A structural damage identification method based on physical information branch convolutional neural network

Abstract

The application provides a structural damage identification method based on a physical information branch convolutional neural network, a physical information branch convolutional neural network with a branch architecture is established, damage features in a characteristic frequency vector and a mode shape matrix are separately extracted, two kinds of damage features are fused, and finally damage identification results are mapped, and when the network is trained, a loss function containing sensitivity analysis of modal parameters and spatial sparsity of damage is used, physical information is combined with the neural network, so that the training of the network not only depends on a limited data set, but also is assisted by the physical information, the utilization efficiency of the network for the limited data set is improved, and the damage identification precision of the network under the limited data set is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of structural damage identification, in particular to a structural damage identification method based on a physical information branch convolutional neural network. BACKGROUND

[0002] Infrastructure inevitably suffers damage during long-term service due to the continuous bearing of various loads, which in turn affects its service life. Structural health monitoring, as an effective means of infrastructure maintenance, its core lies in structural damage identification, that is, the positioning of damage location and the quantification of damage degree. The vibration-based damage identification method has become a research hotspot due to its wide applicability, and its theoretical basis lies in that structural damage will cause changes in vibration response data. By comparing the baseline data of the undamaged state with the measured vibration response data and applying a specific damage identification algorithm, the damage identification result can be obtained.

[0003] In the technical application level, the deep learning method has been widely concerned due to its strong data feature extraction capability. Among them, the damage identification method based on ResNet architecture is representative. This research introduces a deep residual network to effectively overcome the gradient vanishing problem in the training of traditional neural networks, and uses a deeper network structure to extract more accurate damage features, which significantly improves the recognition accuracy.

[0004] It is worth noting that the successful application of neural networks is highly dependent on big data support. However, in actual engineering, there are two dilemmas in obtaining sufficient training data: first, collecting measured data in different scenarios is cost-limited, and researchers cannot intentionally damage the structure to collect data in different damage scenarios; second, although numerical simulation can be used as an alternative data source, for large infrastructure structures with complex modeling, not only is the numerical modeling difficult, but the generation of massive data also poses a serious challenge to computer storage and computing power. Therefore, how to improve the efficiency and accuracy of the algorithm under the condition of limited data has become a key issue that needs to be broken through in the field of neural network-based damage identification. SUMMARY

[0005] To solve the problems existing in the prior art, the present application provides a structural damage identification method based on a physical information branch convolutional neural network,

[0006] The technical scheme of the present application is as follows:

[0007] A structural damage identification method based on a physical information branch convolutional neural network, comprising the following steps:

[0008] Step 1: Establish a finite element model of the structure to be evaluated; generate modal parameters of different damage states using the finite element model; generate training samples for training the physical information branch convolutional neural network according to the modal parameters;

[0009] Step 2: building a physical information branch convolutional neural network; the physical information branch convolutional neural network adopts a branch structure, extracts damage features in the modal parameters respectively, and maps to obtain damage identification results after fusing various damage features;

[0010] Step 3: training the physical information branch convolutional neural network by using the training samples obtained in step 1, wherein the loss function used in the training contains two kinds of physical information, which are sensitivity analysis of modal parameters and spatial sparsity of damage respectively;

[0011] Step 4: obtaining actual modal parameters of the structure to be evaluated, inputting the trained physical information branch convolutional neural network, and obtaining structure damage identification results.

[0012] Further, the modal parameters include characteristic frequency and mode shape.

[0013] Further, the loss function used in the training is:

[0014] Loss = Loss MSE + Loss Phy

[0015] In the formula, Loss MSE represents mean square error loss, Loss Phy represents physical information loss;

[0016] Wherein the physical information loss Loss Phy is:

[0017] Loss Phy = β Mod Loss Mod + β Spa Loss Spa

[0018] In the formula, Loss Mod represents modal parameter loss, Loss Spa represents sparsity loss, β Mod and β Spa represent modal hyperparameters and sparsity hyperparameters.

[0019] Further, the modal parameter loss Loss Mod is:

[0020] Loss Mod = Loss FErr + Loss SErr

[0021] In the formula, Loss FErr represents frequency error loss, LossSErr denotes the mode error loss;

[0022] the frequency error loss Loss FErr is:

[0023]

[0024] where N b denotes the number of samples in the same batch during training, N order denotes the order of the eigenfrequency vector, S λ denotes the sensitivity matrix of the eigenfrequency vector with respect to the damage parameter vector, denotes the predicted value of the damage parameter vector of the t-th sample in the same batch, λ Dt denotes the eigenfrequency vector corresponding to the t-th sample in the same batch of input samples, λ U denotes the eigenfrequency vector corresponding to the structure to be evaluated in the undamaged state;

[0025] the mode error loss Loss SErr is:

[0026]

[0027] where N dof denotes the number of degrees of freedom set, S φ denotes the sensitivity matrix of the mode shape matrix with respect to the damage parameter vector, φ Dt denotes the mode shape matrix corresponding to the t-th sample in the same batch of input samples, φ U denotes the mode shape matrix corresponding to the structure to be evaluated in the undamaged state.

[0028] Further, the sparsity loss Loss Spa is:

[0029]

[0030] where N b denotes the number of samples in the same batch during training, N el denotes the number of elements of the finite element model, denotes the k-th element of the predicted value of the damage parameter vector of the t-th sample in the same batch of input samples, and σ is a hyperparameter in the fractional function F(·).

[0031] Further, the sensitivity matrix S λ of the eigenfrequency vector with respect to the damage parameter vector and the sensitivity matrix S φ of the mode shape matrix with respect to the damage parameter vector are respectively:

[0032]

[0033] wherein is the i-th eigenfrequency in the eigenfrequency vector in the undamaged state, is the i-th column vector in the mode shape matrix in the undamaged state, and j denotes the damage parameter of the j-th element, i.e., the j-th element of the damage parameter vector, j denotes the stiffness matrix of the j-th element, denotes the transpose of.

[0034] Further, in step 1, the process of generating modal parameters of different damage states by using the finite element model is: by setting the stiffness reduction values of different elements in the finite element model as damage parameters, simulating different degrees of damage of different elements, and calculating the modal parameters under different damage states by using the finite element model.

[0035] Further, in step 1, Gaussian random noise is added to the eigenfrequency vector and the mode shape matrix calculated by using the finite element model to simulate measurement noise, and Gaussian random noise is added to the damage parameter vector composed of the damage parameters of different elements in the finite element model to simulate finite element modeling errors.

[0036] Further, in step 1, the eigenfrequency vector and the mode shape matrix with added Gaussian random noise are subjected to maximum-minimum normalization preprocessing to obtain the eigenfrequency vector and the mode shape matrix used for training the physical information branch convolutional neural network.

[0037] Further, in step 2, in the physical information branch convolutional neural network, for the eigenfrequency vector, one-dimensional convolution operation is used to extract damage features therein, and for the mode shape matrix, two-dimensional convolution operation with residual connection is used to extract damage features therein; then the damage features extracted from the two kinds of data are spliced together in the channel dimension, and one-dimensional convolution is used for feature fusion; finally, the fused damage features are mapped to damage identification results by using a fully connected layer.

[0038] Beneficial effects

[0039] The beneficial effects of the present application are: the structural damage identification method based on the physical information branch convolutional neural network provided by the present application improves the damage identification accuracy under limited data set, solves the problem of low damage identification accuracy under limited data set, such as Figure 3As shown; wherein the branch architecture extracts damage features from frequency and mode shape respectively, improving the efficiency of extracting damage features from a limited data set;The modal parameter sensitivity and the spatial sparsity of the damage are added to the loss function as physical information, so that the training of the network not only depends on the limited data set, but also gets the help of the physical information, improves the utilization efficiency of the network to the limited data set, and improves the damage identification precision of the network under the limited data set.

[0040] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and / or realization of the inherent aspects of the application. BRIEF DESCRIPTION OF DRAWINGS

[0041] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which:

[0042] Figure 1 is a network architecture schematic diagram of a physical information branch convolutional neural network of the present application;

[0043] Figure 2 is a finite element model schematic diagram shown in embodiment 1 of the present application;

[0044] Figure 3 is a curve graph of the change of the validation mean square error loss of ResNet and the network (PIBCNN) provided by the present application under different data amounts shown in embodiment 1 of the present application;

[0045] Figure 4 is a curve graph of the change of the training loss and the validation loss of ResNet and the network (PIBCNN) provided by the present application shown in embodiment 1 of the present application;

[0046] Figure 5 is a damage identification result comparison graph of ResNet and the network (PIBCNN) provided by the present application shown in embodiment 1 of the present application;

[0047] Figure 6 is a damage identification actual structure physical map shown in embodiment 2 of the present application;

[0048] Figure 7 is a curve graph of the change of the training loss and the validation loss of ResNet and the network (PIBCNN) provided by the present application shown in embodiment 2 of the present application;

[0049] Figure 8 is a damage identification result comparison graph of ResNet and the network (PIBCNN) provided by the present application shown in embodiment 2 of the present application. DETAILED DESCRIPTION

[0050] The embodiments of the present application are described in detail below, which are exemplary and intended to explain the present application, and cannot be understood as a limitation of the present application.

[0051] Embodiment 1:

[0052] The structural damage identification method based on the physical information branch convolutional neural network proposed in this embodiment includes the following steps:

[0053] Step 1: Establish a finite element model of the structure to be evaluated; generate modal parameters of different damage states using the finite element model; generate training samples for training the physical information branch convolutional neural network according to the modal parameters.

[0054] In this embodiment, MATLAB programming software is used for finite element modeling, and the model schematic diagram is as shown in Figure 2 The model has a total of 54 units. By setting the stiffness reduction values of different units in the finite element model, different degrees of damage of different units are simulated, and the modal parameters under different damage states are calculated using the finite element model. In this embodiment, the modal parameters include characteristic frequencies and mode shapes.

[0055] To simulate the actual measurement situation, in this embodiment, the measured modal order is set to the first 8 orders, and the measured degrees of freedom are the transverse and longitudinal degrees of freedom of the 5th, 7th, 9th, 11th, 13th, and 15th nodes, a total of 12 degrees of freedom. Therefore, the simulated measured characteristic frequency vector λ in this embodiment is an 8x1 vector, and the mode shape matrix φ is a 12x8 matrix.

[0056] In order to simulate different damage conditions, this embodiment sets single damage conditions and multiple damage conditions. In the single damage condition, any unit is set with damage alone, the damage degree (i.e. the stiffness reduction value) ranges from 1% to 20% with a step of 1%, and a total of 54*20=1080 samples are generated. In the multiple damage condition, any two units are set with damage simultaneously, the damage degree ranges from 2% to 20% with a step of 2%, and a total of samples are generated. In summary, there are a total of 144180 samples for network training, verification and testing. Each sample is composed of a corresponding characteristic frequency vector λ and a mode shape matrix φ, and the label is a damage parameter vector θ. The damage parameter vector θ is a 54x1 vector, and each element value is the stiffness reduction value of the corresponding unit, taking a value between 0 and 1.

[0057] In order to more truly reflect the actual situation, on the basis of the above calculation results, 1% and 5% Gaussian random noise is added to the characteristic frequency vector and the mode shape matrix respectively to simulate measurement noise, and 1% Gaussian random noise is added to the damage parameter vector to simulate finite element modeling error. The specific noise adding method is as follows:

[0058] λ Noise = (1 + 0.01 μ) λ

[0059] φ Noise = (1 + 0.05 μ) φ

[0060] θ Noise = θ + 0.01 μ

[0061] where λ Noise , φ Noise , θ Noise are the feature frequency vector, mode shape matrix and damage parameter vector after adding noise respectively, and μ represents a Gaussian random number with mean 0 and variance 1.

[0062] The generated feature frequency vector λ Noise and mode shape matrix φ Noise are respectively preprocessed before inputting into the neural network to eliminate the dimension and make the training more stable. The maximum and minimum normalization preprocessing method is adopted in this embodiment to scale the data to [0, 1], and the specific operation is as shown in the following formula:

[0063]

[0064] where λ Noise is the i-th order feature frequency in λ Noise , max(λ Noise ) and min(λ Noise ) represent the maximum value and the minimum value of the elements in λ norm , respectively, and λ Noise represents the i-th order feature frequency in the normalized feature frequency vector λ Noise ; φ Noise is the i-th row and p-th column element in φ Noise , max(φ norm ) and min(φ norm ) represent the maximum value and the minimum value of the elements in φ norm , respectively, and φ MSE represents the i-th row and p-th column element in the normalized mode shape matrix φ Phy .

[0065] The above samples are randomly shuffled and divided into a training set and a validation set in a ratio of 7:3.

[0066] Step 2: Build a physical information branch convolutional neural network (PIBCNN):

[0067] As shown in the following figure:As shown, the physical information branch convolutional neural network adopts a branch architecture to extract damage features in frequency and mode shape data respectively, and finally maps to damage identification results after fusing the two kinds of damage features.

[0068] Specifically, for the characteristic frequency vector λ norm , one-dimensional convolution operation is used to extract damage features therein, and for the mode shape matrix φ norm , two-dimensional convolution operation with residual connection is used to extract damage features therein; then the damage features extracted from the two kinds of data are spliced together in the channel dimension, and one-dimensional convolution is used for feature fusion; finally, the fused damage features are mapped to the damage identification results by using the full connection layer.

[0069] In this embodiment, the one-dimensional convolution branch and the two-dimensional convolution branch each include five stages, the convolution kernel size of each stage is 2x1 and 2x2 respectively, the channel number is 16k, 32k, 64k, 128k and 128k respectively, and each stage has residual connection. k represents the width factor of the network, which is a hyperparameter.

[0070] Step 3: Train the physical information branch convolutional neural network:

[0071] The training sample data obtained in step 1 is input into the physical information branch convolutional neural network established in step 2 in batches for training and verification.

[0072] The loss function used in training contains two kinds of physical information, which are the sensitivity analysis of modal parameters and the spatial sparsity of damage respectively. The spatial sparsity of damage is described by using a fractional function, and the physical information is combined with the neural network, so that a structural damage identification model with high recognition accuracy can be obtained under the condition of limited training samples.

[0073] The specific loss function is:

[0074] Loss=Loss MSE +Loss Phy

[0075] In the formula, Loss MSE represents the mean square error loss, and Loss Phy represents the physical information loss.

[0076] For the mean square error loss Loss MSE , it is defined as:

[0077]

[0078] In the formula, N el represents the number of elements of the finite element model, and N bdenotes the number of samples in the same batch of input samples during training, denotes the predicted value of the damage parameter vector of the t-th sample in the same batch, θ t denotes the true damage state label of the corresponding sample.

[0079] for the physical information loss Loss Phy As mentioned before, two kinds of physical information are contained, which are the sensitivity analysis of modal parameters and the spatial sparsity of damage, respectively, and are specifically defined as:

[0080] Loss Phy = β Mod Loss Mod + β Spa Loss Spa

[0081] In the formula, Loss Mod denotes the modal parameter loss, Loss Spa denotes the sparsity loss, β Mod and β Spa denote the modal hyperparameter and the sparsity hyperparameter.

[0082] The modal parameter loss Loss Mod in it is defined as:

[0083] Loss Mod = Loss FErr + Loss SErr

[0084] In the formula, Loss FErr denotes the frequency error loss, and Loss SErr denotes the mode error loss.

[0085] For the frequency error loss Loss FErr , it is defined as:

[0086]

[0087] In the formula, N order denotes the order of the characteristic frequency vector, which is 8 in the embodiment; S λ denotes the sensitivity matrix of the characteristic frequency vector with respect to the damage parameter vector, λ Dt denotes the characteristic frequency vector corresponding to the t-th sample in the same batch of input samples, λ U denotes the characteristic frequency vector corresponding to the structure to be evaluated in the undamaged state.

[0088] For the mode error loss Loss SErr , it is defined as:

[0089]

[0090] wherein N dof denotes the number of degrees of freedom set, which is 12 in the embodiment; S φ denotes the sensitivity matrix of the mode shape matrix with respect to the damage parameter vector, φ Dt denotes the mode shape matrix corresponding to the t-th sample in the same batch of input samples, φ U denotes the mode shape matrix corresponding to the structure to be evaluated in the undamaged state.

[0091] wherein the sparsity loss Loss Spa is defined as:

[0092]

[0093] wherein denotes the k-th element in the damage parameter vector prediction value of the t-th sample in the same batch of input samples, and σ is a hyperparameter in the fractional function F(·).

[0094] The fractional function F(·) is defined as:

[0095]

[0096] The sensitivity matrix S of the eigenfrequency vector with respect to the damage parameter vector λ and the sensitivity matrix S of the mode shape matrix with respect to the damage parameter vector φ is defined as:

[0097]

[0098] wherein is the i-th order eigenfrequency in the eigenfrequency vector in the undamaged state, is the i-th column vector in the mode shape matrix in the undamaged state, θ j denotes the damage parameter of the j-th unit, K j denotes the stiffness matrix of the j-th unit, denotes the transpose of.

[0099] In the verification, the mean square error loss function is used for verification, the Adam optimizer is used as the network training optimizer, and the stepwise decay strategy is used as the learning rate decay strategy.

[0100] In order to reflect the improvement of the damage identification precision of the physical information branch convolutional neural network obtained by the application, a ResNet network is established for damage identification in the embodiment, and the loss value curves of the training and verification of the two networks with the number of training times are as shown in Figure 4 It can be seen from the figure that the loss of the network provided by the application in the verification set is lower, that is, the damage identification precision is higher.

[0101] Step 4: Structural damage identification using the trained network:

[0102] For the actual collected structural acceleration data, the modal analysis method is used to convert the acceleration data into a characteristic frequency vector and a mode shape matrix. The characteristic frequency vector and the mode shape matrix are input into the trained physical information branch convolutional neural network to obtain the structural damage identification results, including the damage location and severity. The visualization results are shown in Figure 5 .

[0103] Similarly, for comparison, the characteristic frequency vector and the mode shape matrix of the actual structure are also input into the ResNet for damage identification. The visualization results are also shown in Figure 5 . The results show that the ResNet misjudges the unit 38, while the network provided by the present application accurately identifies the damage at units 33 and 46, and completes more accurate damage identification.

[0104] Example 2:

[0105] This example performs structural damage identification on the actual cantilever beam structure as shown in Figure 6 .

[0106] Step 1: Establish a finite element model of the cantilever beam; generate modal parameters of different damage states using the finite element model; generate training samples for training the physical information branch convolutional neural network according to the modal parameters.

[0107] In this example, except for the acceleration sensor at the fixed end, there are 5 effective acceleration sensors equally spaced on the actual structure cantilever beam to measure the horizontal direction acceleration of the corresponding measuring points, corresponding to the 3rd, 5th, 7th, 9th and 11th nodes of the finite element numerical model. Since the first order frequency of the cantilever beam is difficult to measure, the 2nd to 7th order modes of the cantilever beam measured in the experiment are taken as the modal parameters.

[0108] The training data includes single damage and multiple damage of various damage conditions. In generating the single damage data set, any unit is set with damage alone, the damage degree is equally distributed between 1% and 50%, the distance is 1%, and a total of 500 samples are generated. In generating the multiple damage data set, any two units are set with damage, the damage degree is equally distributed between 2% and 50%, the distance is 2%, and a total of 28125 samples are generated. In addition, any three units are set with damage, the damage degree is equally distributed between 10% and 50%, the distance is 10%, and a total of 15000 samples are generated. In summary, there are a total of 43625 samples of different damage conditions in the data set for training and verifying the neural network. In line with the numerical experiment of embodiment 1, the above data set is randomly shuffled and divided into a training set and a verification set in a ratio of 7:3. And to simulate various uncertainties in the laboratory, 2% Gaussian white noise is added to the frequency in the above data set, and 10% Gaussian white noise is added to the mode shape, and the noise adding mode is consistent with the preprocessing mode of embodiment 1.

[0109] Step 2: build and train the physical information branch convolutional neural network (PIBCNN):

[0110] The building method is consistent with embodiment 1. The training samples are input into the built physical information branch convolutional neural network in batches, and the network is trained and verified in the same way as in embodiment 1. In order to compare with other methods and reflect the improvement of the network provided by the present application in damage identification accuracy, a ResNet network is also established for damage identification. The loss value curves of the two networks with the number of training times are shown in Figure 7 From the figure, it can be seen that the loss of the verification set of the network provided by the present application (PIBCNN) is lower, that is, the damage identification accuracy is higher.

[0111] Step 3: use the trained network for structural damage identification:

[0112] In order to collect the measured modal parameters under different damage conditions, including the characteristic frequency vector and the mode shape matrix, a series of vibration tests are carried out by using the hammering method in this embodiment. First, one end of the beam is fixed on a support, and 6 acceleration sensors are evenly arranged on the beam to measure the horizontal acceleration response. Second, the cantilever beam is struck with a force hammer to make it vibrate. The acceleration signal generated by the vibration is transmitted to the constant source flow adapter and recorded by the dynamic and static strain acquisition and analysis system. Finally, the signal acquisition system and modal analysis system in the desktop computer are used to complete the acquisition and modal analysis of the acceleration signal, and finally the modal parameters are obtained.

[0113] The characteristic frequency vector and mode shape matrix of the real structure collected are input into the PIBCNN and ResNet built and trained, and the visual results of the damage identification calculated are shown in Figure 8The results show that ResNet seriously underestimates the damages at unit 1 and unit 9, while the PIBCNN provided by the application more accurately identifies all the damages, and completes more accurate damage identification.

[0114] Although the embodiments of the present application have been shown and described above, it should be understood by those skilled in the art that the above-mentioned embodiments are exemplary and cannot be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the present application without departing from the principles and purposes of the present application.

Claims

1. A method for structural damage identification based on physical information branch convolutional neural network, characterized in that: The method comprises the following steps: Step 1: establishing a finite element model of a structure to be evaluated; generating modal parameters of different damage states by using the finite element model; and generating training samples for training a physical information branch convolutional neural network according to the modal parameters; The process of generating modal parameters of different damage states by using the finite element model is as follows: setting stiffness reduction values of different units in the finite element model as damage parameters to simulate different degrees of damage of the different units, and calculating modal parameters under different damage states by using the finite element model; the modal parameters include characteristic frequencies and vibration modes; Gaussian random noise is added to the characteristic frequency vector and the vibration mode matrix calculated by using the finite element model to simulate measurement noise, and Gaussian random noise is added to a damage parameter vector composed of the damage parameters of the different units in the finite element model to simulate finite element modeling errors; The characteristic frequency vector and the vibration mode matrix with the added Gaussian random noise are subjected to maximum-minimum normalization pretreatment to obtain the characteristic frequency vector and the vibration mode matrix for training the physical information branch convolutional neural network; Step 2: building a physical information branch convolutional neural network; the physical information branch convolutional neural network adopts a branch structure, extracts damage features in the modal parameters respectively, and maps to obtain damage identification results after fusing various damage features; Step 3: training the physical information branch convolutional neural network by using the training samples obtained in step 1, wherein two kinds of physical information are contained in a loss function used in the training, and the two kinds of physical information are respectively a sensitivity analysis of the modal parameters and a spatial sparsity of damage; Step 4: obtaining actual modal parameters of the structure to be evaluated, inputting the actual modal parameters into the trained physical information branch convolutional neural network, and obtaining a structure damage identification result. 2.The structural damage identification method based on physical information branch convolutional neural network according to claim 1, characterized in that: The loss function used in the training is as follows: wherein denotes the mean square error loss, denotes the physical information loss; where the physical information loss is: wherein denotes the modal parameter loss, denotes the sparsity loss, and denotes the modal hyperparameters and the sparsity hyperparameters. 3.The structural damage identification method based on physical information branch convolutional neural network according to claim 2, characterized in that: The modal parameter loss is: wherein represents the frequency error loss, represents the mode shape error loss; The frequency error loss is: In the formula represents the number of samples in the same batch of input samples during training, represents the order of the characteristic frequency vector, represents the sensitivity matrix of the characteristic frequency vector with respect to the damage parameter vector, represents the predicted value of the damage parameter vector of the sample in the same batch, represents the characteristic frequency vector corresponding to the sample in the same batch of input samples, represents the characteristic frequency vector corresponding to the structure to be evaluated in the undamaged state. The mode error loss is: In the formula represents the number of degrees of freedom set, represents a sensitivity matrix of the mode shape matrix with respect to the damage parameter vector, represents a mode shape matrix corresponding to the i-th sample in the same batch of input samples, represents a mode shape matrix corresponding to the i-th sample in the same batch of input samples, represents a mode shape matrix corresponding to the i-th sample in the same batch of input samples, 4.The structural damage identification method based on physical information branch convolutional neural network according to claim 2, characterized in that: The sparsity loss Is: In the formula represents the number of samples in the same batch of input samples during training, represents the number of units of the finite element model, represents the first element in the damage parameter vector prediction value of the sample in the same batch of input samples, represents the first element in the damage parameter vector prediction value of the is a hyperparameter in the fractional function .

5. The structural damage identification method based on physical information branch convolutional neural network according to claim 3, characterized in that: Sensitivity matrix of the feature frequency vector with respect to the damage parameter vector Sensitivity matrix of the mode shape matrix with respect to the damage parameter vector respectively: In the formula The first element in the feature frequency vector under the undamaged state First-order characteristic frequencies, The first modal matrix in the undamaged state Column vector, Indicates the first Damage parameters of each unit, Indicates the first Stiffness matrix of each element express The transpose of . 6.The structural damage identification method based on physical information branch convolutional neural network according to claim 1, characterized in that: In step 2, for the characteristic frequency vector, one-dimensional convolution operation is used to extract damage features therein, and for the vibration mode matrix, two-dimensional convolution operation with residual connection is used to extract damage features therein; then the damage features extracted from the two kinds of data are spliced together in the channel dimension, and one-dimensional convolution is used for feature fusion; finally, a fully connected layer is used to map the fused damage features to damage identification results.

Images