Cross-Domain Structural Damage Identification Method Based on Digital Twin and Spatiotemporal Feature Fusion

Abstract

The present invention discloses a cross-domain structural damage identification method based on digital twin and spatio-temporal feature fusion, belonging to the technical field of structural health monitoring. The present invention proposes a cross-domain transfer learning damage identification technical solution that fuses a digital twin model and spatio-temporal features of structural multi-sensor data. By abstracting the connection relationship between the measurement points of the digital twin model and the actual physical structure into a graph structure, a spatio-temporal feature learning model for simultaneously processing multi-sensor time-domain data and spatial feature information is designed based on classical time-domain convolution and graph convolutional network, providing a specific model architecture for deeply fusing the spatio-temporal features of the structural digital twin model and multi-sensor data of the actual structure to achieve structural damage detection in cross-data domain scenarios, so as to make up for problems such as insufficient actual damage data and insufficient fusion of structural multi-dimensional feature data.

Description

Technical Field

[0001] The present invention belongs to the technical field of structural health monitoring, and particularly relates to a cross-domain structural damage identification method based on digital twin and spatio-temporal feature fusion. Background Art

[0002] With the development of deep learning technology, the research on using various neural network models for structural damage identification has gradually received extensive attention. Existing research mainly trains neural network models by collecting a large amount of damage data under different damage scenarios, and identifies different types of damage according to the feature mapping of the damage data. However, in actual engineering measurements, structural damage monitoring data is very scarce, and neural network models often have insufficient generalization ability due to too little training data, thus reducing the recognition accuracy.

[0003] In order to make up for the scarcity of measured data, existing technologies generate batch simulation data under multiple damage scenarios as training data to train a certain neural network model by establishing a structural finite element model or a digital twin model, and use a small amount of measured data to fine-tune the neural network model to achieve cross-domain data-driven damage task migration identification. In the current mainstream research methods for fusing cross-domain data and transfer learning, it is considered that structural vibration data is the most potential data for structural damage state identification. Most research is based on structural time-domain vibration data, with the goal of minimizing the difference of cross-domain data, and uses time-domain data or variant data (such as frequency-domain and time-frequency-domain data) for deep learning model training and feature learning, and finally realizes the mapping relationship fitting from structural vibration data to damage state detection. However, these research methods ignore the node space features corresponding to the structural vibration data, and there are few research results on fusing the structural vibration data and the spatial relationship of the positions where the sensors are located to carry out damage identification.

[0004] According to the mechanism of sensor optimal placement, the sensor placement measurement points are a set of the most critical nodes obtained by comprehensively optimizing by combining means such as structural dynamic characteristics and modal analysis. The data set or single training sample data constructed from the vibration data at these nodes only reflects the time history response information of each node, and does not reflect the spatial layout of each node. According to the structural dynamic characteristics, there must be a certain correlation in the dynamic responses of different nodes of the structure that obeys the structural dynamic characteristics, and this spatial correlation may have a potential impact on the time-domain response signals between multi-sensor measurement points. Summary of the Invention

[0005] In view of the above deficiencies in the prior art, relying on the dominant position of structural vibration data, this invention considers the key issues ignored by existing research results and proposes a cross-domain structural damage identification method based on digital twin and spatio-temporal feature fusion to solve the problem that in the current research methods for cross-domain structural damage identification, mainly time-domain, frequency-domain, time-frequency domain or other variant data of structural vibration data are used, and few studies consider fusing the spatial features of multi-sensor measurement points of the structure for damage discrimination.

[0006] To achieve the above invention purpose, the technical solution adopted by this invention is: a cross-domain structural damage identification method based on digital twin and spatio-temporal feature fusion, including the following steps:

[0007] S1. Construct a digital twin model corresponding to the actual engineering structure and optimize the sensor layout based on the digital twin model;

[0008] S2. Construct a two-domain spatio-temporal data set for the actual engineering structure under various damage scenarios;

[0009] The two-domain spatio-temporal data set includes a measured-domain data set and a simulated-domain data set; the measured-domain data set includes the measured time-domain vibration data of the actual engineering structure under various damage scenarios and its corresponding adjacency matrix; the simulated-domain data set includes the simulated time-domain vibration data generated by the corresponding digital twin model of the actual engineering structure under various damage scenarios and its corresponding adjacency matrix; the adjacency matrix characterizes the spatial position association of the sensors after optimized layout;

[0010] S3. Construct a spatio-temporal feature learning network model and train it using the spatio-temporal data set to obtain a cross-domain structural damage identification model;

[0011] The cross-domain structural damage identification model includes an input layer, a time-domain feature extraction layer, a spatial feature extraction layer, and a damage prediction layer;

[0012] S4. Process the time-domain vibration data of the two structures with damage to be identified through a time-domain adaptive model to obtain adaptive time-domain data;

[0013] S5. Input the adaptive time-domain data and its corresponding adjacency matrix into the cross-domain structural damage identification model for processing to obtain a cross-domain structural damage identification result.

[0014] Further, in step S3, the method for the cross-domain structural damage identification model to process the input data is specifically:

[0015] S31. Receive the time-domain vibration data and the adjacency matrix through the input layer and construct them into a spatio-temporal feature matrix;

[0016] S32. Separately convolve the time-domain vibration data of each sensor in the spatio-temporal feature matrix through the time-domain feature extraction layer to extract the corresponding time-domain features;

[0017] S33. Based on the extracted time-domain features, extract the corresponding multi-dimensional spatial features through the spatial feature extraction layer;

[0018] S34. Flatten the multi-dimensional spatial features into one-dimensional damage information through the damage prediction layer and output the damage recognition result.

[0019] Further, in the cross-domain structural damage recognition model, the time-domain feature extraction layer is a one-dimensional convolutional neural network, and the corresponding time-domain features are extracted by separately convolving the time-domain vibration data of each sensor.

[0020] Further, in the cross-domain structural damage recognition model, the spatial feature extraction layer extracts the corresponding multi-dimensional spatial features from the input graph structure;

[0021] In the graph structure, the graph nodes correspond to the positions of the sensors, the features of the graph nodes are the extracted time-domain features, and the connection relationship of the graph nodes is the spatial position association of the sensors represented by the adjacency matrix;

[0022] The spatial feature extraction layer includes a first graph convolutional layer, a first ReLU activation function, a second graph convolutional layer, a second ReLU activation function, and a global average pooling layer connected in sequence;

[0023] The first graph convolutional layer and the second graph convolutional layer both embed a graph attention mechanism.

[0024] Further, in the cross-domain structural damage recognition model, the damage prediction layer includes a fully connected layer and a Softmax function layer, which respectively perform one-dimensional flattening on the input multi-dimensional spatial features and map the score values of the damage prediction to the interval [0,1] to obtain the damage recognition result.

[0025] Further, the training loss function of the cross-domain structural damage recognition model is:

[0026]

[0027]

[0028] In the formula, and respectively represent the score value and the predicted probability value that the th sample belongs to the true damage label , represents that the true damage label of the th sample is , represents the total number of samples, represents the total number of true damage labels.

[0029] Furthermore, in the step S3, the method for training the spatio-temporal feature learning network model using the spatio-temporal data set to obtain the cross-domain structural damage recognition model is as follows:

[0030] Train the spatio-temporal feature learning network model through the simulated domain data set to obtain a pre-trained spatio-temporal feature learning network model;

[0031] Fine-tune the pre-trained spatio-temporal feature learning network model through the measured domain data set to obtain a cross-domain structural damage recognition model.

[0032] Furthermore, in the step S4, the time domain adaptive model includes the first fast Fourier transform, a one-dimensional deep convolutional neural network, and the second Fourier transform;

[0033] The step S4 includes the following sub-steps:

[0034] S41. Perform the first fast Fourier transform on the time domain vibration data of the two structures respectively to obtain the corresponding frequency domain data;

[0035] S42. Input the frequency domain data of the two structures into the corresponding one-dimensional deep convolutional neural network respectively for data alignment;

[0036] S43. Perform the second Fourier transform on the aligned frequency domain data to restore it to the time domain to obtain the adaptive time domain data.

[0037] Furthermore, in the step S4, the loss function of the time domain adaptive model is the MMD loss function.

[0038] The beneficial effects of the present invention are as follows:

[0039] In the existing research on using two-domain data for transfer learning to improve the accuracy of structural damage recognition, the spatial features between the two-domain structures are not considered. The present invention proposes the SfLN model, which integrates the time domain and spatial feature information between key measurement points of the two-domain data, thereby greatly improving the accuracy of cross-domain structural damage recognition, and at the same time providing a new solution for solving the damage recognition of cross-structures. The specific beneficial effects are as follows:

[0040] (1) Propose a cross-domain transfer learning damage recognition technical solution that integrates the digital twin model and the spatio-temporal features of structural multi-sensor data, and solve the problem of incomplete sensor data information (not considering sensor spatial feature information) considered in the field of structural damage recognition.

[0041] (2) A spatio-temporal feature learning model is proposed that can simultaneously process the time-domain data and spatial feature information of multiple sensors, providing a specific model architecture for the spatio-temporal features of the multi-sensor data of the deep fusion structure digital twin model and the actual structure.

[0042] (3) Based on the spatio-temporal feature learning model trained by the structure digital twin model and the measured data, a cross-structure damage identification technology architecture is proposed to solve problems such as the scarcity of damage data in actual structures and the cross-structure damage identification tasks with the same sensor measurement point layout. Description of the Drawings

[0043] Figure 1 It is a flowchart of the cross-domain structure damage identification method based on digital twin and spatio-temporal feature fusion provided by the present invention.

[0044] Figure 2 It is a structural diagram of the spatio-temporal feature learning network model (SfLN) provided by the present invention. Detailed Embodiments

[0045] The following describes the detailed embodiments of the present invention to facilitate those skilled in the art of the present technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the detailed embodiments. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

[0046] The embodiment of the present invention provides a cross-domain structure damage identification method based on digital twin and spatio-temporal feature fusion, as Figure 1 shown, including the following steps:

[0047] S1. Construct a digital twin model corresponding to the actual engineering structure, and optimize the sensor layout based on the digital twin model;

[0048] S2. Construct a two-domain spatio-temporal data set of the actual engineering structure under various damage scenarios;

[0049] S3. Construct a spatio-temporal feature learning network model, and train it with the spatio-temporal data set to obtain a cross-domain structure damage identification model;

[0050] Among them, the cross-domain structure damage identification model includes an input layer, a time-domain feature extraction layer, a spatial feature extraction layer, and a damage prediction layer;

[0051] S4. Process the time-domain vibration data of the two structures with damage to be identified through a time-domain adaptive model to obtain adaptive time-domain data;

[0052] S5. Input the adaptive time-domain data and its corresponding adjacency matrix into the cross-domain structural damage identification model for processing to obtain the cross-structural damage identification results between the two structures.

[0053] In step S1 of the embodiment of the present invention, a digital twin model is constructed according to parameters such as engineering data, dimensions, materials, and environment of the actual engineering structure, and the sensors are optimally arranged using information such as modal vibration mode data of the digital twin model to obtain a reasonable sensor arrangement scheme.

[0054] In step S2 of the embodiment of the present invention, a small amount of measured data in various damage scenarios of the actual engineering structure is collected, as well as a large amount of simulated data in various damage scenarios generated by the digital twin model, and a spatio-temporal dataset is constructed based on the spatial position information of each sensor; specifically, the two-domain spatio-temporal dataset in this embodiment includes a measured-domain dataset and a simulated-domain dataset; the measured-domain dataset includes the measured time-domain vibration data of the actual engineering structure in various damage scenarios and its corresponding adjacency matrix; the simulated-domain dataset includes the simulated time-domain vibration data generated by the corresponding digital twin model of the actual engineering structure in various damage scenarios and its corresponding adjacency matrix; wherein, the adjacency matrix characterizes the spatial position association of the sensors after optimal arrangement.

[0055] In step S3 of the embodiment of the present invention, the Spatiotemporal feature Learning Network model (SfLN) is as Figure 2 shown, and the cross-domain structural damage identification model obtained by training it includes an input layer, a time-domain feature extraction layer, a spatial feature extraction layer, and a damage prediction layer; regarding the spatial position where the sensor is located as a "graph" node and the spatial distribution association between sensors as the edge of the "graph", the local time-domain vibration data and the global node "graph" information are synchronously processed through a one-dimensional convolutional network, a graph convolutional network, and a graph attention mechanism to extract the common time-domain data features of the two-domain structure and the spatial characteristics of the key nodes of the structure from local to global, thereby realizing the structure damage feature learning and task migration that fuse the spatio-temporal multi-dimensional feature information of the structure, and is expected to improve the accuracy and reliability analysis of structure damage identification driven by cross-domain data.

[0056] Specifically, in this embodiment, the method for the cross-domain structural damage identification model to process the input data is specifically as follows:

[0057] S31. Receive the time-domain vibration data and the adjacency matrix through the input layer and construct them into a spatio-temporal feature matrix;

[0058] S32. Separately convolve the time-domain vibration data of each sensor in the spatio-temporal feature matrix through the time-domain feature extraction layer to extract the corresponding time-domain features;

[0059] S33. Extract corresponding multi-dimensional spatial features through the spatial feature extraction layer based on the extracted time-domain features;

[0060] S34. Flatten the multi-dimensional spatial features into one-dimensional damage information through the damage prediction layer and output the damage recognition result.

[0061] In step S31 of this embodiment, the input layer, as an important layer for the cross-domain structural damage recognition model to collect structural spatio-temporal feature data, mainly inputs the time-domain vibration data matrix collected under different damage scenarios of the structure and the adjacency matrix reflecting the spatial position association of sensors. Among them, the time-domain vibration data matrix only reflects the time-history response information of the key nodes of the structure and does not contain the spatial attributes of the sensors, while the adjacency matrix only reflects the spatial relationship of the key nodes of the structure and does not contain any sensor time-history data. For the convenience of subsequent module processing, in step S31, the two are combined to construct a spatio-temporal feature matrix that simultaneously contains the structural time-history data and the node spatial feature information, which is expressed as:

[0062]

[0063] Among them, is a time-domain data matrix of dimension ( represents the number of sensors, represents the number of data points collected), is an adjacency matrix of dimension (the adjacency matrix is an important step in graph convolution and also an important step in integrating spatial information. The adjacency matrix converts the physical connection relationship between sensors into a matrix. For example, if sensor 1 and sensor 2 are physically connected, the adjacency matrix constructs A 12 = 1), then contains the complete spatio-temporal feature information under different damage scenarios of the structure. Thus, it is input into the subsequent structure of SfLN, which is conducive to exploring the restrictive relationship between the time-domain response of the key nodes of the structure and their spatial features, and deeply integrating the structural spatio-temporal feature information for damage feature extraction and recognition.

[0064] In step S32 of this embodiment, in the cross-domain structural damage recognition model, the time-domain feature extraction layer is a one-dimensional convolutional neural network (1DCNN), which extracts the corresponding time-domain features by performing a separate convolution operation on the time-domain vibration data of each sensor.

[0065] Specifically, taking the sensor data of a channel under a certain damage scenario of the structure as an example, its calculation method is as follows:

[0066]

[0067] Among them, * represents the convolution operation, Represent the acceleration vibration response data of a certain sensor, represent the convolution kernel, represent the bias, represent the time-domain feature data of the sensor output.

[0068] In step S33 of this embodiment, in the cross-domain structure damage identification model, the spatial feature extraction layer extracts corresponding multi-dimensional spatial features from the input graph structure; in the graph structure, the graph nodes correspond to the positions of the sensors, the features of the graph nodes are the extracted time-domain features, and the connection relationship of the graph nodes is the spatial position association of the sensors represented by the adjacency matrix;

[0069] The spatial feature extraction layer includes a first graph convolutional layer (GCN1), a first ReLU activation function (ReLU1), a second graph convolutional layer (GCN2), a second ReLU activation function (ReLU2), and a global average pooling layer (GAP) connected in sequence; at the same time, a graph attention mechanism is embedded in both the first graph convolutional layer (GCN1) and the second graph convolutional layer (GCN2).

[0070] Among them, by embedding the graph attention mechanism, it is possible to better focus on the importance between each sensor measurement point, further assisting the graph convolutional network to better extract the spatial information of the structure.

[0071] In step S34 of this embodiment, in the cross-domain structure damage identification model, the damage prediction layer includes a fully connected layer and a Softmax function layer, which respectively perform one-dimensional flattening on the input multi-dimensional spatial features and map the score values of the damage prediction to the [0,1] interval to obtain the damage identification result.

[0072] Specifically, in this embodiment, deep learning is used for structure damage identification. Essentially, it can be reduced to a multi-classification problem of mapping damage positions or degrees from structural response data. In the present invention, after the time-domain feature matrix is processed by the spatial feature extraction module, multi-dimensional spatial features are obtained, which need to be flattened into one-dimensional damage information to complete the damage prediction. Therefore, the damage prediction layer mainly includes two parts: a fully connected layer and a Softmax function layer, which respectively complete information flattening and map the score values of the damage prediction to the [0,1] interval to obtain the damage identification result.

[0073] In the embodiment of the present invention, the calibration method of the considered structural true damage label always satisfies the probability distribution, while the damage prediction score values output by the fully connected layer do not necessarily satisfy this condition, resulting in the incomparability between the damage prediction values and the true values. Therefore, first, the Softmax function is used to map the damage prediction score values to the probability distribution in the interval [0, 1], and then the cross-entropy loss function is used to measure the difference between the probability distribution of the damage prediction and the true damage, and the performance of the SfLN model is optimized in reverse with the goal of minimizing the loss value. Therefore, the training loss function of the cross-domain structural damage identification model in this embodiment is as follows:

[0074]

[0075]

[0076] In the formula, and respectively represent the score value and the predicted probability value that the th sample belongs to the true damage label , represents that the true damage label of the th sample is , represents the total number of samples, represents the total number of true damage labels.

[0077] Through the above operations of the Softmax function layer and calculating the loss function, the discrimination ability of the SfLN model for different damage data can be measured, and the loss function value can directly reflect the parameter update process of the SfLN model. When a large loss value is obtained, it indicates that the model has a poor prediction effect on individual damage categories, prompting the model to further strengthen the learning of such damage features. In addition, during the training process, the Softmax cross-entropy loss can provide stable gradients, helping the SfLN model to maintain good stability and robustness when facing different damage features.

[0078] In step S3 of the embodiment of the present invention, the method for training the spatio-temporal feature learning network model by using the transfer learning strategy and obtaining the cross-domain structural damage identification model is as follows:

[0079] The spatio-temporal feature learning network model is trained through the simulated domain dataset to obtain a pre-trained spatio-temporal feature learning network model;

[0080] The pre-trained spatio-temporal feature learning network model is fine-tuned through the measured domain dataset to obtain a cross-domain structural damage identification model.

[0081] In step S4 of the embodiment of the present invention, when completing the cross-structure loss identification task, it is necessary to perform a time-domain data domain adaptation on the time-domain acceleration data of the two structures first. The specific method of time-domain adaptation is as follows: Perform a Fast Fourier Transformation (FFT) on the time-domain vibration signals collected under the healthy states of the two structures to obtain frequency-domain data, and then send the frequency-domain data into a One-Dimensional Deep Convolutional Neural Network (1DDCNN) to narrow the data distribution characteristics between the two structures. The data after narrowing is then restored to the time domain through an Inverse Fast Fourier Transform (IFFT).

[0082] Therefore, in step S4 of the embodiment of the present invention, a time-domain adaptation model is designed, which includes a first Fast Fourier Transform, a one-dimensional deep convolutional neural network, and a second Fourier Transform; specifically, the method for obtaining the adaptive time-domain data in step S4 is as follows:

[0083] S41. Perform a first Fast Fourier Transform on the time-domain vibration data of the two structures respectively to obtain the corresponding frequency-domain data;

[0084] S42. Input the frequency-domain data of the two structures into the corresponding one-dimensional deep convolutional neural networks respectively for data alignment;

[0085] S43. Perform a second Fourier Transform on the aligned frequency-domain data to restore it to the time domain to obtain the adaptive time-domain data.

[0086] In this embodiment, the loss function of the time-domain adaptation model is the MMD loss function, and the data alignment degree is measured by the Max Mean Discrepancy (MMD) method.

[0087] In this embodiment, MMD is used as the loss function of the time-domain adaptation model to guide its training, and the trained model is saved. The data of all states of the two structures collected subsequently all pass through this time-domain adaptation model to reduce the gap between the time-domain data of the two structures. Finally, the spatial feature data and the processed time-domain data are sent into the SfLN again, and the cross-structure damage identification between the two structures can be realized.

[0088] It should be noted that the cross-domain structure damage identification method provided in the present invention can be used for the structure damage identification across the measured domain and the simulation domain (cross-data domain), or for the cross-structure damage identification of two similar structures (cross-structure domain).

[0089] In an embodiment of the present invention, an example for verifying the effect of the above cross-domain structure damage identification method is provided. By comparing the two-dimensional convolutional neural network model (2DCNN) that does not consider spatial features and the SfLN model using the same data, the comparison table of damage identification accuracy rates is shown in Table 1.

[0090] Table 1 Comparison table of accuracy rates, recall rates, and F1 values

[0091]

[0092] In actual experiments, we set up 4 specific damage scenarios, including bolt loosening and bolt detachment, etc., and used hammer excitation to simulate the natural excitation of the actual structure. 7 sensors were arranged on the structure to collect vibration acceleration signals at different positions. It can be seen from Table 1 that the identification accuracy rate and various indicators of the SfLN in the present invention are significantly higher than those of the two-dimensional convolutional neural network model (2DCNN) that does not consider spatial features.

[0093] In the present invention, specific embodiments are applied to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

[0094] Those of ordinary skill in the art will realize that the embodiments described herein are for helping the reader understand the principle of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention according to the technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. Cross - domain Structural Damage Identification Method Based on Digital Twin and Spatiotemporal Feature Fusion, Characterized in that, It includes the following steps: S1. Construct a digital twin model corresponding to the actual engineering structure, and optimize the sensor layout based on the digital twin model; S2. Construct a two - domain spatiotemporal data set for the actual engineering structure under various damage scenarios; The two - domain spatiotemporal data set includes a measured - domain data set and a simulated - domain data set; the measured - domain data set includes the measured time - domain vibration data of the actual engineering structure under various damage scenarios and its corresponding adjacency matrix; the simulated - domain data set includes the simulated time - domain vibration data generated by the corresponding digital twin model of the actual engineering structure under various damage scenarios and its corresponding adjacency matrix; the adjacency matrix characterizes the spatial position association of the sensors after optimized layout; S3. Construct a spatiotemporal feature learning network model and train it using the spatiotemporal data set to obtain a cross - domain structural damage identification model; The cross - domain structural damage identification model includes an input layer, a time - domain feature extraction layer, a spatial feature extraction layer, and a damage prediction layer; S4. Process the time - domain vibration data of the two structures with damage to be identified through a time - domain adaptive model to obtain adaptive time - domain data; S5. Input the adaptive time - domain data and its corresponding adjacency matrix into the cross - domain structural damage identification model for processing to obtain the cross - domain structural damage identification result; In step S4, the time - domain adaptive model includes a first fast Fourier transform, a one - dimensional deep convolutional neural network, and a second Fourier transform; Step S4 includes the following sub - steps: S41. Perform the first fast Fourier transform on the time - domain vibration data of the two structures respectively to obtain the corresponding frequency - domain data; S42. Input the frequency - domain data of the two structures into the corresponding one - dimensional deep convolutional neural networks respectively for data alignment; S43. Perform the second Fourier transform on the aligned frequency - domain data to restore it to the time - domain to obtain the adaptive time - domain data.

2. The cross - domain structural damage identification method based on digital twin and spatiotemporal feature fusion according to claim 1, Characterized in that, In step S3, the method for the cross - domain structural damage identification model to process the input data is specifically: S31. Receive the time - domain vibration data and the adjacency matrix through the input layer and construct them into a spatiotemporal feature matrix; S32. Perform individual convolution on the time - domain vibration data of each sensor in the spatiotemporal feature matrix through the time - domain feature extraction layer to extract the corresponding time - domain features; S33. Based on the extracted time - domain features, extract the corresponding multi - dimensional spatial features through the spatial feature extraction layer; S34. Flatten the multi - dimensional spatial features into one - dimensional damage information through the damage prediction layer and output the damage identification result.

3. The cross - domain structural damage identification method based on digital twin and spatiotemporal feature fusion according to claim 1, Characterized in that, In the cross - domain structural damage identification model, the time - domain feature extraction layer is a one - dimensional convolutional neural network, and the corresponding time - domain features are extracted by performing individual convolution operations on the time - domain vibration data of each sensor.

4. The cross - domain structural damage identification method based on digital twin and spatiotemporal feature fusion according to claim 1, Characterized in that, In the cross-domain structure damage identification model, the spatial feature extraction layer extracts corresponding multi-dimensional spatial features from the input graph structure; In the graph structure, the graph nodes correspond to the positions of the sensors, the features of the graph nodes are the extracted time-domain features, and the connection relationship of the graph nodes is the spatial position association of the sensors represented by the adjacency matrix; The spatial feature extraction layer includes a first graph convolutional layer, a first ReLU activation function, a second graph convolutional layer, a second ReLU activation function, and a global average pooling layer connected in sequence; The graph attention mechanism is embedded in both the first graph convolutional layer and the second graph convolutional layer.

5. The cross-domain structure damage identification method based on digital twin and spatio-temporal feature fusion according to claim 1, wherein, in the cross-domain structure damage identification model, the damage prediction layer includes a fully connected layer and a Softmax function layer, which respectively perform one-dimensional flattening on the input multi-dimensional spatial features and map the score values of damage prediction to the interval [0,1] to obtain the damage identification result.

6. The cross-domain structure damage identification method based on digital twin and spatio-temporal feature fusion according to claim 1, wherein, The training loss function of the cross-domain structure damage identification model is as follows: In the formula, and respectively represent the score value and predicted probability value that the -th sample belongs to the true damage label . represents that the true damage label of the -th sample is , represents the total number of samples, represents the total number of true damage labels.

7. The cross-domain structure damage identification method based on digital twin and spatio-temporal feature fusion according to claim 1, wherein, in step S3, the method for training the spatio-temporal feature learning network model by using the spatio-temporal data set to obtain the cross-domain structure damage identification model is as follows: The spatio-temporal feature learning network model is trained through the simulation domain data set to obtain a pre-trained spatio-temporal feature learning network model; The pre-trained spatio-temporal feature learning network model is fine-tuned through the measured domain data set to obtain the cross-domain structure damage identification model.

8. The cross-domain structure damage identification method based on digital twin and spatio-temporal feature fusion according to claim 1, wherein, in step S4, the loss function of the time-domain adaptive model is the MMD loss function.

Images