performing wind load identification, using the trained deep learning model to output the corresponding DCT coefficients based on the input observed response power spectral matrix, and reconstructing the power spectral matrix of the wind load according to the DCT coefficients; After the deep learning model is trained, the processing module is further used for: Based on the transfer learning strategy, for wind loads at different frequency discrete points, initializing the model at the target frequency with the model parameters trained by the wind load based on the source frequency, and fine-tuning the model parameters with the target frequency data to obtain the deep learning model corresponding to the wind load at the target frequency.

7. An electronic device, characterized in that, Including: A memory that stores execution instructions; and a processor that executes the execution instructions stored in the memory, such that the processor executes the method according to any one of claims 1 to 5.

8. A readable storage medium, characterized in that, Execution instructions are stored in the readable storage medium, and when the execution instructions are executed by a processor, they are used to implement the method according to any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 5.

Images