Structural damage identification method based on physical information branch convolutional neural network

Abstract

The invention provides a structural damage identification method based on a physical information branch convolutional neural network, and the method comprises the steps: building a physical information branch convolutional neural network with a branch architecture, respectively and independently extracting damage features in a feature frequency vector and a vibration mode matrix, fusing the two damage features, and finally carrying out the mapping, thereby obtaining a damage identification result. And during network training, a loss function containing two kinds of physical information including sensitivity analysis of modal parameters and spatial sparsity of damage is adopted, and the 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, and the training accuracy of the network is improved. The utilization efficiency of the network on 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 invention relates to the technical field of structural damage identification, and in particular to a structural damage identification method based on a physical information branch convolutional neural network. Background Art

[0002] During long-term service, infrastructure will inevitably suffer damage due to the continuous loads it bears, which in turn affects its service life. Structural health monitoring, as an effective means of infrastructure maintenance, focuses on structural damage identification, namely, locating the damage location and quantifying the damage extent. Vibration-based damage identification methods have become a research hotspot due to their wide applicability. Their theoretical basis is that structural damage will cause changes in vibration response data. By comparing non-destructive baseline data with measured vibration response data and applying specific damage identification algorithms, damage identification results can be obtained.

[0003] In terms of technological application, deep learning methods have attracted widespread attention due to their powerful data feature extraction capabilities. Among them, the damage identification method based on the ResNet architecture is a representative example. By introducing a deep residual network, this research effectively overcomes the vanishing gradient problem in traditional neural network training. By utilizing a deeper network structure to extract more precise damage features, it significantly improves identification accuracy.

[0004] It is worth noting that the successful application of neural networks is highly dependent on big data support. However, in actual engineering, obtaining sufficient training data faces a dual dilemma: First, there are cost constraints in collecting measured data from different scenarios, and researchers cannot intentionally damage structures indefinitely to collect data under different damage scenarios. Second, although numerical simulation can be used as an alternative data source, for large-scale infrastructure structures with complex modeling, not only is numerical modeling difficult, but the generation of massive amounts of data also poses severe challenges to computer storage and computing power. Therefore, how to improve algorithm utilization efficiency and recognition accuracy under limited data conditions has become a key issue that urgently needs to be broken through in the field of damage identification based on neural networks. Summary of the Invention

[0005] In order to solve the problems existing in the prior art, the present invention proposes a structural damage identification method based on physical information branch convolutional neural network.

[0006] The technical solution of the present invention is:

[0007] A structural damage identification method based on a physical information branch convolutional neural network comprises 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; and generate training samples for training a physical information branch convolutional neural network based on 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 to extract the damage features in the modal parameters respectively, and after fusing various damage features, map to obtain a damage identification result;

[0010] Step 3: Using the training samples obtained in step 1, the physical information branch convolutional neural network is trained, wherein the loss function used in the training includes two physical information, namely, sensitivity analysis of modal parameters and spatial sparsity of damage;

[0011] Step 4: Obtain the actual modal parameters of the structure to be evaluated and input them into the trained physical information branch convolutional neural network to obtain the structural damage identification results.

[0012] Furthermore, the modal parameters include characteristic frequencies and vibration shapes.

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

[0014] Loss=Loss MSE +Loss Phy

[0015] Where Loss MSE Represents the mean square error loss, Loss Phy Indicates physical information loss;

[0016] The physical information loss Loss Phy for:

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

[0018] Where Loss Mod Represents the modal parameter loss, Loss Spa represents the sparsity loss, β Mod and β Spa represents the modality hyperparameter and the sparsity hyperparameter.

[0019] Furthermore, the modal parameter loss Loss Mod for:

[0020] Loss Mod =Loss FErr +Loss SErr

[0021] Where Loss FErr Indicates frequency error loss, LossSErr represents the mode error loss;

[0022] The frequency error loss Loss FErr for:

[0023]

[0024] Where N b Indicates the number of samples in the same batch of input samples during training, N order Indicates the order of the characteristic frequency vector, S λ represents the sensitivity matrix of the eigenfrequency vector relative to the damage parameter vector, represents the predicted value of the damage parameter vector of the t-th sample in the same batch, λ Dt Represents the characteristic frequency vector corresponding to the t-th sample in the same batch of input samples, λ U Represents the eigenfrequency vector corresponding to the structure to be evaluated in the undamaged state;

[0025] The vibration mode error loss Loss SErr for:

[0026]

[0027] Where N dof Indicates the number of degrees of freedom, S φ represents the sensitivity matrix of the mode shape matrix relative to the damage parameter vector, φ Dt Represents the vibration mode matrix corresponding to the t-th sample in the same batch of input samples, φ U Represents the mode shape matrix corresponding to the structure to be evaluated in the undamaged state.

[0028] Furthermore, the sparsity loss Loss Spa for:

[0029]

[0030] Where N b Indicates the number of samples in the same batch of input samples during training, N el represents the number of elements in the finite element model, represents the kth element in the damage parameter vector prediction value of the tth sample in the same batch of input samples, and σ is a hyperparameter in the fractional function F(·).

[0031] Furthermore, the sensitivity matrix S of the characteristic frequency vector relative to the damage parameter vector λ The sensitivity matrix S of the harmonic mode matrix relative to the damage parameter vector φ They are:

[0032]

[0033] In the formula is the i-th order eigenfrequency in the eigenfrequency vector in the undamaged state, is the i-th column vector in the vibration mode matrix in the undamaged state, θ j represents the damage parameter of the jth unit, that is, the jth element of the damage parameter vector, K j represents the stiffness matrix of the jth element, express The transpose of .

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

[0035] Furthermore, in step 1, Gaussian random noise is added to the characteristic frequency vector and mode matrix calculated using the finite element model to simulate measurement noise, and Gaussian random noise is added to the damage parameter vector composed of damage parameters of different units in the finite element model to simulate finite element modeling error.

[0036] Furthermore, in step 1, the characteristic frequency vector and mode shape matrix with added Gaussian random noise are preprocessed by maximum and minimum normalization to obtain the characteristic frequency vector and mode shape matrix for training the physical information branch convolutional neural network.

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

[0038] Beneficial effects

[0039] The beneficial effects of the present invention are as follows: the structural damage identification method based on physical information branch convolutional neural network provided by the present invention improves the damage identification accuracy under limited data sets and solves the problem of low damage identification accuracy under limited data sets, such as Figure 3As shown in the figure, the branch architecture extracts damage features from frequency and vibration mode respectively, which improves the efficiency of extracting damage features from a limited data set. The modal parameter sensitivity and the spatial sparsity of damage are added to the loss function as physical information, so that the network training not only relies on a limited data set, but also obtains the help of physical information, which improves the network's efficiency in utilizing the limited data set and improves the network's damage identification accuracy under a limited data set.

[0040] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments with reference to the following drawings, in which:

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

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

[0044] Figure 3 Graph showing how the verification mean square error loss of ResNet and the network (PIBCNN) provided by the present invention changes with the number of training times under different data amounts shown in Example 1 of the present invention;

[0045] Figure 4 Graph showing the training loss and validation loss of the ResNet shown in Example 1 of the present invention and the network (PIBCNN) provided by the present invention as a function of the number of training times;

[0046] Figure 5 This is a comparison chart of damage identification results of the ResNet shown in Example 1 of the present invention and the network (PIBCNN) provided by the present invention;

[0047] Figure 6 This is a diagram of the actual structure of damage identification shown in Example 2 of the present invention;

[0048] Figure 7 Graph showing the training loss and validation loss of the ResNet shown in Example 2 of the present invention and the network (PIBCNN) provided by the present invention as a function of the number of training times;

[0049] Figure 8 This is a comparison chart of the damage identification results of the ResNet shown in Example 2 of the present invention and the network (PIBCNN) provided by the present invention. DETAILED DESCRIPTION

[0050] The following describes in detail embodiments of the present invention. The embodiments are exemplary and intended to explain the present invention, but are not to be construed as limiting the present invention.

[0051] Example 1:

[0052] A structural damage identification method based on a physical information branched 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; use the finite element model to generate modal parameters of different damage states; and generate training samples for training a physical information branch convolutional neural network based on the modal parameters.

[0054] In this embodiment, finite element modeling is performed using MATLAB programming software, and the model schematic diagram is shown in FIG. Figure 2 The model has 54 units in total. By setting the stiffness reduction values of different units in the finite element model, different degrees of damage to 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 actual measurement conditions, this example sets the measured modal orders to the first eight, and the measured degrees of freedom to the lateral and longitudinal degrees of freedom of nodes 5, 7, 9, 11, 13, and 15, for a total of 12 degrees of freedom. Therefore, the simulated eigenfrequency vector λ is an 8×1 vector, and the mode shape matrix φ is a 12×8 matrix.

[0056] In order to simulate different damage conditions, this embodiment sets single damage condition and multiple damage condition. In single damage condition, any element is individually damaged, and the damage degree (i.e. stiffness reduction value) ranges from 1% to 20%, with a step size of 1%, thus generating 54*20=1080 samples in total. In multiple damage condition, any two elements are simultaneously damaged, and the damage degree ranges from 2% to 20%, with a step size of 2%, thus generating 54*20=1080 samples in total. In summary, a total of 144,180 samples were used for network training, validation, and testing. The theoretical composition of each sample is the corresponding eigenfrequency vector λ and mode matrix φ, with the label being the damage parameter vector θ. The damage parameter vector θ is a 54×1 vector, with each element value being the stiffness reduction value of the corresponding element, ranging from 0 to 1.

[0057] To more realistically reflect the actual situation, based on the above calculation results, 1% and 5% Gaussian random noise are added to the eigenfrequency vector and mode matrix respectively to simulate measurement noise. At the same time, 1% Gaussian random noise is added to the damage parameter vector to simulate finite element modeling error. The specific method of adding noise is as follows:

[0058] λ Noise =(1+0.01μ)λ

[0059] φ Noise =(1+0.05μ)φ

[0060] θ Noise =θ+0.01μ

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

[0062] The generated eigenfrequency vector λ Noise Harmonic mode matrix φ Noise Before being input into the neural network, they are preprocessed separately to eliminate the dimension and make the training more stable. This embodiment uses the maximum and minimum normalization preprocessing method to scale the data to [0,1]. The specific operation is shown in the following formula:

[0063]

[0064] in is λ Noise The i-th order characteristic frequency, max(λ Noise ) and min(λ Noise ) represent λ Noise The maximum and minimum values of the elements in , Represents the normalized eigenfrequency vector λ norm The i-th order characteristic frequency in ; is φ Noise The element in row i and column p, max(φ Noise ) and min(φ Noise ) represent φ Noise The maximum and minimum values of the elements in , represents the normalized mode matrix φ norm The element in row i and column p.

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

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

[0067] like Figure 1As shown, the physical information branch convolutional neural network adopts a branch architecture to extract damage features in frequency and vibration mode data respectively, and finally maps the two damage features to the damage identification results after fusing them.

[0068] Specifically, for the characteristic frequency vector λ norm , use one-dimensional convolution operation to extract the damage characteristics, for the vibration matrix φ norm , a two-dimensional convolution operation with residual connection is used to extract the damage features; then the damage features extracted from the two 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 the damage recognition results.

[0069] In this embodiment, the one-dimensional convolution branch and the two-dimensional convolution branch include five stages respectively. The convolution kernel size of each stage is 2×1 and 2×2 respectively, and the number of channels is 16k, 32k, 64k, 128k and 128k respectively. Each stage has a residual connection, and 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 are 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 types of physical information, namely the sensitivity analysis of modal parameters and the spatial sparsity of damage. The spatial sparsity of damage is characterized by 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 limited training samples.

[0073] The specific loss function is:

[0074] Loss=Loss MSE +Loss Phy

[0075] Where Loss MSE Represents the mean square error loss, Loss Phy Indicates physical information loss.

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

[0077]

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

[0079] For physical information loss Loss Phy As mentioned above, it contains two kinds of physical information, namely the sensitivity analysis of modal parameters and the spatial sparsity of damage, which are specifically defined as:

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

[0081] Where Loss Mod Represents the modal parameter loss, Loss Spa represents the sparsity loss, β Mod and β Spa represents the modality hyperparameter and the sparsity hyperparameter.

[0082] The modal parameter loss Loss Mod Defined as:

[0083] Loss Mod =Loss FErr +Loss SErr

[0084] Where Loss FErr Indicates frequency error loss, Loss SErr Represents the mode error loss.

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

[0086]

[0087] Where N order represents the order of the characteristic frequency vector, which is 8 in this embodiment; S λ represents the sensitivity matrix of the characteristic frequency vector relative to the damage parameter vector, λ Dt Represents the characteristic frequency vector corresponding to the t-th sample in the same batch of input samples, λ U Represents the eigenfrequency vector corresponding to the structure to be evaluated in the undamaged state.

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

[0089]

[0090] Where N dof Indicates the number of degrees of freedom, which is 12 in this embodiment; S φ represents the sensitivity matrix of the mode shape matrix relative to the damage parameter vector, φ Dt Represents the vibration mode matrix corresponding to the t-th sample in the same batch of input samples, φ U Represents the mode shape matrix corresponding to the structure to be evaluated in the undamaged state.

[0091] The sparsity loss Loss Spa Defined as:

[0092]

[0093] In the formula represents the kth element in the damage parameter vector prediction value of the tth 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] Sensitivity matrix S of the eigenfrequency vector relative to the damage parameter vector λ The sensitivity matrix S of the harmonic mode matrix relative to the damage parameter vector φ Defined as:

[0097]

[0098] In the formula is the i-th order eigenfrequency in the eigenfrequency vector in the undamaged state, is the i-th column vector in the vibration mode matrix in the undamaged state, θ j represents the damage parameter of the jth unit, K j represents the stiffness matrix of the jth element, express The transpose of .

[0099] The mean square error loss function is used for verification, the Adam optimizer is used as the network training optimizer, and the learning rate decay strategy adopts the step-by-step decay strategy.

[0100] In order to demonstrate the improvement of damage recognition accuracy of the physical information branch convolutional neural network obtained in the present invention, a ResNet network was also established in this embodiment for damage recognition. The curves of the training and verification loss values of the two networks with the number of training times are shown in Figure 2. Figure 4 As shown in the figure, it can be seen that the network provided by the present invention has lower loss in the verification set, that is, the damage identification accuracy is higher.

[0101] Step 4: Use the trained network to identify structural damage:

[0102] For the actual collected structural acceleration data, the modal analysis method is used to convert the acceleration data into characteristic frequency vectors and mode shape matrices. The characteristic frequency vectors and mode shape matrices 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 Figure 5 In display.

[0103] For comparison, the eigenfrequency vector and mode matrix of the actual structure are also input into ResNet for damage identification, and the visualization results are also shown in Figure 5 The results show that ResNet misidentifies unit 38, while the network provided by the present invention accurately identifies the damage at units 33 and 46, achieving more accurate damage identification.

[0104] Example 2:

[0105] This embodiment is for Figure 6 The actual cantilever beam structure shown is used for structural damage identification:

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

[0107] In this embodiment, excluding the accelerometer at the fixed end, a total of five effective accelerometers were evenly spaced on the actual cantilever beam to measure horizontal acceleration at corresponding measurement points, corresponding to nodes 3, 5, 7, 9, and 11 of the finite element numerical model. Because the first-order frequency of the cantilever beam is difficult to measure, the experimentally measured second- to seventh-order modes of the cantilever beam were used as modal parameters.

[0108] The training data includes single damage and multiple damage conditions. When generating a single damage dataset, any unit is individually set to be damaged, and the damage degree is evenly distributed between 1% and 50%, with a distance of 1%, generating a total of 500 samples. When generating a multi-damage dataset, any two units are set to be damaged, and the damage degree is evenly distributed between 2% and 50%, with a distance of 2%, generating a total of 28,125 samples. In addition, any three units are set to be damaged, and the damage degree is evenly distributed between 10% and 50%, with a distance of 10%, generating a total of 15,000 samples. In summary, there are 43,625 samples with different damage conditions in the dataset for training and validating the neural network. Consistent with the numerical experiment in Example 1, the above dataset is randomly shuffled and divided into a training set and a validation set in a ratio of 7:3. In addition, to simulate various uncertainties in the laboratory, in the above dataset, 2% Gaussian white noise is added to the frequency and 10% Gaussian white noise is added to the vibration mode. The noise addition method and preprocessing method are consistent with Example 1.

[0109] Step 2: Build and train the Physical Information Branched Convolutional Neural Network (PIBCNN):

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

[0111] Step 3: Use the trained network to identify structural damage:

[0112] In order to collect the measured modal parameters under different damage conditions, including the characteristic frequency vector and vibration mode matrix, this embodiment uses the hammer method to conduct a series of vibration tests. First, one end of the beam is fixed to the bracket, and 6 acceleration sensors are equidistantly arranged on the beam to measure its horizontal acceleration response. Secondly, the cantilever beam is struck with a hammer to cause it to vibrate. The acceleration signal generated by the vibration is transmitted to the constant source current adapter and recorded by the dynamic and static strain acquisition and analysis system. Finally, the acceleration signal is acquired and modally analyzed using the signal acquisition system and modal analysis system in the desktop computer to obtain the modal parameters.

[0113] The acquired eigenfrequency vector and mode matrix of the real structure are input into the built and trained PIBCNN and ResNet, and the calculated damage identification visualization results are shown in Figure 8The results show that ResNet seriously underestimates the damage at unit 1 and unit 9, while the PIBCNN provided by the present invention accurately identifies all damages and achieves more accurate damage identification.

[0114] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and purpose of the present invention.

Claims

1. A structural damage identification method based on a physical information branch convolutional neural network, characterized by: The following steps are involved: 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; and generate training samples for training a physical information branch convolutional neural network based on the modal parameters. Step 2: Building a physical information branch convolutional neural network; the physical information branch convolutional neural network adopts a branch structure to extract the damage features in the modal parameters respectively, and after fusing various damage features, map to obtain a damage identification result; Step 3: Using the training samples obtained in step 1, the physical information branch convolutional neural network is trained, wherein the loss function used in the training includes two physical information, namely, sensitivity analysis of modal parameters and spatial sparsity of damage; Step 4: Obtain the actual modal parameters of the structure to be evaluated and input them into the trained physical information branch convolutional neural network to obtain the structural damage identification results.

2. The structural damage identification method based on a physical information branched convolutional neural network according to claim 1, characterized in that: The modal parameters include characteristic frequencies and vibration shapes.

3. The structural damage identification method based on a physical information branched convolutional neural network according to claim 2, characterized in that: The loss function used during training is: Loss=Loss MSE +Loss Phy Where Loss MSE Represents the mean square error loss, Loss Phy Indicates physical information loss; The physical information loss Loss Phy for: Loss Phy =β Mod Loss Mod +β Spa Loss Spa Where Loss Mod Represents the modal parameter loss, Loss Spa represents the sparsity loss, β Mod and β Spa represents the modality hyperparameter and the sparsity hyperparameter.

4. The structural damage identification method based on a physical information branched convolutional neural network according to claim 3, characterized in that: The modal parameter loss Loss Mod for: Loss Mod =Loss FErr +Loss SErr Where Loss FErr Indicates frequency error loss, Loss SErr represents the mode error loss; The frequency error loss Loss FErr for: Where N b Indicates the number of samples in the same batch of input samples during training, N order Indicates the order of the characteristic frequency vector, S λ represents the sensitivity matrix of the eigenfrequency vector relative to the damage parameter vector, represents the predicted value of the damage parameter vector of the t-th sample in the same batch, λ Dt Represents the characteristic frequency vector corresponding to the t-th sample in the same batch of input samples, λ U Represents the eigenfrequency vector corresponding to the structure to be evaluated in the undamaged state; The vibration mode error loss Loss SErr for: Where N dof Indicates the number of degrees of freedom, S φ represents the sensitivity matrix of the mode shape matrix relative to the damage parameter vector, φ Dt Represents the vibration mode matrix corresponding to the t-th sample in the same batch of input samples, φ U Represents the mode shape matrix corresponding to the structure to be evaluated in the undamaged state.

5. The structural damage identification method based on physical information branched convolutional neural network according to claim 3 is characterized by: The sparsity loss Loss Spa for: Where N b Indicates the number of samples in the same batch of input samples during training, N el represents the number of elements in the finite element model, represents the kth element in the damage parameter vector prediction value of the tth sample in the same batch of input samples, and σ is a hyperparameter in the fractional function F(·).

6. The structural damage identification method based on physical information branched convolutional neural network according to claim 4, characterized in that: characteristic frequency The sensitivity matrix S of the vector relative to the damage parameter vector λ The sensitivity matrix S of the harmonic mode matrix relative to the damage parameter vector φ They are: In the formula is the i-th order eigenfrequency in the eigenfrequency vector in the undamaged state, is the i-th column vector in the vibration mode matrix in the undamaged state, θ j represents the damage parameter of the jth unit, K j represents the stiffness matrix of the jth element, express The transpose of .

7. The structural damage identification method based on physical information branched convolutional neural network according to claim 1, characterized in that: In step 1, the process of generating modal parameters of different damage states using the finite element model is as follows: by setting the stiffness reduction values of different units in the finite element model as damage parameters, simulating different degrees of damage of different units, and using the finite element model to calculate the modal parameters under different damage states.

8. The structural damage identification method based on physical information branched convolutional neural network according to claim 7, characterized in that: In step 1, Gaussian random noise is added to the characteristic frequency vector and mode matrix calculated using the finite element model to simulate measurement noise, and Gaussian random noise is added to the damage parameter vector composed of damage parameters of different units in the finite element model to simulate finite element modeling error.

9. The structural damage identification method based on physical information branched convolutional neural network according to claim 8, characterized in that: In step 1, the characteristic frequency vector and mode shape matrix with added Gaussian random noise are preprocessed by maximum and minimum normalization to obtain the characteristic frequency vector and mode shape matrix for training the physical information branch convolutional neural network.

10. The structural damage identification method based on physical information branched convolutional neural network according to claim 1, characterized in that: In step 2, in the physical information branch convolutional neural network, for the characteristic frequency vector, a one-dimensional convolution operation is used to extract the damage features therein, and for the vibration mode matrix, a two-dimensional convolution operation with residual connection is used to extract the damage features therein; then the damage features extracted from the two types 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 the damage identification results.

Images