Bridge structure damage identification method and system based on cross-channel attention

Abstract

The invention discloses a bridge structure damage identification method and system based on cross-channel attention, and the method comprises the steps: obtaining a bridge structure sample data set based on vibration acceleration measurement; a self-adaptive embedding mechanism module and a cross-channel attention mechanism module are introduced, and a bridge structure damage identification model is constructed; and based on the bridge structure damage identification model, carrying out damage identification on the bridge structure sample data set to obtain a bridge structure damage identification result. According to the method, the expansion rate and the receptive field can be adaptively adjusted according to the input response change, the space-time correlation among the multi-channel response data is better mined, and the bridge structure damage identification precision is improved. The bridge structure damage identification method and system based on cross-channel attention can be widely applied to the technical field of bridge structure damage identification.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge structure damage identification, and in particular to a bridge structure damage identification method and system based on cross-channel attention. Background Art

[0002] Traditional damage-sensitive feature extraction can be achieved through modal analysis, statistical characteristics analysis, signal decomposition, etc. However, the implementation effect of these methods depends to a certain extent on the validity of the initial value assumptions or prior knowledge. With the development of computer technology, with convolutional neural networks as the core algorithm, the research field has formed two major research hotspots: crack recognition based on machine vision and damage parameter recognition based on vibration response. However, traditional convolutional neural networks use fixed-size (and parameter) convolution kernels for response embedding, which has problems such as limited receptive field range and weak global attention ability, which easily leads to insufficient accuracy and poor stability of structural damage recognition representation of the extracted features. Summary of the invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a bridge structure damage identification method and system based on cross-channel attention, which can adaptively adjust the expansion rate and receptive field according to the input response changes, better explore the spatiotemporal correlation between multi-channel response data, and improve the accuracy of bridge structure damage identification.

[0004] The first technical solution adopted by the present invention is: a bridge structure damage identification method based on cross-channel attention, comprising the following steps: Based on vibration acceleration measurement, a sample data set of bridge structures is obtained; Introduce adaptive embedding mechanism module and cross-channel attention mechanism module to build a bridge structure damage recognition model; Based on the bridge structure damage identification model, damage identification is performed on the bridge structure sample data set to obtain the bridge structure damage identification result.

[0005] Furthermore, the step of obtaining a sample data set of a bridge structure based on vibration acceleration measurement specifically includes: Obtain the target bridge structure to build a finite element model and perform damage simulation to obtain the overall stiffness matrix of the target bridge structure after damage; Based on the overall stiffness matrix of the target bridge structure after damage, the response data of the target bridge structure is obtained by applying Gaussian white noise excitation and performing response calculation using the Newmark method; The vertical acceleration of each node is extracted from the response data of the target bridge structure to obtain a sample data set of the bridge structure.

[0006] Furthermore, the bridge structure damage identification model specifically includes a feature extraction layer, a feature compression layer and a classification layer, and the feature extraction layer, the feature compression layer and the classification layer are connected in sequence, wherein: The feature extraction layer includes an adaptive embedding mechanism module and a cross-channel attention mechanism module, the adaptive embedding mechanism module includes an overlapping block embedding layer and an adaptive expansion overlapping block embedding layer, the cross-channel attention mechanism module includes a plurality of cross-channel attention layers and a plurality of gated feedforward network layers, and the cross-channel attention layer includes a normalization layer, a first deep convolution layer, a second deep convolution layer, a third deep convolution layer, a first Softmax function normalization layer and a convolution layer; The feature compression layer includes a first maximum pooling layer, a second maximum pooling layer and a third maximum pooling layer; The classification layer includes a flattening layer, a first fully connected layer, a second fully connected layer and a second Softmax function normalization layer.

[0007] Furthermore, the step of performing damage identification on the bridge structure sample data set based on the bridge structure damage identification model to obtain the bridge structure damage identification result specifically includes: Inputting the bridge structure sample data set into the bridge structure damage identification model; Based on the feature extraction layer of the bridge structure damage identification model, feature extraction processing is performed on the bridge structure sample data set to obtain the bridge structure damage feature map; Based on the feature compression layer of the bridge structure damage identification model, feature compression processing is performed on the bridge structure damage feature map to obtain a compressed bridge structure damage feature map; Based on the classification layer of the bridge structure damage identification model, feature mapping and data classification processing are performed on the compressed bridge structure damage feature map to obtain the bridge structure damage identification result.

[0008] Furthermore, the feature extraction layer based on the bridge structure damage identification model performs feature extraction processing on the bridge structure sample data set to obtain the bridge structure damage feature map, which specifically includes: Inputting the bridge structure sample data set into the feature extraction layer of the bridge structure damage identification model; Based on the adaptive embedding mechanism module of the feature extraction layer, feature information extraction is performed on the bridge structure sample data set to obtain feature information of the bridge structure sample data; Based on the cross-channel attention mechanism module of the feature extraction layer, the feature information of the bridge structure sample data is subjected to attention feature calculation and nonlinear activation processing to obtain the damage feature map of the bridge structure.

[0009] Furthermore, the adaptive embedding mechanism module based on the feature extraction layer performs feature information extraction processing on the bridge structure sample data set to obtain feature information of the bridge structure sample data, which specifically includes: Input the bridge structure sample data set into the adaptive embedding mechanism module of the feature extraction layer; Based on the overlapping block embedding layer of the adaptive embedding mechanism module, local feature extraction is performed on the bridge structure sample data set to obtain local continuous information of the bridge structure sample data; Based on the adaptive expansion overlapping block embedding layer of the adaptive embedding mechanism module, the global feature extraction processing is performed on the bridge structure sample data set to obtain the global feature information of the bridge structure sample data; The local continuous information of the bridge structure sample data and the global characteristic information of the bridge structure sample data are combined to obtain the characteristic information of the bridge structure sample data.

[0010] Furthermore, the cross-channel attention mechanism module based on the feature extraction layer performs attention feature calculation and nonlinear activation processing on the feature information of the bridge structure sample data to obtain the step of the bridge structure damage feature map, which specifically includes: Input the feature information of the bridge structure sample data into the cross-channel attention mechanism module of the feature extraction layer; Based on the cross-channel attention layer of the cross-channel attention mechanism module, the cross-channel attention feature calculation is performed on the feature information of the bridge structure sample data to obtain a preliminary bridge structure damage feature map; Based on the gated feedforward network layer of the cross-channel attention mechanism module, the preliminary bridge structure damage feature map is processed by nonlinear mapping activation to obtain the bridge structure damage feature map.

[0011] Furthermore, the cross-channel attention layer based on the cross-channel attention mechanism module performs cross-channel attention feature calculation on the feature information of the bridge structure sample data to obtain a preliminary bridge structure damage feature map, which specifically includes: Input the characteristic information of the bridge structure sample data into the cross-channel attention layer of the cross-channel attention mechanism module; Based on the normalization layer of the cross-channel attention layer, the feature information of the bridge structure sample data is normalized to obtain the normalized feature information of the bridge structure sample; Based on the first deep convolution layer, the second deep convolution layer and the third deep convolution layer of the cross-channel attention layer, the normalized bridge structure sample feature information is respectively subjected to deep convolution calculation to obtain the query matrix of the bridge structure sample feature information, the key matrix of the bridge structure sample feature information and the value matrix of the bridge structure sample feature information; The key matrix of the bridge structure sample feature information is subjected to matrix transposition processing and combined with the query matrix of the bridge structure sample feature information, and input into the first Softmax function normalization layer, and the weight value of the bridge structure sample feature information is output; The value matrix of the bridge structure sample feature information is multiplied by the weight value of the bridge structure sample feature information and input into the convolution layer for convolution processing. The convolution result is added point by point to the feature information of the bridge structure sample data to obtain a preliminary bridge structure damage feature map.

[0012] The second technical solution adopted by the present invention is: a bridge structure damage identification system based on cross-channel attention, comprising: The first module is used to obtain a sample data set of a bridge structure based on vibration acceleration measurement; The second module is used to introduce the adaptive embedding mechanism module and the cross-channel attention mechanism module to build a bridge structure damage recognition model; The third module is used to perform damage identification on the bridge structure sample data set based on the bridge structure damage identification model to obtain the bridge structure damage identification result.

[0013] The beneficial effects of the method and system of the present invention are as follows: the present invention obtains a sample data set of a bridge structure based on vibration acceleration measurement, further introduces an adaptive embedding mechanism module and a cross-channel attention mechanism module, constructs a bridge structure damage identification model, performs damage identification on the sample data set of the bridge structure, and obtains a bridge structure damage identification result. The adaptive embedding mechanism module can adaptively adjust the expansion rate and the receptive field according to the input response change, and the cross-channel attention mechanism module can better mine the spatiotemporal correlation between multi-channel response data, and combine the gating mechanism to control the information flow, thereby improving the quality and stability of the output and improving the accuracy of bridge structure damage identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of steps of a bridge structure damage identification method based on cross-channel attention of the present invention; Figure 2 It is a structural block diagram of a bridge structure damage identification system based on cross-channel attention of the present invention; Figure 3 It is a schematic diagram of a bridge structure damage identification model (AECCA-former) provided in a specific embodiment of the present invention; Figure 4 is a schematic diagram of the comparison of receptive fields of different embedding mechanisms provided by a specific embodiment of the present invention; Figure 5 is a schematic diagram of a cross-channel attention mechanism module provided by a specific embodiment of the present invention; Figure 6It is a schematic diagram of a structural damage identification process based on AECCA-former provided by a specific embodiment of the present invention; Figure 7 is a schematic diagram of a finite element model of an elastic support beam provided in a specific embodiment of the present invention; Figure 8 It is a schematic diagram of two existing models provided by a specific embodiment of the present invention; Fig. 9 Schematic diagrams of training curves of three methods provided in specific embodiments of the present invention, wherein (a) represents a training set, and (b) represents a test set; Fig.10 is a schematic diagram of the specific accuracy of three methods provided by a specific embodiment of the present invention on a test set; Fig.11 1 is a schematic diagram of the specific accuracy of the three methods on the test set under different noise levels provided by the specific embodiment of the present invention, wherein (a) represents a noise level of 1%, (b) represents a noise level of 5%, and (c) represents a noise level of 10%; Fig.12 is a schematic diagram of a Z24 bridge sensor arrangement provided by a specific embodiment of the present invention, wherein (a) represents a plan view, and (b) represents a cross-sectional view; Fig.13 is a schematic diagram of the accuracy of recognition results of different working conditions provided by a specific embodiment of the present invention; Fig.14 Schematic diagram of confusion matrix of different sensor combination schemes in a test set provided by a specific embodiment of the present invention, wherein (a) represents the first combination, (b) represents the second combination, (c) represents the third combination, and (d) represents the fourth combination; Fig.15 It is a result schematic diagram of visualizing the intermediate layer results provided by a specific embodiment of the present invention. DETAILED DESCRIPTION

[0015] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0016] Reference Figure 1 The present invention provides a bridge structure damage identification method based on cross-channel attention, which comprises the following steps: S100, obtaining a sample data set of a bridge structure based on vibration acceleration measurement; Specifically, a finite element model of the target bridge structure is constructed and a damage simulation is performed to obtain the overall stiffness matrix of the target bridge structure after damage; based on the overall stiffness matrix of the target bridge structure after damage, the response data of the target bridge structure is obtained by applying Gaussian white noise excitation and performing response calculation through the Newmark method; the vertical acceleration of each node is extracted from the response data of the target bridge structure to obtain a sample data set of the bridge structure.

[0017] S200, introduce adaptive embedding mechanism module and cross-channel attention mechanism module to build a bridge structure damage recognition model; Specifically, Figure 3 As shown, the bridge structure damage identification model specifically includes a feature extraction layer, a feature compression layer and a classification layer, and the feature extraction layer, the feature compression layer and the classification layer are connected in sequence, wherein the feature extraction layer includes an adaptive embedding mechanism module and a cross-channel attention mechanism module, the adaptive embedding mechanism module includes an overlapping block embedding layer and an adaptive expanded overlapping block embedding layer, the cross-channel attention mechanism module includes several cross-channel attention layers and several gated feedforward network layers, the cross-channel attention layer includes a normalization layer, a first deep convolutional layer, a second deep convolutional layer, a third deep convolutional layer, a first Softmax function normalization layer and a convolutional layer; the feature compression layer includes a first maximum pooling layer, a second maximum pooling layer and a third maximum pooling layer; the classification layer includes a flattening layer, a first fully connected layer, a second fully connected layer and a second Softmax function normalization layer.

[0018] S300: Based on the bridge structure damage identification model, damage identification is performed on the bridge structure sample data set to obtain a bridge structure damage identification result.

[0019] Further, it should be noted that step S300 includes steps S310 to S340, wherein: S310, inputting the bridge structure sample data set into the bridge structure damage identification model; S320, based on the feature extraction layer of the bridge structure damage identification model, performing feature extraction processing on the bridge structure sample data set to obtain a bridge structure damage feature map; Furthermore, step S320 includes steps S321 to S323, wherein: S321, inputting the bridge structure sample data set into the feature extraction layer of the bridge structure damage identification model; S322, based on the adaptive embedding mechanism module of the feature extraction layer, extracting feature information from the bridge structure sample data set to obtain feature information of the bridge structure sample data; Specifically, the bridge structure sample data set is input into the adaptive embedding mechanism module of the feature extraction layer; based on the overlapping block embedding layer of the adaptive embedding mechanism module, local feature extraction processing is performed on the bridge structure sample data set to obtain local continuous information of the bridge structure sample data; based on the adaptive expansion overlapping block embedding layer of the adaptive embedding mechanism module, global feature extraction processing is performed on the bridge structure sample data set to obtain global feature information of the bridge structure sample data; the local continuous information of the bridge structure sample data and the global feature information of the bridge structure sample data are combined to obtain the feature information of the bridge structure sample data.

[0020] In this embodiment, the adaptive embedding mechanism realizes feature embedding by fusing the Overlap Patch Embedding layer (OPE) and the Adaptive Dilated Overlap Patch Embedding layer (ADOPE). OPE is a traditional embedding method that divides the response sequence into several overlapping segments by sliding the convolution kernel on the response sequence. This method helps the model capture local continuous information in the original sequence, but its block encoding method is fixed, and the receptive field of the convolution kernel is only within the convolution kernel size range, such as Figure 4 As shown in (a) in , it is easy to miss global information. The ADOPE method uses dilated convolution instead of ordinary convolution, as shown in Figure 4 As shown in (b) in the figure, the receptive field of the convolution kernel can be expanded without increasing the size of the convolution kernel, thereby improving the global feature capture performance of the embedding layer.

[0021] For the time course response of a channel, the mathematical expressions of conventional convolution and dilated convolution are as follows: ; In the above formula, and Represents the convolution before and after response data points. is the weight of the convolution kernel, Represents the size of the convolution kernel, is the dilation rate. By dilating the convolution, the convolution kernel weight parameter is not changed. In the case of In the embodiment of the present invention, the expansion rate of ADOPE is adaptively selected according to the input response sequence, and the specific implementation method is shown in Table 1.

[0022] Table 1 ADOPE method pseudo code

[0023] S323. Based on the cross-channel attention mechanism module of the feature extraction layer, attention feature calculation and nonlinear activation processing are performed on the feature information of the bridge structure sample data to obtain a bridge structure damage feature map.

[0024] Specifically, the feature information of the bridge structure sample data is input into the cross-channel attention mechanism module of the feature extraction layer; based on the cross-channel attention layer of the cross-channel attention mechanism module, the cross-channel attention features of the feature information of the bridge structure sample data are calculated to obtain a preliminary bridge structure damage feature map; based on the gated feedforward network layer of the cross-channel attention mechanism module, the preliminary bridge structure damage feature map is subjected to nonlinear mapping activation processing to obtain a bridge structure damage feature map.

[0025] Among them, the characteristic information of the bridge structure sample data is input into the cross-channel attention layer of the cross-channel attention mechanism module; Based on the normalization layer of the cross-channel attention layer, the feature information of the bridge structure sample data is normalized to obtain the normalized bridge structure sample feature information; based on the first deep convolution layer, the second deep convolution layer and the third deep convolution layer of the cross-channel attention layer, the normalized bridge structure sample feature information is respectively subjected to deep convolution calculation to obtain the query matrix of the bridge structure sample feature information, the key matrix of the bridge structure sample feature information and the value matrix of the bridge structure sample feature information; the key matrix of the bridge structure sample feature information is subjected to matrix transposition processing and combined with the query matrix of the bridge structure sample feature information, and input into the first Softmax function normalization layer, and the weight value of the bridge structure sample feature information is output; the value matrix of the bridge structure sample feature information is multiplied by the weight value of the bridge structure sample feature information and input into the convolution layer for convolution processing, and the convolution result is added point by point with the feature information of the bridge structure sample data to obtain a preliminary bridge structure damage feature map.

[0026] In this embodiment, the attention mechanism is the core of the Transformer neural network. The original multi-head self-attention mechanism emphasizes the similarity between samples, does not consider the correlation between different channels, and its time complexity is , Represents the signal length.

[0027] The embodiment of the present invention adopts a cross-channel attention mechanism (such as Figure 5 As shown in Figure 2, the original multi-head self-attention mechanism is replaced to effectively connect the signals on different channels and apply the attention mechanism along the channel dimension. In this method, the time complexity is ,in is the number of measurement channels. Usually, the number of measurement channels is much smaller than the length of the measurement signal. Therefore, the cross-channel attention mechanism can improve the computational efficiency. The specific implementation of the cross-channel attention mechanism is as follows: ; In the above formula, and are the input and output feature maps, respectively. It is a learnable parameter that adaptively adjusts the dot product amplitude during the training process. The Softmax function normalizes the attention weights.

[0028] The matrix is ​​passed through the batch normalized tensor Compared with the self-attention mechanism feature map , feature map of the cross-channel attention mechanism , with smaller spatial complexity. The feature map calculated by the cross-channel attention mechanism can emphasize the relationship between different channels while retaining the sensitivity to time series, which helps to enhance the accuracy and stability of the model in structural damage identification of multi-channel response data.

[0029] In addition, the method of the embodiment of the present invention further integrates the gating mechanism into the feedforward network, and designs another parallel path to use the GELU function for nonlinear activation to improve the output quality of the model. , the mathematical expression of this strategy is: ; In the above formula, represents element-wise multiplication, Represents the nonlinear activation of the GELU function, represents batch normalization, Represent the first, second, and third convolution layers respectively. They represent the second and third layers of depth-wise separable convolution respectively.

[0030] S330, based on the feature compression layer of the bridge structure damage identification model, performing feature compression processing on the bridge structure damage feature map to obtain a compressed bridge structure damage feature map; S340. Based on the classification layer of the bridge structure damage identification model, feature mapping and data classification processing are performed on the compressed bridge structure damage feature map to obtain a bridge structure damage identification result.

[0031] In this embodiment, the feature map extracted by the feature extraction layer is input into a compressor composed of a pooling layer to achieve further compression. After feature extraction and compression, the original data is mapped to the hidden layer feature space, and the expression of the pooling layer is: ; In the above formula, It represents the first The first The output value of the secondary pooling, In the feature map Channels, by position As the center, is the input value of the pooling window size, Indicates the range of pooling, from arrive The maximum value of the range.

[0032] The fully connected layer acts as a classification layer to map the learned distributed feature representation to the sample label space to achieve the classification of the input data. The expression of the fully connected layer is: ; In the above formula, Indicates the channel number, represents the length of the input sequence, represents the pooling window size, represents the step length, It is The weight matrix of the layer, It is The bias vector for the layer.

[0033] The detailed composition of the feature map compression and classification layer is shown in Table 2.

[0034] Table 2 Detailed composition of feature map compression and classification layers

[0035] In summary, if Figure 6As shown, the embodiment of the present invention proposes a new structural damage identification method (AECCA-former) that integrates adaptive embedding and cross-channel attention mechanism based on vibration acceleration measurement. The method proposes an adaptive embedding mechanism to improve the problem of limited receptive field of convolution kernel and improve the performance of feature index characterization; secondly, the cross-channel attention mechanism is introduced to make full use of the spatial correlation of multi-channel measurement and improve the damage sensitivity of feature index. The adaptive embedding mechanism uses overlapping block embedding (OPE) and adaptive dilated overlapping block embedding (ADOPE) for feature vector embedding at the same time, and can adaptively adjust the dilation rate and receptive field according to the input response change. In the cross-channel attention mechanism, the cross-covariance is used to calculate the attention feature map of the query matrix Q, the key matrix K and the value matrix V to better mine the spatiotemporal correlation between multi-channel response data; at the same time, the gating mechanism is combined to control the flow of information and improve the quality and stability of the output. The local damage of the bridge is identified, and artificial noise is added at the same time to study the recognition accuracy and noise robustness of the method.

[0036] Further, the embodiments of the present invention are described by specific experimental values: Use Figure 7 The 10-element finite element model with a total length of 25 meters is shown to verify the proposed method. The element type is a two-node four-degree-of-freedom plane bending beam element. The beam cross section is rectangular, the elastic modulus is 210 GPa, and the density is 7800 kg / m 3 , with an area of ​​0.48 m 2 , the moment of inertia is 0.0256 m 4 . Vertical spring stiffness , torsion spring stiffness The numbers in the box are the element numbers, and the other numbers are the node numbers. Use the Rayleigh damping of the following formula and and The values ​​of are set to To simulate the weak damping effect, the expression is: ; The detailed dynamic characteristics of the numerical model are shown in Table 3. When simulating damage, only the stiffness change caused by the damage is considered, and the overall stiffness matrix after damage is Expressed as ,in, and Respectively represent The lossless element stiffness matrix and damage factor of each element, , Indicates the number of units.

[0037] Table 3 Dynamic characteristics of the numerical model

[0038] The elastic support beam data set has five damage levels of 10%, 20%, 30%, 40% and 50% and a healthy state. The response calculation uses the Newmark method, the excitation mode is Gaussian white noise excitation, the duration is 3 s, the standard deviation is 200 N, and each working condition is repeated 15 times. The excitation is low-pass filtered using an eighth-order Butterworth filter with a cutoff frequency of 512 Hz to simulate the load in the real environment. The sampling time step is 2 -13 s, and downsampled to 1024 Hz, only the vertical acceleration of each node is measured, and the data size of each measurement is 9×3072. In summary, the data set has 11 classification categories: 10 damage locations and one healthy state. In order to maintain the balance of the data set, the healthy state is repeatedly calculated 5 times corresponding to the five damage degrees of each damage location. Finally, the size of the entire data set is 7425, that is, the number of measurement points is 9×the number of working conditions is 11×the number of damage degrees is 5×the number of repeated calculations is 15.

[0039] The original data set is normalized to a standard normal distribution with a mean of 0 and a standard deviation of 1 through Z-score standardization to eliminate the impact of different acceleration levels caused by different channels and different sampling times. Then, the data set is expanded by combining repeated sampling and subset sampling to increase the number of samples. The specific steps are: 1) Divide the dataset into training, validation, and test sets using a ratio of 0.7: 0.2: 0.1; increase their sample sizes to the nearest power of 2 by repeated sampling; 2) Randomly select a segment of length 1024 from each 3072 sample to increase the diversity of the sample, that is, extract a 1 s sub-signal from a 3 s signal.

[0040] In order to verify that the proposed method has higher damage identification accuracy, the proposed method is compared with the Multi-Scale Convolutional Neural Network (MCNN) and the Joint Convolutional Neural Network and Bidirectional Long Short-Term Memory Network (CNN-BiLSTM). The two methods differ from the proposed method only in the selection of the feature extraction layer. The specific flow chart is as follows Figure 8 shown.

[0041] In the embodiment of the present invention, cross entropy is used as the damage function to estimate the difference between the actual damage location and the predicted damage probability distribution. The Adam optimizer is used to update the gradient, and the learning rate is adjusted in a piecewise decay manner, decreasing by 10% every 50 epochs. The hyperparameter settings of the three models are consistent, and the specific settings are shown in Table 4.

[0042] Table 4 Hyperparameter settings

[0043] CNN-BiLSTM, MCNN, and AECCA-former used 200, 200, and 134 epochs to complete the training respectively. The corresponding training curves in the training set and the validation set are shown in the following figure. Fig. 9 As shown. As can be seen from the figure, the training curves of each model on the training set and the validation set are basically consistent, indicating that the model training process is reasonable and effective, and there is no overfitting phenomenon. Among them, AECCA-former has the fastest convergence speed, and CNN-BiLSTM has obvious oscillation in the early stage of training, and requires more iterations to achieve stability. The results in Table 5 show that AECCA-former has the highest recognition accuracy, reaching 98.83%, which fully proves that the method of the embodiment of the present invention has stronger feature extraction capabilities than the other two methods, and can achieve more accurate structural damage identification. More specifically, Fig.10 The accuracy of the three methods under different working conditions is demonstrated, which is calculated based on the percentage of the number of successful identifications of a certain damage location to the total number of identifications of the damage location. It can be seen that the three methods have the highest recognition accuracy for the healthy state. However, under different damage conditions, the recognition accuracy of the method of the embodiment of the present invention always remains above 95%. In contrast, the recognition accuracy of the other two methods under most damage conditions is lower than 95%, especially in the conditions for damage to units 2, 6, 8 and 9, the recognition accuracy has dropped significantly. Among them, the accuracy of CNN has dropped most significantly. In summary, the method of the embodiment of the present invention effectively improves the ability to extract damage-sensitive features by coupling the adaptive embedding mechanism with the cross-channel attention mechanism, and is superior to traditional methods in terms of training stability and recognition accuracy, showing better robustness and reliability.

[0044] Table 5 Accuracy of three methods on the test set

[0045] By formula Gaussian white noise is added to simulate the measurement uncertainty of the time history response, where: denote the acceleration with and without noise, respectively, is the maximum acceleration amplitude of each channel, represents the noise level, To obey Gaussian distributed random numbers. The embodiment of the present invention considers three different noise levels of 1%, 5% and 10%, and the recognition accuracy of each model on the noisy test set is shown in Table 6. It can be seen that as the noise level increases, the damage recognition accuracy of each model decreases to varying degrees. However, compared with other methods, AECCA-former is least affected by noise. At a high noise level of 10%, the recognition accuracy of this method on the test set only decreases by about 8% compared to the noise-free case, and it can still maintain a high recognition accuracy. In addition, the recognition accuracy of the three methods for different damage conditions is shown in the following table. Fig.11 As shown. It can be further seen that, except for the recognition accuracy of unit 1 and unit 2 which is slightly lower than 90%, the recognition accuracy of AECCA-former in other damage conditions exceeds 90%, showing good noise robustness. In contrast, under the influence of noise, the recognition accuracy of all damage conditions of the other two methods has dropped significantly, especially that of CNN. The recognition accuracy of CNN for most conditions is lower than 80%. This shows that the method of the embodiment of the present invention combines the adaptive embedding mechanism and the cross-channel attention mechanism, which can improve the noise robustness of the method and has a higher recognition accuracy for damage conditions of different positions and degrees.

[0046] Table 6 Accuracy of the three methods on the test set under different noise levels

[0047] Further elaboration combined with Z24 engineering application, Z24 bridge ( Fig.12 ) is a three-span prestressed concrete continuous box girder bridge with a main span of 30 meters and a side span of 14 meters. Located in Switzerland, the bridge was built in 1963. Before it was dismantled and rebuilt in 1998, it was monitored for about 10 months and subjected to progressive damage tests in a controlled manner, with a variety of health monitoring conditions constructed. The monitoring data of the bridge is often used to test the application effect of structural damage identification methods in real bridges.

[0048] Table 7 Damage conditions

[0049] The time history response of the Z24 bridge is collected by 4 accelerometers arranged in the vertical direction and 4 accelerometers arranged in the non-vertical direction. The sampling frequency is 100 Hz, and the sampling duration of each acceleration is the first 11 minutes of each hour. A total of 65,536 data points are collected. The 60,000 data points before 1 a.m. are intercepted to construct the data set, and the data is upsampled to 1024 Hz through cubic spline interpolation to keep the size of the input data consistent with the numerical simulation. The 10 damage conditions included in the data set are shown in Table 7, where condition 0 represents the healthy reference state. Finally, the size of the generated data set is , where 2000 represents 2000 pieces of data, 8 represents 8 channels, and 3072 represents 3072 data points.

[0050] All sensor data sets are divided into training set, validation set and test set, and the damage recognition accuracy of the method of the embodiment of the present invention for different working conditions of the test set is obtained as follows: Fig.13 As shown. It can be seen that among the 10 working conditions, the recognition accuracy of AECCA-former for 7 of them reached 100%, indicating that this method can effectively distinguish different causes of structural state changes and is feasible in actual engineering. Among them, working conditions 1 to 3 and working condition 5 are structural damage, which usually cause obvious changes in bridge response. AECCA-former can accurately capture this change. Compared with working conditions 6 and 8, the degree of damage to the bridge structure in working conditions 7 and 9 gradually accumulates, which is more obviously reflected in the bridge response, so they can be effectively identified. In contrast, the recognition accuracy of working conditions 4, 6 and 8 is slightly worse, but it can also be maintained at about 95%. Among them, the abutment landslide in working condition 4 is caused by heavy rain erosion, which has a more obvious impact on horizontal vibration and a smaller impact on bending stiffness. This may be the reason for the relatively low recognition accuracy of this working condition. Working conditions 6 and 8 have a greater impact on the structural bearing capacity, but a smaller impact on the overall stiffness of the structure. In addition, the number of anchor head failures and prestressed tendon ruptures is small, and the impact on the bridge response in a short period of time is small. This may be the reason for the relatively low identification accuracy of these two working conditions.

[0051] In order to examine the impact of different numbers and arrangements of sensors on the method of the embodiment of the present invention, four different sensor combinations are selected based on existing measurements as shown in Table 8.

[0052] The accuracy of the four sensor combination schemes on the test set were 98.44%, 91.8%, 90.62% and 94.92% respectively. It can be seen that when the number of sensors decreases, the recognition accuracy of the method of the embodiment of the present invention for different damage conditions decreases, but it can still maintain more than 90%. In addition, the combination of sensors in different positions also affects the recognition accuracy. In the three tests with 4 sensors as a combination, the scheme using only half of the sensors on the left side has the highest recognition accuracy. This is because the combination scheme has a more balanced sensor layout and includes sensors in both vertical and non-vertical directions. The measured response can provide more damage information. This also shows that the accuracy of structural damage recognition depends not only on the number of sensors, but also on the location of the sensor layout.

[0053] Table 8 Different sensor combinations

[0054] The specific recognition accuracy of different working conditions is as follows: Fig.14 The confusion matrix is ​​shown in Figure 1. The ordinate represents the true label and the abscissa represents the predicted label. The more obvious the diagonal nature of the confusion matrix is, the more accurate the recognition result is. It should be noted that the number of test samples for each working condition in the confusion matrix is ​​not consistent. This is because the test set data is randomly sampled and contains all working conditions, but the number of samples for each working condition is not balanced. Fig.14 It can be seen that when only vertical sensors are arranged, there are obvious non-zero values ​​in the non-diagonal positions of the confusion matrix in the rows of working conditions 4 and 8, indicating that these working conditions are more obviously misjudged. This is because the landslide of the bridge abutment in working condition 4 has a more obvious impact on the horizontal vibration, and the vertical sensor is difficult to capture the characteristics of this horizontal vibration. The poor recognition of working condition 8 is related to the small change in response caused by this factor. In contrast, when non-vertical sensors are arranged, the model misjudged multiple labels and performed the worst. In addition, except for combination 3, other combination schemes did not misjudge the damaged working condition as a healthy state. This shows that vertical sensors have higher efficiency in capturing structural information and can significantly improve the accuracy of damage identification. Therefore, by comprehensively considering the number, layout location and direction of sensors, the structural damage identification scheme can be better optimized to improve the accuracy and robustness of damage identification.

[0055] In order to further visualize the intermediate process of feature extraction, compression and classification, the embodiment of the present invention uses the UMAP visualization method to display the internal working mechanism of the proposed AECCA-former, where: Fig.15 (a) in the figure represents the original data. Fig.15 (b) in the figure represents the feature extraction layer. Fig.15 (c) in the figure represents the feature compression layer. Fig.15 (d) in the figure represents the feature classification layer. Fig.15 As shown in (a) in the figure, in the input layer, data points of different categories are mixed together, making it difficult to distinguish the structural state. As feature extraction proceeds, the model effectively extracts key features from the input data through multi-scale embedding and cross-channel attention mechanisms. Fig.15 (b) to Fig.15 In (c), samples of different categories begin to separate gradually. As the extracted features are further compressed and classified, the distance between points of different categories increases, while points of the same category gradually get closer. Finally, the model's results at the classification layer ( Fig.15 (d) in the figure shows that the boundaries between different categories are very clear, and samples of the same category are highly concentrated in the center of their respective categories.

[0056] In summary, the embodiment of the present invention uses the Transformer neural network as the core framework, and adopts an adaptive embedding mechanism of fused overlapping block embedding and expanded overlapping block embedding in the feature extraction layer to achieve adaptive embedding of feature vectors, thereby improving the flexibility of block coding to cope with the needs of structural damage identification under different working conditions. In addition, a cross-channel attention mechanism is used in the Transformer encoder for feature extraction, which enhances the feature extraction capability of the proposed method on multi-channel measurement responses. Numerical simulations of elastically supported beam bridges show that AECCA-former has better feature extraction performance and can provide more reliable, accurate and robust structural damage identification results under complex damage conditions. The practical results of the Z24 Bridge project show that AECCA-former can effectively identify real bridge damage, and a small number of sensors has little effect on the accuracy of structural damage identification.

[0057] Therefore, the embodiment of the present invention proposes a new bridge structure damage identification method that integrates adaptive embedding and cross-channel attention mechanism, namely AECCA-former. This method combines OPE and ADOPE in the feature extraction layer to perform multi-scale adaptive embedding of input responses. At the same time, the cross-channel attention mechanism and gating mechanism are integrated to improve the algorithm's ability to process multi-channel responses. Compared with the prior art, the embodiment of the present invention has the following advantages: 1) The adaptive embedding mechanism and cross-channel attention mechanism effectively improve the problem of limited receptive field of convolution kernel in traditional methods. At the same time, they can adaptively adjust the expansion rate according to the response input to meet the needs of structural damage identification under different working conditions. Compared with traditional convolution methods, the proposed method shows higher recognition accuracy and noise robustness.

[0058] 2) Under different actual bridge working conditions, the proposed method has a recognition accuracy slightly lower than 95% in only two of the working conditions, showing good classification performance for different causes of structural damage. At the same time, when reducing sensor measurements, the proposed method can still provide a recognition accuracy of more than 90%, which is expected to be applied in actual engineering.

[0059] 3) The visualization results of the feature extraction, compression and classification layers show that the proposed method can achieve a clear distinction between different damage categories. At the same time, the clustering effect of damage categories is obvious, indicating that the features extracted from the acceleration response have good damage sensitivity.

[0060] Reference Figure 2 , a bridge structure damage identification system based on cross-channel attention, including: The first module 201 is used to obtain a sample data set of a bridge structure based on vibration acceleration measurement; The second module 202 is used to introduce an adaptive embedding mechanism module and a cross-channel attention mechanism module to build a bridge structure damage recognition model; The third module 203 is used to perform damage identification on the bridge structure sample data set based on the bridge structure damage identification model to obtain a bridge structure damage identification result.

[0061] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0062] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A bridge structure damage identification method based on cross-channel attention, characterized in that: The following steps are involved: Based on vibration acceleration measurement, a sample data set of bridge structures is obtained; Introduce adaptive embedding mechanism module and cross-channel attention mechanism module to build a bridge structure damage recognition model; The bridge structure damage identification model specifically includes a feature extraction layer, a feature compression layer and a classification layer, and the feature extraction layer, the feature compression layer and the classification layer are connected in sequence, wherein: The feature extraction layer includes an adaptive embedding mechanism module and a cross-channel attention mechanism module, the adaptive embedding mechanism module includes an overlapping block embedding layer and an adaptive expansion overlapping block embedding layer, the cross-channel attention mechanism module includes a plurality of cross-channel attention layers and a plurality of gated feedforward network layers, and the cross-channel attention layer includes a normalization layer, a first deep convolution layer, a second deep convolution layer, a third deep convolution layer, a first Softmax function normalization layer and a convolution layer; The feature compression layer includes a first maximum pooling layer, a second maximum pooling layer and a third maximum pooling layer; The classification layer includes a flattening layer, a first fully connected layer, a second fully connected layer and a second Softmax function normalization layer; Based on the bridge structure damage identification model, damage identification is performed on the bridge structure sample data set to obtain the bridge structure damage identification result.

2. According to claim 1, a bridge structure damage identification method based on cross-channel attention is characterized in that: The step of obtaining a sample data set of a bridge structure based on vibration acceleration measurement specifically includes: Obtain the target bridge structure to build a finite element model and perform damage simulation to obtain the overall stiffness matrix of the target bridge structure after damage; Based on the overall stiffness matrix of the target bridge structure after damage, the response data of the target bridge structure is obtained by applying Gaussian white noise excitation and performing response calculation using the Newmark method; The vertical acceleration of each node is extracted from the response data of the target bridge structure to obtain a sample data set of the bridge structure.

3. According to claim 2, a bridge structure damage identification method based on cross-channel attention is characterized in that: The step of performing damage identification on the bridge structure sample data set based on the bridge structure damage identification model to obtain the bridge structure damage identification result specifically includes: Inputting the bridge structure sample data set into the bridge structure damage identification model; Based on the feature extraction layer of the bridge structure damage identification model, feature extraction processing is performed on the bridge structure sample data set to obtain the bridge structure damage feature map; Based on the feature compression layer of the bridge structure damage identification model, feature compression processing is performed on the bridge structure damage feature map to obtain a compressed bridge structure damage feature map; Based on the classification layer of the bridge structure damage identification model, feature mapping and data classification processing are performed on the compressed bridge structure damage feature map to obtain the bridge structure damage identification result.

4. According to claim 3, a bridge structure damage identification method based on cross-channel attention is characterized in that: The feature extraction layer based on the bridge structure damage identification model performs feature extraction processing on the bridge structure sample data set to obtain the bridge structure damage feature map. This step specifically includes: Inputting the bridge structure sample data set into the feature extraction layer of the bridge structure damage identification model; Based on the adaptive embedding mechanism module of the feature extraction layer, feature information extraction is performed on the bridge structure sample data set to obtain feature information of the bridge structure sample data; Based on the cross-channel attention mechanism module of the feature extraction layer, the feature information of the bridge structure sample data is subjected to attention feature calculation and nonlinear activation processing to obtain the damage feature map of the bridge structure.

5. According to claim 4, a bridge structure damage identification method based on cross-channel attention is characterized in that: The step of extracting feature information from the bridge structure sample data set by the adaptive embedding mechanism module based on the feature extraction layer to obtain feature information of the bridge structure sample data specifically includes: Input the bridge structure sample data set into the adaptive embedding mechanism module of the feature extraction layer; Based on the overlapping block embedding layer of the adaptive embedding mechanism module, local feature extraction is performed on the bridge structure sample data set to obtain local continuous information of the bridge structure sample data; Based on the adaptive expansion overlapping block embedding layer of the adaptive embedding mechanism module, the global feature extraction processing is performed on the bridge structure sample data set to obtain the global feature information of the bridge structure sample data; The local continuous information of the bridge structure sample data and the global characteristic information of the bridge structure sample data are combined to obtain the characteristic information of the bridge structure sample data.

6. According to claim 5, a bridge structure damage identification method based on cross-channel attention is characterized in that: The cross-channel attention mechanism module based on the feature extraction layer performs attention feature calculation and nonlinear activation processing on the feature information of the bridge structure sample data to obtain the bridge structure damage feature map. This step specifically includes: Input the feature information of the bridge structure sample data into the cross-channel attention mechanism module of the feature extraction layer; Based on the cross-channel attention layer of the cross-channel attention mechanism module, the cross-channel attention feature calculation is performed on the feature information of the bridge structure sample data to obtain a preliminary bridge structure damage feature map; Based on the gated feedforward network layer of the cross-channel attention mechanism module, the preliminary bridge structure damage feature map is processed by nonlinear mapping activation to obtain the bridge structure damage feature map.

7. According to claim 6, a bridge structure damage identification method based on cross-channel attention is characterized in that: The cross-channel attention layer based on the cross-channel attention mechanism module performs cross-channel attention feature calculation on the feature information of the bridge structure sample data to obtain a preliminary bridge structure damage feature map. Specifically, the step includes: Input the characteristic information of the bridge structure sample data into the cross-channel attention layer of the cross-channel attention mechanism module; Based on the normalization layer of the cross-channel attention layer, the feature information of the bridge structure sample data is normalized to obtain the normalized feature information of the bridge structure sample; Based on the first deep convolution layer, the second deep convolution layer and the third deep convolution layer of the cross-channel attention layer, the normalized bridge structure sample feature information is respectively subjected to deep convolution calculation to obtain the query matrix of the bridge structure sample feature information, the key matrix of the bridge structure sample feature information and the value matrix of the bridge structure sample feature information; The key matrix of the bridge structure sample feature information is subjected to matrix transposition processing and combined with the query matrix of the bridge structure sample feature information, and input into the first Softmax function normalization layer, and the weight value of the bridge structure sample feature information is output; The value matrix of the bridge structure sample feature information is multiplied by the weight value of the bridge structure sample feature information and input into the convolution layer for convolution processing. The convolution result is added point by point to the feature information of the bridge structure sample data to obtain a preliminary bridge structure damage feature map.

8. A bridge structure damage identification system based on cross-channel attention, characterized in that: Includes the following modules: The first module is used to obtain a sample data set of a bridge structure based on vibration acceleration measurement; The second module is used to introduce the adaptive embedding mechanism module and the cross-channel attention mechanism module to build a bridge structure damage recognition model; The bridge structure damage identification model specifically includes a feature extraction layer, a feature compression layer and a classification layer, and the feature extraction layer, the feature compression layer and the classification layer are connected in sequence, wherein: The feature extraction layer includes an adaptive embedding mechanism module and a cross-channel attention mechanism module, the adaptive embedding mechanism module includes an overlapping block embedding layer and an adaptive expansion overlapping block embedding layer, the cross-channel attention mechanism module includes a plurality of cross-channel attention layers and a plurality of gated feedforward network layers, and the cross-channel attention layer includes a normalization layer, a first deep convolution layer, a second deep convolution layer, a third deep convolution layer, a first Softmax function normalization layer and a convolution layer; The feature compression layer includes a first maximum pooling layer, a second maximum pooling layer and a third maximum pooling layer; The classification layer includes a flattening layer, a first fully connected layer, a second fully connected layer and a second Softmax function normalization layer; The third module is used to perform damage identification on the bridge structure sample data set based on the bridge structure damage identification model to obtain the bridge structure damage identification result.

Images