Structural full-degree-of-freedom vibration mode identification method based on a small amount of measurement point acceleration data and generative adversarial network

Abstract

The present invention proposes a method for identifying structural full-degree-of-freedom vibration shapes based on a small amount of measurement point acceleration data and a generative adversarial network. First, a generative adversarial network model is trained to learn a large data set of simulated vibration shapes of a structural finite element model under various possible values ​​of physical parameters. The generator input is a randomly sampled vector z in a low-dimensional space, and the output is a high-dimensional simulated full-degree-of-freedom vibration shape data φ. Secondly, a mapping relationship between vibration shapes and monitoring data is established based on dynamic knowledge, and an optimization objective function is constructed. Finally, the trained generator G represents the characteristics of the full-degree-of-freedom vibration shape of the structure through a relatively low-dimensional vector z, and the optimal full-degree-of-freedom vibration shape is obtained by directly searching in the z space. This method uses the generator as a physical model-driven constraint item of the vibration shape φ to be identified, and automatically learns physical information related to vibration shape identification from finite element simulation data to improve the ill-posedness of the inversion problem of full-degree-of-freedom modal identification.

Description

Technical Field

[0001] The present invention belongs to the technical field of structural modal recognition and machine learning, and in particular relates to a method for structural full-degree-of-freedom vibration mode recognition based on a small amount of measurement point acceleration data and a generative adversarial network. Background Art

[0002] my country has made great achievements in infrastructure construction, and its scale ranks first in the world. However, under the combined effects of environmental erosion, material aging, long-term effects of loads and component defects, infrastructure inevitably accumulates damage, resulting in a continuous reduction in its bearing capacity, a continuous loss of service functions, and even endangering the safe use of the structure. Ensuring the service safety of major infrastructure and extending its life is a major issue related to the national economy and people's livelihood. Structural health monitoring effectively ensures the service safety and life of major infrastructure by deploying large-scale, distributed sensor networks and installing data acquisition, transmission, management and analysis systems to perceive, diagnose and evaluate the damage evolution and safety status of the structure on site. In structural health monitoring, structural modal identification uses the collected structural response data and possible excitation loads to estimate the mathematical model and parameters of the structural dynamic system, which is conducive to a deep understanding of the structure and is crucial to ensuring the service safety of the structure. It is the core issue and important content of structural health monitoring.

[0003] Modal parameters such as structural vibration modes are the bridge between the dynamic response of the structure and physical parameters such as stiffness and damping. In structural modal identification, we often need to infer the structural modal parameters from the dynamic response of the structure under unknown loads, and further identify the physical parameters of the structure to perform health diagnosis and safety assessment of the structure. The spatial resolution of the structural vibration mode, that is, the number of degrees of freedom identified, determines the accuracy of the identified physical parameters (such as the resolution of structural stiffness identification). However, in structural modal identification, there are many false modes and the spatial resolution of the vibration mode (number of degrees of freedom) is low due to the arrangement of sensor measurement points. Therefore, the vibration mode vector containing more degrees of freedom, especially the full-degree-of-freedom vibration mode, becomes an important source of ill-posedness in the structural system identification problem due to its high-dimensional characteristics. For each structure, although the vibration mode of the structural system has a certain simple fixed mode, it does not meet the sparsity characteristics, so the sparse inversion of zero-norm and one-norm constraints cannot be used. In addition, it is difficult to produce an effective constraint effect using the two-norm regularization term. In the case of severe incompleteness of observation data information, in order to further limit the solution space of structural vibration modes, theoretically, the structural finite element model can be used to numerically simulate the full-degree-of-freedom vibration mode data of the structure based on various possible physical parameters, that is, the solution is reduced to all vibration modes with "structural dynamics physical meaning", and a reasonable structural vibration mode solution is searched in an exhaustive manner, but it is unrealistic due to the huge computational complexity. Therefore, we need to further find effective constraint means to achieve accurate identification of the full-degree-of-freedom vibration mode of the structure, which is an important way to improve the accuracy and robustness of structural physical parameter identification. The deep learning generative model developed in recent years provides an effective means to solve this problem.

[0004] In 2014, Goodfellow et al. proposed the Generative Adversarial Network (GAN). Based on the idea of ​​zero-sum game in game theory, the GAN innovatively uses adversarial training methods to automatically learn the distribution of real data in a certain latent space (which is generally a low-dimensional vector space), and shows excellent data generation performance. The trained GAN can sample from the latent space and generate a large amount of data close to the real distribution. In the past five years, GAN has received a lot of attention and innovation in the fields of super-resolution, image editing, and one-dimensional data enhancement. Compared with traditional generative models, GAN uses a discriminator to determine whether the generated data is realistic. Therefore, there is no need to assume the distribution of real data. Instead, it samples in a low-dimensional vector space that obeys a random distribution, and uses adversarial training to make the generator learn the mapping relationship between the low-dimensional vector (that is, the generator's latent feature space vector) and the distribution of the real training data set, thereby obtaining a generative model that can generate a large amount of realistic data. The generator can capture the implicit physical features of the training data set through a low-dimensional latent feature space, and has obvious learning and capture capabilities for those features that have concise feature patterns but are difficult to describe with explicit mathematical expressions.

[0005] Considering that the structural vibration mode has a non-analytical fixed pattern, and the generative adversarial network has a powerful low-dimensional vector representation ability in high-dimensional data generation tasks such as audio, image and video, the present invention studies the high-dimensional data of the full-degree-of-freedom vibration mode of the structure, and uses the generative adversarial network model to learn a large data set of simulated vibration modes of a certain structural finite element model under various possible values ​​of the structural physical parameters. The input of the generator is a randomly sampled vector z in the low-dimensional space, and the output is a high-dimensional simulated full-degree-of-freedom vibration mode data. If the generator model can obtain the characteristics of the full-degree-of-freedom vibration mode of the structure through a relatively low-dimensional vector z, it can directly search in the z space to obtain the optimal one, and it is not necessary to directly identify the full-degree-of-freedom vibration mode. Therefore, the generator model here serves as a physical model-driven constraint item for identifying the vibration mode, that is, automatically learning useful vibration mode information from the finite element simulation data to improve the ill-posedness of the inverse problem of vibration mode identification. The present invention will establish a generative adversarial network model that maps a large data set of full-degree-of-freedom vibration modes that takes into account the uncertainty of structural parameters into a low-dimensional representation vector. Based on the physical constraints of the structural vibration mode dynamics implied in the model and the optimization of the low-dimensional vector z, the identification of structural full-degree-of-freedom vibration modes based on generative adversarial networks and its corresponding system identification method are studied. Summary of the invention

[0006] The present invention aims to solve the problems in the prior art and proposes a method for identifying the full-degree-of-freedom vibration mode of a structure based on a small amount of measurement point acceleration data and a generative adversarial network. The method is suitable for identifying the full-degree-of-freedom mode vibration mode of a structure based on a small amount of spatial sensor monitoring data.

[0007] The present invention is implemented by the following technical scheme. The present invention proposes a method for identifying the full-degree-of-freedom vibration mode of a structure based on a small amount of measurement point acceleration data and a generative adversarial network. The method comprises the following steps:

[0008] Step 1: Randomly sample the physical parameters of the structural finite element model, calculate and establish a large data set of full-degree-of-freedom modal vibration shapes that includes the uncertainty of the structural physical parameters and contains the physical constraints of the structural dynamics;

[0009] Step 2: Build a generative adversarial network model and train the model to map the full-degree-of-freedom vibration shape φ′ considering the uncertainty of structural parameters into a low-dimensional vector z, obtain a generator G with the ability to represent the full-degree-of-freedom vibration shape in low dimensions, and realize the extraction of structural dynamics physical constraint information from a large data set of full-degree-of-freedom vibration shapes;

[0010] Step 3: Based on the trained structural vibration shape generator, establish the transfer learning criteria for similar structures, adopt appropriate transfer learning methods according to specific situations, and realize the rapid training of the structural vibration shape generator G;

[0011] Step 4. Based on the free vibration response of the structure and the obtained generative adversarial network model that contains the physical constraints of the structural vibration mode dynamics and has the ability to represent the full-degree-of-freedom vibration mode in a low-dimensional manner, the low-dimensional vector z is easily converged to the true solution during the optimization process, and the structural response of a small number of measurement points is used to realize the identification of the full-degree-of-freedom modal vibration mode.

[0012] Furthermore, the step 1 is specifically as follows:

[0013] Step 1.1, based on sensitivity analysis, determine the sensitive physical parameters that affect the change of structural vibration mode;

[0014] Step 1.2: For these physical parameters, a large number of Gaussian random samplings are performed with the nominal value of the physical model as the center and a certain variance is set. In addition, random sampling is also performed on other physical parameters;

[0015] Step 1.3: Based on the physical parameters obtained from each sampling, use finite element simulation to calculate multi-order full-degree-of-freedom modal vibration shapes, and establish a large data set of full-degree-of-freedom vibration shapes that includes the uncertainty of structural physical parameters and contains the physical constraints of structural dynamics.

[0016] Furthermore, the step 2 is specifically as follows:

[0017] Step 2.1: Construct a generative adversarial network model. The constructed generative adversarial network model consists of two parts: a generator G and a discriminator D.

[0018] Step 2.2: Based on the established model, the discriminator aims to determine whether the input data comes from the real data set, while the generator aims to generate realistic data to confuse the discriminator's judgment, and continuously improves the generator's generation ability through a dynamic game process;

[0019] Step 2.3, extract features from the full-degree-of-freedom vibration shapes generated by the generator and the “real” full-degree-of-freedom vibration shapes simulated by the finite element, and then calculate the FID distance metric in the feature space to quantify the data generation quality;

[0020] Step 2.4, establish a dimension selection method for the low-dimensional vector z, obtain a generator model G with the ability to represent the full-degree-of-freedom vibration mode in low dimensions, and realize the mapping of the large data set of high-dimensional full-degree-of-freedom vibration modes in finite element simulation to the low-dimensional vector z.

[0021] Furthermore, the functions of the generator G and the discriminator D are respectively: the input of the generator is a low-dimensional vector z randomly sampled from the implicit space, and this vector is projected through the generator network to "generate" full-degree-of-freedom vibration shape data φ=G(z); the input of the discriminator is the "real" full-degree-of-freedom vibration shape data φ′ calculated by the structural finite element simulation and the vibration shape data φ generated by the generator, and the discriminator network outputs the discriminant result as true or false.

[0022] Furthermore, the minimum dimension of the low-dimensional vector z corresponding to the stable increasing trend of the FID indicator value is selected as the appropriate dimension.

[0023] Furthermore, in step three, there are two major types of transfer learning methods for deep learning models: fine-tuning and feature extractor freezing; the fine-tuning method uses the trained model network parameters to re-fine-tune the target task to improve the transfer learning effect; the feature extractor freezing method fixes most of the hidden layer parameters in the network to the parameter values ​​of the corresponding positions in the trained model network, and retrains other layers to save network training time to the maximum extent; for the specific modal feature learning task of the target structure required, observe the changing trend of the FID indicator, test the two methods and make a selection.

[0024] Furthermore, the step 4 is specifically as follows:

[0025] Step 4.1: Use the observed structural vibration response data y′ to establish the optimization objective function L for modal identification z; Based on vibration theory, the mechanical relationship between structural vibration response data and structural vibration mode parameters is analyzed; According to the modal superposition principle, the structural vibration response is approximately expressed as the product sum of the corresponding modal coordinates of each order and the structural modal vibration mode; The relationship between the structural dynamic response and the structural vibration mode in the frequency domain is represented by a linear mapping;

[0026] Step 4.2: Substitute the low-dimensional vector z randomly sampled from the implicit space of the generator G as the initial value into the objective function L z ; According to the back propagation theory, the gradient caused by the objective function error is calculated, and the value of the low-dimensional vector z is updated by the gradient descent method, and it is iterated continuously until the iteration converges to obtain the optimal solution of the low-dimensional vector Finally, the projection through the vibration shape generator G is obtained As a result of the identification of the modal shapes of the structure with all degrees of freedom.

[0027] Furthermore, in step 4.2, the objective function is specifically:

[0028] L z =||A·F[G(z)]-y′||2+λ·L p (z)

[0029] Among them, F is the structural dynamics relationship, A is the linear observation matrix, and L p is the selected penalty term and λ is the corresponding hyperparameter.

[0030] The present invention proposes an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method for identifying the full-degree-of-freedom vibration mode of a structure based on a small amount of measurement point acceleration data and a generative adversarial network are implemented.

[0031] The present invention proposes a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the method for identifying the full-degree-of-freedom vibration mode of a structure based on a small amount of measurement point acceleration data and a generative adversarial network are implemented.

[0032] Beneficial effects of the present invention:

[0033] 1. Compared with the traditional method, the full-degree-of-freedom modal vibration shape identification method based on generative adversarial network described in the present invention uses the modal vibration shape generator as the physical constraint of modal recognition. It can identify the full-degree-of-freedom modal vibration shape of the structure based on only a small number of spatial sparse point sensors, which reduces the monitoring cost and improves the recognition precision. It is suitable for application in actual structural health monitoring systems.

[0034] 2. The present invention utilizes a generative adversarial network generator that has learned a large data set of diversified modal vibration shapes as a physical constraint, establishes a mapping relationship between low-dimensional vectors and sensor observation data, and realizes full-degree-of-freedom modal vibration shape recognition based on a small number of measurement points by optimizing the low-dimensional vectors, significantly reduces the dimensionality of the understanding space, alleviates the ill-posedness of the modal recognition inversion problem, and enhances the speed and precision of modal recognition.

[0035] 3. The method of the present invention is highly robust to noise, and the physical constraints contained in the trained generator ensure the physical authenticity of the modal recognition results, which can help eliminate false modes in the data, achieve better robustness than traditional modal recognition methods, and reduce the requirements of modal recognition methods for the quality of monitoring data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0037] Figure 1 It is a schematic diagram of the structural full-degree-of-freedom vibration mode identification method based on a small amount of measurement point acceleration data and a generative adversarial network according to the present invention.

[0038] Figure 2 This is a schematic diagram of the results of identifying the first eight vibration modes of a simply supported beam structure using the method of the present invention under complete observation.

[0039] Figure 3 This is a schematic diagram of an example of generated data for training a full-degree-of-freedom modal vibration shape generator for a truss structure according to the present invention.

[0040] Figure 4 It is a schematic diagram of the basic mode identification results of the truss structure corresponding to the method of the present invention under sparse observation. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] The purpose of the present invention is to solve the problem that the existing modal identification methods can only obtain the modal vibration shape values ​​of a few degrees of freedom at the measurement point position. This problem will lead to the inability to eliminate the generated false modes and reduce the use value of the modal identification method. Therefore, the present invention proposes a method based on generative adversarial network learning of a large data set of full-degree-of-freedom modal vibration shapes that contains structural physical parameter uncertainties and contains structural dynamics physical constraints, captures the constraints on the physical characteristics of the modal vibration shapes in the data set, trains a generator of full-degree-of-freedom modal vibration shapes, and uses the generator as a full-degree-of-freedom modal vibration shape identification method based on only a small number of measurement point observation data with physical constraints. By utilizing the characteristic that low-dimensional vectors are easy to converge to the true solution during the optimization process, this method not only enhances the speed, accuracy, and spatial resolution of modal identification, but also makes the features in the reconstructed results more physically realistic, automatically achieves robustness to false modes contained in data noise, and has a wider range of applicability in actual engineering. For example Figure 1 The process principle of training the generative adversarial network model and iteratively optimizing the identification of full-degree-of-freedom modal vibration shapes is shown. For spatially sparse point monitoring data in structural health monitoring, since adding sensors will significantly increase the system cost, a method for identifying high-resolution modal vibration shapes based on only a small number of measurement points is urgently needed. At the same time, the data variability of actual civil engineering structures is large, and it is necessary to be sensitive to false modes and resist noise interference, so it is suitable for the method of the present invention.

[0043] The present invention proposes a method for identifying structural full-degree-of-freedom vibration modes based on a small amount of measurement point acceleration data and a generative adversarial network, which specifically includes the following steps:

[0044] Step 1: Randomly sample the physical parameters of the structural finite element model, calculate and establish a large data set of full-degree-of-freedom modal vibration shapes that includes the uncertainty of the structural physical parameters and contains the physical constraints of the structural dynamics;

[0045] Step 2: Construct a generative adversarial network model and train the model to map the full-degree-of-freedom vibration shape φ′ considering the uncertainty of structural parameters into a low-dimensional vector z, obtain a generator G with the ability to represent the full-degree-of-freedom vibration shape in low dimensions, and realize the extraction of structural dynamics physical constraint information from a large data set of full-degree-of-freedom vibration shapes;

[0046] Step 3: The key to the vibration mode recognition of the method is the full-degree-of-freedom vibration mode characteristics of a certain structure learned by the generator G. Therefore, it is only applicable to the vibration mode recognition of the corresponding structure when training the generative adversarial network. In order to expand the applicability of the method to other structures, the knowledge that the vibration modes of the same type of structures often have similar characteristics is used. Based on the trained structural vibration mode generator, a transfer learning criterion for the same type of structure is established, and a suitable transfer learning method is used for specific situations to achieve rapid training of the structural vibration mode generator G;

[0047] Step 4. Based on the free vibration response of the structure and the obtained generative adversarial network model that contains the physical constraints of the structural vibration mode dynamics and has the ability to represent the full-degree-of-freedom vibration mode in a low-dimensional manner, the low-dimensional vector z is easily converged to the true solution during the optimization process, and the structural response of only a small number of measurement points is used to realize the identification of the full-degree-of-freedom modal vibration mode.

[0048] In order to make the generative adversarial network model have better generalization ability, the amount of simulation data set should be as large as possible, and the uncertainty of the parameter values ​​of the structural physical model should be fully considered. The specific step 1 is:

[0049] Step 1.1, based on sensitivity analysis, determine the sensitive physical parameters that affect the change of structural vibration mode;

[0050] Step 1.2: For these physical parameters, a large number of Gaussian random samplings are performed with the nominal value of the physical model as the center and a certain variance is set. In addition, random sampling is also performed on other physical parameters;

[0051] Step 1.3: Based on the physical parameters obtained from each sampling, use finite element simulation to calculate multi-order full-degree-of-freedom modal vibration shapes, and establish a large data set of full-degree-of-freedom vibration shapes that includes the uncertainty of structural physical parameters and contains the physical constraints of structural dynamics.

[0052] The step 2 is specifically as follows:

[0053] Step 2.1, construct a generative adversarial network model, which consists of two parts: the generator G and the discriminator D. The functions of these two parts are as follows: the input of the generator is a low-dimensional vector z randomly sampled from the implicit space, and this vector is projected through the generator network to "generate" full-degree-of-freedom vibration shape data φ = G(z); the input of the discriminator is the "real" full-degree-of-freedom vibration shape data φ′ calculated by the structural finite element simulation and the vibration shape data φ generated by the generator, and the discriminator network outputs the discriminant result as true or false;

[0054] Step 2.2: Based on the established model, the discriminator aims to determine whether the input data comes from the real data set as accurately as possible, while the generator aims to generate realistic data as much as possible to confuse the discriminator's judgment, and continuously improve the generator's generation ability through a dynamic game process;

[0055] Step 2.3, in order to comprehensively evaluate the physical authenticity and diversity of the generated data, extract features from the full-degree-of-freedom vibration shapes generated by the generator and the "real" full-degree-of-freedom vibration shapes simulated by finite element simulation, and then calculate the FID (Fréchet Inception Distance) distance metric in the feature space to quantify the data generation quality;

[0056] Step 2.4, establish a suitable method for selecting the dimension of the low-dimensional vector z, obtain a generator model G with the ability to represent the full-degree-of-freedom vibration mode in low dimensions, and realize the mapping of the high-dimensional full-degree-of-freedom vibration mode data set of the finite element simulation to the low-dimensional vector z. Among them, if the dimension of z is too low, some key features of the full-degree-of-freedom vibration mode of the structure may not be learned; if the dimension of z is too high, it may affect the subsequent optimization and recognition effect. In order to make the dimension of the low-dimensional vector z as low as possible and ensure the quality of the generated structure full-degree-of-freedom vibration mode data, it is necessary to test the dimensions of different low-dimensional vectors z. Continuously adjust the network architecture and training hyperparameters, and select the appropriate dimension based on the minimum dimension of the low-dimensional vector z corresponding to the stable trend of the FID indicator value increase.

[0057] The step three is specifically as follows:

[0058] There are two main types of transfer learning methods for deep learning models: fine-tuning and feature extractor freezing. The fine-tuning method uses the trained model network parameters to re-fine-tune the target task to improve the transfer learning effect. The feature extractor freezing method fixes most of the hidden layer parameters in the network to the parameter values ​​of the corresponding positions in the trained model network, and retrains other layers to save network training time to the maximum extent. For the modal feature learning task of the specific target structure required, observe the change trend of the FID indicator, and test the two methods for selection.

[0059] The step 4 is specifically as follows:

[0060] Step 4.1: Use the observed structural vibration response data y′ to establish the optimization objective function L for modal identification z . Based on vibration theory, the mechanical relationship between structural vibration response data and structural vibration parameters is analyzed. According to the modal superposition principle, under common service loads (environmental excitation), the structural vibration response can be approximately expressed as the product sum of the corresponding modal coordinates of each order and the structural modal vibration mode. This relationship means that the relationship between the structural dynamic response and the structural vibration mode in the frequency domain can be expressed by a simple linear mapping, so that the following objective function can be constructed:

[0061] L z =||A·F[G(z)]-y′||2+λ·L p (z)

[0062] Among them, F is the structural dynamics relationship, A is the linear observation matrix, and L p is the selected penalty term, and λ is the corresponding hyperparameter. The main function of this penalty term is to prevent the low-dimensional vector z from deviating too much from the prior distribution selected during training (for example, the standard Gaussian distribution corresponds to the two-norm penalty term) during the optimization process;

[0063] Step 4.2: Substitute the low-dimensional vector z randomly sampled from the implicit space of the generator G as the initial value into the objective function L z According to the back propagation theory, the gradient caused by the objective function error is calculated, and the value of the low-dimensional vector z is updated by the gradient descent method. Iterate continuously until the iteration converges to obtain the optimal solution of the low-dimensional vector Finally, the projection through the vibration shape generator G is obtained As a result of the identification of the modal shapes of the structure with all degrees of freedom.

[0064] Example 1

[0065] This implementation mode is to apply the present invention to a certain simply supported beam bridge, and to train a generative adversarial network model for the full-degree-of-freedom modal vibration mode identification method for the simply supported beam structure, to identify the full-degree-of-freedom modal vibration mode under complete observation, and to verify the optimization error of the method of the present invention (excluding the error caused by incomplete information). Figure 1 A schematic diagram of a structural full-degree-of-freedom vibration mode identification method based on a small amount of measurement point acceleration data and a generative adversarial network according to the present invention is given; Figure 2 The results of the identification of the first eight vibration modes of the simply supported beam structure using the method of the present invention under complete observation are given.

[0066] The step 1 is specifically as follows: the number of degrees of freedom of the full-degree-of-freedom vibration mode of the simply supported beam bridge finite element model used is 81, and the dimension of the implicit characteristic space vector z is 10.

[0067] The step 2 is specifically as follows: referring to the Deep Convolution Generative Adversarial Networks (DCGAN), constructing a generator G and a discriminator D for training.

[0068] The step three is specifically: using the Adam optimization algorithm to minimize the objective function L z , iterate 200 times and select the one with the smallest error as the identification result of the mode vibration shape.

[0069] The specific step 4 is as follows: given the complete observed response data, the identification results of the method of the present invention for the first 8 vibration modes of the structure are as follows: Figure 2 As shown in the figure. The results show that the vibration mode of the generator generated by substituting the 10-dimensional low-dimensional vector z into the generative adversarial network can correspond more accurately to the vibration mode of the complete model of the real system, which preliminarily verifies the good generation ability of the generative adversarial network for the vibration mode of the structural system. Therefore, the identification of the 81-dimensional full-degree-of-freedom vibration mode can be directly realized by optimizing the 10-dimensional low-dimensional vector z, which can significantly reduce the ill-posedness of vibration mode identification under incomplete structural response data.

[0070] Example 2

[0071] This implementation mode is to apply the present invention to a truss structure, utilize the full-degree-of-freedom modal vibration shape identification method based on the generative adversarial network of the present invention, train the generative adversarial network model for the truss structure, identify the full-degree-of-freedom modal vibration shape under sparse observation, and verify the actual effect of the method of the present invention. Figure 3 An example of generated data for training a full-degree-of-freedom modal vibration generator for a truss structure according to the present invention is given. Figure 4 The results of basic mode identification of truss structures corresponding to the method of the present invention under sparse observation are given.

[0072] The step 1 is specifically as follows: the number of degrees of freedom of the full degree of freedom vibration mode of the truss finite element model used is 328, and the dimension of the implicit characteristic space vector z is 50. The number of measurement points is 25% of the number of vertical degrees of freedom, and a measurement noise of 5% level is considered in the response data.

[0073] The step 2 is specifically as follows: referring to the DCGAN model, constructing a generator G and a discriminator D for training.

[0074] The step three is specifically: using the Adam optimization algorithm to minimize the objective function L z , iterate 200 times and select the one with the smallest error as the identification result of the mode vibration shape.

[0075] The specific step 4 is as follows: the learning generation example of the structural vibration mode of the method of the present invention is as follows Figure 3 As shown in Figure 1, it shows that the generator successfully captures the physical characteristics of the vibration mode. Given the sparsely observed response data, the recognition results of the basic vibration mode of the structure by the method of the present invention are as follows: Figure 4 As shown, it is demonstrated that the full-degree-of-freedom modal vibration type identification method of the present invention is practical and effective.

[0076] The present invention proposes an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method for identifying the full-degree-of-freedom vibration mode of a structure based on a small amount of measurement point acceleration data and a generative adversarial network are implemented.

[0077] The present invention proposes a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the method for identifying the full-degree-of-freedom vibration mode of a structure based on a small amount of measurement point acceleration data and a generative adversarial network are implemented.

[0078] The memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct RAM bus RAM (DR RAM). It should be noted that the memory of the method described in the present invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0079] In the above embodiments, it 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 instructions. When the computer instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (digital subscriber line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a high-density digital video disc (DVD)), or a semiconductor medium (eg, a solid state disc (SSD)).

[0080] In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in a processor or an instruction in the form of software. The steps of the method disclosed in conjunction with the embodiment of the present application can be directly embodied as a hardware processor for execution, or a combination of hardware and software modules in a processor for execution. The software module can be located in a storage medium mature in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in a memory, and the processor reads the information in the memory and completes the steps of the above method in conjunction with its hardware. To avoid repetition, it is not described in detail here.

[0081] It should be noted that the processor in the embodiment of the present application can be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method embodiment can be completed by an integrated logic circuit of hardware in the processor or an instruction in the form of software. The above processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to perform, or the hardware and software modules in the decoding processor can be combined and performed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in a memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0082] The above is a detailed introduction to the structural full-degree-of-freedom vibration mode identification method based on a small amount of measurement point acceleration data and a generative adversarial network proposed in the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A structural full-degree-of-freedom vibration mode identification method based on a small amount of measurement point acceleration data and a generative adversarial network, characterized in that: The method comprises the following steps: Step 1: Randomly sample the physical parameters of the structural finite element model, calculate and establish a large data set of full-degree-of-freedom modal vibration shapes that includes the uncertainty of the structural physical parameters and contains the physical constraints of the structural dynamics; Step 2: Construct a generative adversarial network model and train the model to map the full-degree-of-freedom vibration shape φ' considering the uncertainty of structural parameters into a low-dimensional vector z, obtain a generator G with the ability to represent the full-degree-of-freedom vibration shape in low dimensions, and realize the extraction of structural dynamics physical constraint information from a large data set of full-degree-of-freedom vibration shapes; Step 3: Based on the trained structural vibration shape generator, establish the transfer learning criteria for similar structures, adopt appropriate transfer learning methods according to specific situations, and realize the rapid training of the structural vibration shape generator G; Step 4: Based on the free vibration response of the structure and the obtained generative adversarial network model that contains the physical constraints of the structural vibration mode dynamics and has the ability to represent the full-degree-of-freedom vibration mode in low dimension, the low-dimensional vector z is easily converged to the true solution during the optimization process, and the structural response of a small number of measurement points is used to realize the identification of the full-degree-of-freedom modal vibration mode; The step 1 is specifically as follows: Step 1.1, based on sensitivity analysis, determine the sensitive physical parameters that affect the change of structural vibration mode; Step 1.2: For these physical parameters, a large number of Gaussian random samplings are performed with the nominal value of the physical model as the center and a certain variance is set. In addition, random sampling is also performed on other physical parameters; Step 1.3: Based on the physical parameters obtained from each sampling, use finite element simulation to calculate multi-order full-degree-of-freedom modal vibration shapes, and establish a large data set of full-degree-of-freedom vibration shapes that includes the uncertainty of structural physical parameters and contains the physical constraints of structural dynamics.

2. The method according to claim 1, characterized in that The step 2 is specifically as follows: Step 2.1: Construct a generative adversarial network model. The constructed generative adversarial network model consists of two parts: a generator G and a discriminator D. Step 2.2: Based on the established model, the discriminator aims to determine whether the input data comes from the real data set, while the generator aims to generate realistic data to confuse the discriminator's judgment, and continuously improves the generator's generation ability through a dynamic game process; Step 2.3, extract features from the full-degree-of-freedom vibration shapes generated by the generator and the "real" full-degree-of-freedom vibration shapes simulated by finite element simulation, and then calculate the FID distance metric in the feature space to quantify the data generation quality; Step 2.4, establish a dimension selection method for the low-dimensional vector z, obtain a generator model G with the ability to represent the full-degree-of-freedom vibration mode in low dimensions, and realize the mapping of the large data set of high-dimensional full-degree-of-freedom vibration modes in finite element simulation to the low-dimensional vector z.

3. The method according to claim 2, characterized in that The functions of the generator G and the discriminator D are as follows: the input of the generator is a low-dimensional vector z randomly sampled from the implicit space, and this vector is projected through the generator network to "generate" full-degree-of-freedom vibration shape data φ=G(z); the input of the discriminator is the "real" full-degree-of-freedom vibration shape data φ' calculated by the structural finite element simulation and the vibration shape data φ generated by the generator, and the discriminator network outputs the judgment result as true or false.

4. The method according to claim 2, characterized in that: The minimum dimension of the low-dimensional vector z corresponding to the stable increasing trend of the FID indicator value is selected as the appropriate dimension.

5. The method according to claim 1, characterized in that In step 3, there are two main types of transfer learning methods for deep learning models: fine-tuning and feature extractor freezing. The fine-tuning method uses the trained model network parameters to re-fine-tune the target task to improve the transfer learning effect. The feature extractor freezing rule is to fix most of the hidden layer parameters in the network to the parameter values ​​of the corresponding positions in the trained model network, and retrain other layers to save network training time to the maximum extent; for the modal feature learning task of the specific target structure, observe the changing trend of the FID indicator, and test two types of methods for selection.

6. The method according to claim 1, characterized in that The step 4 is specifically as follows: Step 4.1: Use the observed structural vibration response data y' to establish the optimization objective function L for modal identification z ; Based on vibration theory, the mechanical relationship between structural vibration response data and structural vibration mode parameters is analyzed; According to the modal superposition principle, the structural vibration response is approximately expressed as the product sum of the corresponding modal coordinates of each order and the structural modal vibration mode; The relationship between the structural dynamic response and the structural vibration mode in the frequency domain is represented by a linear mapping; Step 4.2: Substitute the low-dimensional vector z randomly sampled from the implicit space of the generator G as the initial value into the objective function L z ; According to the back propagation theory, the gradient caused by the objective function error is calculated, and the value of the low-dimensional vector z is updated by the gradient descent method, and it is iterated continuously until the iteration converges to obtain the optimal solution of the low-dimensional vector Finally, the projection through the vibration shape generator G is obtained As a result of the identification of the modal shapes of the structure with all degrees of freedom.

7. The method according to claim 6, characterized in that In step 4.2, the objective function is specifically: L z =‖A·F[G(Z)]-y′||2+λ·L p (z) Among them, F is the structural dynamics relationship, A is the linear observation matrix, and L p is the selected penalty term and λ is the corresponding hyperparameter.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

9. A computer-readable storage medium for storing computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images