A method and system for identifying bridge flood risk locations

Abstract

This invention discloses a method and system for identifying bridge flood risk locations. The method includes: selecting feature extraction locations on the bridge as nodes; extracting the equivalent stress and total deformation values ​​of each node under different water depth and flow velocity combinations to construct a dataset; constructing a LAG-Block module, which is used to capture local detailed features in the spatial domain through convolution and establish global dependencies in the feature domain through an attention mechanism; sequentially connecting the convolution module, LAG-Block module, first channel attention module, first global average pooling, and classifier to construct a risk identification model; training the risk identification model; and using the trained risk identification model to identify bridge flood risk locations. The advantage of this invention is that it achieves bridge flood risk location identification.

Description

Technical Field

[0001] This invention relates to the field of deep learning, specifically to a method and system for identifying the location of bridge flood risks. Background Technology

[0002] In recent years, against the backdrop of global climate change and increasing watershed development, extreme rainfall and flash floods have become more frequent, rapid, and prolonged, posing a more severe challenge to the service safety of bridges spanning rivers under flood conditions. Numerous engineering examples demonstrate that flood-induced scouring, debris impacts, and asymmetric water flow loads often lead to bridge foundation instability, superstructure damage, and even complete bridge collapse, causing not only severe traffic disruptions and economic losses but also posing a significant threat to the lives and property of residents along the affected areas. Accurately identifying high-risk locations of bridge structures before or during floods, and thus providing support for flood control reinforcement, pre-disaster planning, and disaster early warning, has become a critical scientific and engineering problem that urgently needs to be solved in the field of bridge disaster prevention and mitigation.

[0003] Traditional research on the effects of flooding on bridges has largely relied on physical model experiments and numerical simulations based on RANS equations to systematically analyze local pier scour, flood bypass, and hydrodynamic responses. In recent years, high-precision numerical methods such as fine-macro data interaction and multi-field coupling have been introduced into the study of bridge flood damage mechanisms, providing relatively detailed flow velocity fields, fluid pressure fields, and scour evolution processes at the structural scale. However, these methods typically focus on the response assessment of a few typical sections or components, and still rely on empirical interpretation for characterizing the spatial distribution of flood risk at the "whole-bridge scale," especially for the automatic identification of risk locations. Meanwhile, visual detection technologies for the appearance damage of hydraulic and civil structures are developing rapidly. For apparent cracks in dams and bridges, the paper "Zhou Xun, Li Yonglong, Zhou Yingyue, et al. A method for dam surface crack detection based on an improved DeepLabV3+ network [J]. Journal of Tsinghua University (Natural Science Edition), 2023, 63(7): 1153-1163." proposes a dam surface crack segmentation method based on an improved DeepLabV3+ network, which has achieved good results in complex backgrounds, weak cracks, and multi-scale target recognition. The paper "Cha YJ, Choi W, Büyüköztürk O. Deep learning-based crack damage detection using convolutional neural networks [J]. Computer-Aided Civil and Infrastructure Engineering, 2017, 32(5): 361-378." applies convolutional neural networks (CNN) and fully convolutional networks (FCN) to the automatic identification and segmentation of cracks in concrete and steel structures, and gradually extends it to vibration response-driven damage identification and state classification. However, most existing studies focus on identifying apparent defects that have already occurred, with inputs often being static images or monitoring signals. The goal is to determine the location and scale of cracks. There is a lack of risk identification methods for key bridge nodes under flood conditions, which to some extent limits the application depth of deep learning methods in bridge flood control design and operational risk management. Summary of the Invention

[0004] The technical problem to be solved by this invention is how to identify the risks of key bridge nodes under flood conditions.

[0005] This invention solves the above-mentioned technical problems through the following technical means: a method for identifying the location of bridge flood risk, comprising the following steps:

[0006] S1. Select feature extraction locations on the bridge as nodes, extract the equivalent stress and total deformation values ​​of each node under different water depth and flow velocity combinations, and construct a dataset. S2. Construct the LAG-Block module. The LAG-Block module is used to capture local detailed features in the spatial domain through convolution and to establish global dependencies in the feature domain through attention mechanism. The convolution module, LAG-Block module, first channel attention module, first global average pooling and classifier are sequentially connected to form the risk recognition model. S3. Train the risk identification model; S4. Use the trained risk identification model to identify the location of bridge flood risk.

[0007] Furthermore, the working process of the LAG-Block module is as follows: The input feature matrix is ​​represented as an input feature tensor. The input feature tensor undergoes dimensionality transformation to obtain dimensionality-transformed input features. The dimensionality-transformed input features are then convolved with depthwise separable convolutions to obtain local features. These local features are mapped through a first convolutional layer to obtain query vectors, key vectors, and value vectors. Attention weights are calculated using the query vectors and key vectors. The global features are obtained by multiplying the attention weights by the value vectors. The global features are then linearly projected to restore the channel dimensions, resulting in a projected global feature map. The projected global feature map is then compared with the local features obtained through linear projection. The results of the enhancement function are summed to obtain the enhanced feature map. The enhanced feature map is then processed by the second convolutional layer and then by the first activation function to obtain the gated weight tensor. The gated weight tensor is multiplied by the enhanced feature map to obtain the gated feature map. The gated feature map is input into the second channel attention module to obtain the channel-weighted feature map. The channel-weighted feature map is then passed through a multilayer perceptron and a global response normalization layer to obtain the multilayer perceptron output feature. The multilayer perceptron output feature is added to the channel-weighted feature map and then subjected to a dimension reduction operation to obtain a feature map with the same shape as the input, which is used as the final output feature of the LAG-Block module.

[0008] Furthermore, the first channel attention module and the second channel attention module have the same structure, and the processing procedure of the second channel attention module is as follows: The gated feature map is subjected to a second global average pooling to obtain the channel description vector. The channel description vector is then processed by a third convolutional layer and a second activation function, and then by a fourth convolutional layer and a third activation function to obtain the channel attention weight vector. The channel attention weight vector is multiplied pixel-by-pixel with the gated feature map to obtain the channel-weighted feature map.

[0009] Furthermore, the convolution module and the first to fourth convolutional layers are all 1×1 convolutions.

[0010] Furthermore, the classifier consists of a fifth convolutional layer, a fourth activation function, a Dropout operation, and a sixth convolutional layer connected in sequence.

[0011] Furthermore, the fifth and sixth convolutional layers are 1×1 convolutions.

[0012] Further, S3 includes: The parameters of the risk identification model are continuously adjusted, and the value of the overall optimization objective is calculated. Training stops when the value of the overall optimization objective is minimized, resulting in a well-trained risk identification model. The formula for calculating the overall optimization objective is as follows:

[0013] in, The objective is to optimize the parameters; The model loss function; The total number of training samples; Total number of categories; For the first The sample at the th The smoothed true label on the class; For the first Each sample is predicted as follows The probability of a class; Represents the regularization coefficient. The parameters represent the risk identification model. Denotes the square of the L2 norm. This is an L2 regularization term.

[0014] The present invention also provides a bridge flood risk location identification system, comprising: The dataset construction module is used to select feature extraction locations on the bridge as nodes, extract the equivalent stress and total deformation values ​​of each node under different water depth and flow velocity combinations, and construct the dataset. The model building module is used to construct the LAG-Block module. The LAG-Block module is used to capture local detailed features in the spatial domain through convolution and to establish global dependencies in the feature domain through an attention mechanism. The convolution module, LAG-Block module, first channel attention module, first global average pooling and classifier are sequentially connected to form the risk recognition model. The model training module is used to train the risk identification model; The risk identification module is used to identify the location of bridge flood risks using a trained risk identification model.

[0015] Furthermore, the working process of the LAG-Block module is as follows: The input feature matrix is ​​represented as an input feature tensor. The input feature tensor undergoes dimensionality transformation to obtain dimensionality-transformed input features. The dimensionality-transformed input features are then convolved with depthwise separable convolutions to obtain local features. These local features are mapped through a first convolutional layer to obtain query vectors, key vectors, and value vectors. Attention weights are calculated using the query vectors and key vectors. The global features are obtained by multiplying the attention weights by the value vectors. The global features are then linearly projected to restore the channel dimensions, resulting in a projected global feature map. The projected global feature map is then compared with the local features obtained through linear projection. The results of the enhancement function are summed to obtain the enhanced feature map. The enhanced feature map is then processed by the second convolutional layer and then by the first activation function to obtain the gated weight tensor. The gated weight tensor is multiplied by the enhanced feature map to obtain the gated feature map. The gated feature map is input into the second channel attention module to obtain the channel-weighted feature map. The channel-weighted feature map is then passed through a multilayer perceptron and a global response normalization layer to obtain the multilayer perceptron output feature. The multilayer perceptron output feature is added to the channel-weighted feature map and then subjected to a dimension reduction operation to obtain a feature map with the same shape as the input, which is used as the final output feature of the LAG-Block module.

[0016] Furthermore, the first channel attention module and the second channel attention module have the same structure, and the processing procedure of the second channel attention module is as follows: The gated feature map is subjected to a second global average pooling to obtain the channel description vector. The channel description vector is then processed by a third convolutional layer and a second activation function, and then by a fourth convolutional layer and a third activation function to obtain the channel attention weight vector. The channel attention weight vector is multiplied pixel-by-pixel with the gated feature map to obtain the channel-weighted feature map.

[0017] Furthermore, the convolution module and the first to fourth convolutional layers are all 1×1 convolutions.

[0018] Furthermore, the classifier consists of a fifth convolutional layer, a fourth activation function, a Dropout operation, and a sixth convolutional layer connected in sequence.

[0019] Furthermore, the fifth and sixth convolutional layers are 1×1 convolutions.

[0020] Furthermore, the model training module is also used for: The parameters of the risk identification model are continuously adjusted, and the value of the overall optimization objective is calculated. Training stops when the value of the overall optimization objective is minimized, resulting in a well-trained risk identification model. The formula for calculating the overall optimization objective is as follows:

[0021] in, The objective is to optimize the parameters; The model loss function; The total number of training samples; Total number of categories; For the first The sample at the th The smoothed true label on the class; For the first Each sample is predicted as follows The probability of a class; Represents the regularization coefficient. The parameters represent the risk identification model. Denotes the square of the L2 norm. This is an L2 regularization term.

[0022] The advantages of this invention are: (1) This invention proposes a risk identification model that integrates local convolution, multi-head attention and channel gating mechanisms, realizing deep expression of flood risk features and node-level location identification. The model structure can effectively capture local salient features and integrate cross-node global information, providing a more expressive feature framework for risk identification under complex working conditions, and realizing risk identification of key bridge nodes under flood conditions.

[0023] (2) This invention combines the "physical field information" such as the three-dimensional hydrodynamic field and scour field (equivalent stress and total deformation value) obtained from flood numerical simulation with a deep neural network. From the perspective of "damage not yet or just occurring", it automatically identifies the potential high-risk locations of bridges under different flood conditions, providing a supplement to traditional assessment methods based on experience or monitoring signals. This technical approach provides a widely applicable foundation for bridge flood safety assessment, risk warning threshold setting, and intelligent auxiliary decision-making.

[0024] (3) The LAG-Block module of the present invention captures local detail features through convolution in the spatial domain, establishes global dependencies through attention mechanism in the feature domain, and adaptively adjusts feature flow through channel attention and gating mechanism, thereby achieving synergistic enhancement of local and global information.

[0025] (4) The LAG-Block module of this invention has collaborative modeling capabilities. Specifically, traditional CNNs are good at capturing local stress gradients, while Transformer architectures are good at capturing long-range dependencies. The LAG-Block module solves the problem of modeling the coupling relationship between local stress anomalies and overall bridge stiffness distribution under flood conditions by fusing depthwise separable convolution (local) with multi-head attention (global). It can simultaneously identify the local scour response of the piers and the overall stress transmission law between nodes. In addition, the LAG-Block module of this invention has the active filtering capability of the gating mechanism. Specifically, under complex flood conditions, the input data contains a large number of redundant or low-contribution features. The gating mechanism is equivalent to a physical saliency filter, which can dynamically adjust the ratio of local and global feature flows, automatically suppress interference caused by numerical simulation noise or minor physical quantities, and enable the model to accurately lock the feature dimension with the highest risk contribution. Secondly, the LAG-Block module of this invention also has spatial topology awareness capabilities. Specifically, the introduced convolutional position bias compensates for the lack of spatial geometric relationship awareness in traditional self-attention mechanisms. This enables the model to learn the relative positional logic of bridge nodes in three-dimensional space (such as the relative topology of supports and main beams, piers and water-facing surfaces), thereby avoiding risky position misjudgments caused by the loss of coordinate information.

[0026] (5) The risk identification model of this invention uses four types of features of nodes as input: water depth, flow velocity, equivalent stress, and total deformation. These features are derived from multi-condition numerical simulations and can reflect the basic stress state of nodes under different flood conditions. To make the input more suitable for the high-dimensional feature space of the subsequent network, the model completes channel mapping through 1×1 convolution, which avoids the influence of differences in physical quantity scales while ensuring the consistency of feature dimensions. The mapped tensor enters the feature extraction backbone composed of LAG-Block modules. The LAG-Block module combines local convolution, multi-head attention, position bias, and channel gating mechanisms in the same structure, enabling the network to simultaneously acquire local response patterns and global correlation features across nodes. Depthwise separable convolution focuses on capturing local stress-displacement changes, multi-head attention is used to characterize long-range dependencies between nodes, while position bias and channel attention further enhance the spatial sensitivity and channel selection ability of feature expression, thereby obtaining more discriminative node representations. After completing the backbone feature extraction, the network weights the fused features through the first channel attention module to evaluate the contribution of each channel to the overall response. This step can suppress interference from local noise, allowing the model to focus more on structural features in the final classification stage and achieve more stable risk location identification.

[0027] (6) The classifier of this invention consists of 1×1 convolution, GELU activation, and Dropout, and maps the features to a specific class space at the end. The introduction of GELU makes the classifier more flexible in response to high-dimensional feature changes, while Dropout improves the model's generalization ability on multi-condition data. Through this structure, the risk identification model transforms features into risk location classification results, realizing a complete mapping process from simulated data to risk labels. Attached Figure Description

[0028] Figure 1 This is a local pressure cloud map of a bridge in a bridge flood risk location identification method disclosed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the LAG-Block module of a bridge flood risk location identification method disclosed in an embodiment of the present invention. Figure 2 (a) shows the calculation flowchart of the LAG-Block module. Figure 2 (b) is a schematic diagram of the SE module; Figure 3 This is a schematic diagram of the risk identification model structure in a bridge flood risk location identification method disclosed in an embodiment of the present invention; Figure 4(a) is a comparative experimental bar chart of a bridge flood risk location identification method disclosed in the embodiment of the present invention, and Figure 4(b) is an ablation experimental bar chart. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Example 1 Embodiment 1 of the present invention provides a method for identifying the location of bridge flood risk, comprising the following steps: S1, Dataset Construction like Figure 1 The image shows a local pressure cloud map of the bridge. Based on the equivalent stress and deformation distribution simulation analysis under the conditions of a water depth of 4.0 m and a flow velocity of 8 m / s, it can be seen that the main stress and deformation of the bridge are concentrated in the bearings, piers, mid-span of the main girder, and the upstream area of ​​the bridge structure. Based on this, 10 nodes, including 4 bearing nodes, 1 mid-span node of the main girder, 1 upstream node, and 4 pier nodes, were selected as feature extraction locations to represent the main stress characteristics of the bridge at both global and local scales, reflecting the differences in the response of different structural components to flood loads.

[0031] At each node location, simulation results were extracted for different combinations of water depths (1.0 m, 1.5 m, 2.0 m, 2.5 m, 4.0 m) and flow velocities (1 m / s, 2 m / s, 4 m / s, 6 m / s, 8 m / s) to obtain the equivalent stress and total deformation of the node. Considering both fluid interaction characteristics and structural response parameters, the input feature matrix was defined as follows:

[0032] in, Water depth (m); The velocity is (m / s). The nodal equivalent stress (MPa); The total deformation of the node is expressed in mm.

[0033] S2, Risk Identification Model Construction To achieve efficient extraction and discrimination of bridge features under flood conditions, this invention proposes the LAG-Block (Local-Attention-Gated Block) module. In the spatial domain, it captures local detail features through convolution; in the feature domain, it establishes global dependencies through an attention mechanism; and it adaptively adjusts the feature flow through channel attention and gating mechanisms, thereby achieving synergistic enhancement of local and global information. Figure 2 As shown, Figure 2 (a) shows the calculation flowchart of the LAG-Block module. Figure 2 (b) is a schematic diagram of the SE module (channel attention module).

[0034] Input feature matrix Represented as input feature tensor: (1) in: The input feature tensor; Indicates characteristics; Batch size; Number of channels; The height and width of the feature map.

[0035] To facilitate feature relationship modeling at the node level, the input feature tensor undergoes dimensionality transformation: (2) in: The input features are after dimensionality transformation. This is used for subsequent local convolution and attention operations.

[0036] To map the channel dimension to the node dimension, facilitating feature dependency computation along the node direction, local features are extracted and computed using Residual Depthwise Separable Convolution (ResDWConv). (3) in: For local features, preserve local spatial relationships; It is a depthwise separable convolution function used for local feature extraction.

[0037] Residual connections are used to preserve original information and suppress gradient vanishing, capturing stress and displacement characteristics changes in the local neighborhood of nodes, thus reflecting the local response patterns of the bridge structure under flood impact. After obtaining local features, the model further utilizes an attention mechanism to construct global dependencies between nodes, enhancing its ability to perceive the overall stress coupling pattern of the structure. Specifically, local features are first mapped to... Three feature spaces: (4) in: These are the query vector, key vector, and value vector, respectively. This represents a 1×1 convolution, where the number of output channels is three times the number of input channels; This means that the convolution output feature map is divided into three parts along the channel dimension, which are used as the query vector, key vector and value vector respectively.

[0038] Calculate attention weights: (5) in: Attention weights; For single-head feature dimensions, typically 10 ... , For the number of attention heads.

[0039] Obtain the global dependencies, and then obtain the global features through weighted summation: (6) in: This is a global feature.

[0040] Then, the channel dimension is restored using linear projection: (7) in: This is the projected global feature map; , which means outputting the projection weight matrix, mapping the attention results back to the original channel dimension.

[0041] To enhance the model's ability to perceive the spatial distribution and local positional relationships of nodes, a convolutional positional bias is introduced. A 3×3 depthwise convolution is used to model the spatial coordinates and neighborhood relationships of nodes, which are then added to the residuals of the original features. This compensates for the standard attention mechanism's insufficient perception of spatial topology, enabling the model to effectively identify the overall force coupling patterns across nodes. The calculation formula is as follows:

[0042] (8) in: To enhance the feature map; For the local location enhancement function, in this embodiment, a 3×3 depthwise separable convolution function is used.

[0043] To adaptively adjust the feature response intensity after fusing global features, a gating mechanism is introduced: (9) (10) in: For gating weight tensors; Use the Sigmoid activation function; It is a feature map after gating adjustment, which has adaptive channel response capability.

[0044] To enhance feature resolution at the channel level, the gated feature maps are input into the SE module for further optimization of channel feature weights. This module first performs global average pooling: (11) in: This is the channel description vector after global average pooling; This is global average pooling.

[0045] First, a 1×1 convolution and GELU activation function are used for channel dimensionality reduction and non-linear activation; then, a 1×1 convolution and sigmoid activation function are used for channel dimensionality increase and non-linear activation. (12) in: This is the channel attention weight vector; This is the GELU activation function.

[0046] Weighted by input features: (13) in: The feature map is weighted by channel, highlighting important features; This is a pixel-by-pixel multiplication.

[0047] Subsequently, a multilayer perceptron (MLP) and a global response normalization layer (GRN) are used to enhance the representation of nonlinear features: (14) in: The output features of the multilayer perceptron; For global response normalization layer, It is a multilayer perceptron.

[0048] And by maintaining gradient stability through the residual path, the final output is obtained: (15) in: This indicates the output features of the LAG-Block module, which represent the final fusion result.

[0049] Finally, the output tensor undergoes dimension reduction: (16) in: The final output tensor of the LAG-Block module; This is a dimension restoration operation.

[0050] We obtain a feature map with the same shape as the input: .

[0051] Under different water depths and flow velocities during floods, the stress and displacement responses of nodes exhibit significant differences with varying operating conditions, containing crucial information about the structural state under flood load. To more effectively utilize these mechanical characteristics obtained from simulation data and achieve reliable identification of risk locations at the node level, a LAG-BRINet network structure was constructed, enabling it to extract, fuse, and discriminate input features layer by layer. The overall structure of the LAG-BRINet model is as follows: Figure 3As shown, the model takes four types of features as input: water depth, flow velocity, equivalent stress, and total deformation of the nodes. These features are derived from multi-condition numerical simulations and can reflect the basic stress state of the nodes under different flood conditions. To make the input more suitable for the high-dimensional feature space of the subsequent network, the model completes channel mapping through 1×1 convolution, which avoids the influence of differences in physical quantity scales while ensuring the consistency of feature dimensions. The mapped tensor enters the feature extraction backbone composed of the LAG-Block module. The LAG-Block module combines local convolution, multi-head attention, position bias, and channel gating mechanisms in the same structure, enabling the network to simultaneously acquire local response patterns and global correlation features across nodes. Depthwise separable convolution focuses on capturing local stress-displacement changes, multi-head attention is used to characterize long-range dependencies between nodes, while position bias and channel attention further enhance the spatial sensitivity and channel selection ability of feature expression, thereby obtaining more discriminative node representations.

[0052] After extracting the backbone features, the network weights the fused features using a channel attention module to evaluate the contribution of each channel to the overall response. This step suppresses interference from local noise, allowing the model to focus more on structural features in the final classification stage, resulting in more stable risk location identification. The final classifier consists of 1×1 convolutions, GELU activation, and Dropout, mapping features to specific class spaces at the end. The introduction of GELU makes the classifier more flexible in responding to high-dimensional feature variations, while Dropout improves the model's generalization ability on multi-condition data. Through this structure, LAG-BRINet transforms features into risk location classification results, achieving a complete mapping process from simulated data to risk labels.

[0053] The node outputs of the entire network can be represented as: (17) in: The features after channel mapping are completed by 1×1 convolution; This represents the local and global fusion features extracted by the LAG-Block module, which combines depthwise separable convolution, attention mechanism, positional bias and channel gating to achieve multi-scale feature representation. These are features reweighted by the channel attention module to enhance the response of critical channels; This represents the feature after global average pooling. For end-point classifiers; This is the final location risk output. It should be noted that the principle behind the final location risk output is as follows: each element in the input feature matrix corresponds to the water depth, flow velocity, point equivalent stress, and total deformation data for each node. Therefore, when using the risk identification model for prediction, the model predicts a probability value for each element in the input feature matrix through a classifier. The location of the node with bridge flood risk is determined based on the index value corresponding to the maximum probability value among all probability values.

[0054] S3, Risk Identification Model Training The training process employs a cross-entropy loss function and introduces label smoothing to mitigate the impact of class noise on model convergence. The model training objective is to minimize the label-smoothed cross-entropy loss function, and its overall optimization objective can be expressed as: (18) in, The objective is to optimize the parameters; The model loss function; The total number of training samples; Total number of categories; For the first The sample at the th The smoothed true label on the class; For the first Each sample is predicted as follows The probability of a class; Represents the regularization coefficient. The parameters represent the risk identification model. Denotes the square of the L2 norm. This is an L2 regularization term used to prevent overfitting.

[0055] By using the Adam optimizer and adaptive learning rate scheduling (ReduceLROnPlateau) to update parameters, the model achieves stable convergence and good generalization with a limited number of samples.

[0056] S4. Use the trained risk identification model to identify the location of bridge flood risk.

[0057] To verify the effectiveness of the risk identification model in identifying flood risk locations under multiple conditions, its performance was compared with commonly used deep learning models. Structural ablation experiments were conducted to analyze the impact of different modules on model performance. Under the same training samples and hyperparameter conditions, traditional convolutional neural networks (CNN), long short-term memory networks (LSTM), and multilayer perceptrons (MLP) were selected as control models and compared with the LAG-BRINet network proposed in this invention. The comparison results are shown in Table 1.

[0058] Table 1 Comparative Experimental Results

[0059] As shown in the table, LAG-BRINet outperforms the control model in all four metrics. Compared to traditional CNN and LSTM models, the accuracy of the LAG-BRINet model is improved by approximately 3.7% and 1.6%, respectively, and the F1 score is improved by approximately 3%. To further analyze the impact of key internal modules of LAG-BRINet on overall performance, three ablation experiments were designed: removing the LAG-Block module, removing the SE module, removing global average pooling, and removing node properties. The results of the ablation experiments are shown in Table 2.

[0060] Table 2 Ablation Experiment Results

[0061] The ablation experiments show that LAG-BRINet achieves the best performance across all metrics, with an accuracy of 93% and an F1 score of 91%. Removing the LAG-Block main module significantly reduces the overall model performance, with the F1 score decreasing by approximately 3.9%. Removing the SE module reduces both accuracy and precision by approximately 2.5%. Removing global average pooling or canceling node properties also leads to a decrease in all metrics. Overall, the performance improvement of the LAG-BRINet model stems from the synergistic effect of multiple modules: local convolution and attention mechanisms provide spatial feature fusion capabilities, the SE module optimizes feature weight distribution, global average pooling achieves feature compression and aggregation, and node properties enhance the model's structural sensitivity.

[0062] Analysis of the comparative and ablation experiments reveals that the performance advantage of the LAG-BRINet model primarily stems from the synergistic effect of local convolutional feature extraction, global attention fusion, adaptive channel weights, and node attribute modeling. Based on these mechanisms, the LAG-BRINet model can effectively capture stress and deformation characteristics under different water depths and flow velocities, thereby achieving effective identification of risk locations under bridge flooding. Detailed bar charts of the comparative and ablation experiments are shown in Figures 4(a) and 4(b), where Figure 4(a) is the comparative experiment bar chart and Figure 4(b) is the ablation experiment bar chart.

[0063] This invention addresses the challenge of identifying risks at key bridge nodes under flood conditions by constructing an intelligent method for identifying bridge flood risk locations based on fluid-structure interaction numerical simulation and deep learning. To address the locality, multi-scale nature, and cross-node correlation of structural responses under flood conditions, a LAG-BRINet model integrating local convolution, multi-head attention, and channel gating mechanisms is proposed. This model achieves deep representation of flood risk features and node-level location identification. The model structure effectively captures salient local features and integrates global cross-node information, providing a more expressive feature framework for risk identification under complex conditions. The model identification process and comparative experiments demonstrate that the method based on the coupling of numerical simulation and deep learning can achieve rapid identification of flood risks at key bridge nodes, supplementing traditional assessment methods based on experience or monitoring signals. This technical approach provides a generalizable foundation for bridge flood safety assessment, risk warning threshold setting, and intelligent auxiliary decision-making.

[0064] Example 2 Based on Embodiment 1, Embodiment 2 of the present invention also provides a bridge flood risk location identification system, comprising: The dataset construction module is used to select feature extraction locations on the bridge as nodes, extract the equivalent stress and total deformation values ​​of each node under different water depth and flow velocity combinations, and construct the dataset. The model building module is used to construct the LAG-Block module. The LAG-Block module is used to capture local detailed features in the spatial domain through convolution and to establish global dependencies in the feature domain through an attention mechanism. The convolution module, LAG-Block module, first channel attention module, first global average pooling and classifier are sequentially connected to form the risk recognition model. The model training module is used to train the risk identification model; The risk identification module is used to identify the location of bridge flood risks using a trained risk identification model.

[0065] Specifically, the working process of the LAG-Block module is as follows: The input feature matrix is ​​represented as an input feature tensor. The input feature tensor undergoes dimensionality transformation to obtain dimensionality-transformed input features. The dimensionality-transformed input features are then convolved with depthwise separable convolutions to obtain local features. These local features are mapped through a first convolutional layer to obtain query vectors, key vectors, and value vectors. Attention weights are calculated using the query vectors and key vectors. The global features are obtained by multiplying the attention weights by the value vectors. The global features are then linearly projected to restore the channel dimensions, resulting in a projected global feature map. The projected global feature map is then compared with the local features obtained through linear projection. The results of the enhancement function are summed to obtain the enhanced feature map. The enhanced feature map is then processed by the second convolutional layer and then by the first activation function to obtain the gated weight tensor. The gated weight tensor is multiplied by the enhanced feature map to obtain the gated feature map. The gated feature map is input into the second channel attention module to obtain the channel-weighted feature map. The channel-weighted feature map is then passed through a multilayer perceptron and a global response normalization layer to obtain the multilayer perceptron output feature. The multilayer perceptron output feature is added to the channel-weighted feature map and then subjected to a dimension reduction operation to obtain a feature map with the same shape as the input, which is used as the final output feature of the LAG-Block module.

[0066] More specifically, the first channel attention module and the second channel attention module have the same structure, and the processing procedure of the second channel attention module is as follows: The gated feature map is subjected to a second global average pooling to obtain the channel description vector. The channel description vector is then processed by a third convolutional layer and a second activation function, and then by a fourth convolutional layer and a third activation function to obtain the channel attention weight vector. The channel attention weight vector is multiplied pixel-by-pixel with the gated feature map to obtain the channel-weighted feature map.

[0067] More specifically, the convolution module and the first to fourth convolutional layers are all 1×1 convolutions.

[0068] Furthermore, the classifier consists of a fifth convolutional layer, a fourth activation function, a Dropout operation, and a sixth convolutional layer connected in sequence.

[0069] More specifically, the fifth and sixth convolutional layers are 1×1 convolutions.

[0070] Specifically, the model training module is also used for: The parameters of the risk identification model are continuously adjusted, and the value of the overall optimization objective is calculated. Training stops when the value of the overall optimization objective is minimized, resulting in a well-trained risk identification model. The formula for calculating the overall optimization objective is as follows:

[0071] in, The objective is to optimize the parameters; The model loss function; The total number of training samples; Total number of categories; For the first The sample at the th The smoothed true label on the class; For the first Each sample is predicted as follows The probability of a class; Represents the regularization coefficient. The parameters represent the risk identification model. Denotes the square of the L2 norm. This is an L2 regularization term.

[0072] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying the location of bridge flood risk, characterized in that, Includes the following steps: S1. Select feature extraction locations on the bridge as nodes, extract the equivalent stress and total deformation values ​​of each node under different water depth and flow velocity combinations, and construct a dataset. S2. Construct the LAG-Block module. The LAG-Block module is used to capture local detailed features in the spatial domain through convolution and to establish global dependencies in the feature domain through an attention mechanism. The convolution module, LAG-Block module, first channel attention module, first global average pooling, and classifier are sequentially connected to form the risk identification model. The working process of the LAG-Block module is as follows: The input feature matrix is ​​represented as an input feature tensor. The input feature tensor is dimensionally transformed to obtain dimensionally transformed input features. The dimensionally transformed input features are then subjected to depthwise separable convolution, and the result of the convolution is added to the dimensionally transformed input features to obtain local features. The local features are then mapped through the first convolutional layer to obtain query vectors, key vectors, and value vectors. Attention weights are calculated using the query vectors and key vectors. The global features are obtained by multiplying the attention weights by the value vectors. The global features are restored to channel dimensions through linear projection to obtain the projected global feature map. The projected global feature map is added to the result of the local location enhancement function to obtain the enhanced feature map. The enhanced feature map is processed by the second convolutional layer and then passed through the first activation function to obtain the gated weight tensor. The gated weight tensor is multiplied with the enhanced feature map to obtain the gated feature map. The gated feature map is input into the second channel attention module to obtain the channel-weighted feature map. The channel-weighted feature map is passed through the multilayer perceptron and the global response normalization layer to obtain the multilayer perceptron output feature. The multilayer perceptron output feature is added to the channel-weighted feature map and then subjected to dimension restoration to obtain a feature map with the same shape as the input, which is the final output feature of the LAG-Block module. S3. Train the risk identification model; S4. Use the trained risk identification model to identify the location of bridge flood risk.

2. The method for identifying the location of bridge flood risk according to claim 1, characterized in that, The first channel attention module and the second channel attention module have the same structure. The processing procedure of the second channel attention module is as follows: The gated feature map is subjected to a second global average pooling to obtain the channel description vector. The channel description vector is then processed by a third convolutional layer and a second activation function, and then by a fourth convolutional layer and a third activation function to obtain the channel attention weight vector. The channel attention weight vector is multiplied pixel-by-pixel with the gated feature map to obtain the channel-weighted feature map.

3. The method for identifying the location of bridge flood risk according to claim 2, characterized in that, The convolutional module and the first to fourth convolutional layers are all 1×1 convolutions.

4. The method for identifying the location of bridge flood risk according to claim 1, characterized in that, The classifier consists of a fifth convolutional layer, a fourth activation function, a Dropout operation, and a sixth convolutional layer connected in sequence.

5. The method for identifying the location of bridge flood risk according to claim 4, characterized in that, The fifth and sixth convolutional layers are 1×1 convolutions.

6. The method for identifying the location of bridge flood risk according to claim 1, characterized in that, S3 includes: The parameters of the risk identification model are continuously adjusted, and the value of the overall optimization objective is calculated. Training stops when the value of the overall optimization objective is minimized, resulting in a well-trained risk identification model. The formula for calculating the overall optimization objective is as follows: in, The objective is to optimize the parameters; The model loss function; The total number of training samples; Total number of categories; For the first The sample at the th The smoothed true label on the class; For the first Each sample is predicted as follows The probability of a class; Represents the regularization coefficient. The parameters represent the risk identification model. Denotes the square of the L2 norm. This is an L2 regularization term.

7. A bridge flood risk location identification system, characterized in that, include: The dataset construction module is used to select feature extraction locations on the bridge as nodes, extract the equivalent stress and total deformation values ​​of each node under different water depth and flow velocity combinations, and construct the dataset. The model building module is used to construct the LAG-Block module. The LAG-Block module captures local detail features in the spatial domain through convolution and establishes global dependencies in the feature domain through an attention mechanism. The convolution module, LAG-Block module, first channel attention module, first global average pooling module, and classifier are sequentially connected to form the risk recognition model. The working process of the LAG-Block module is as follows: The input feature matrix is ​​represented as an input feature tensor. The input feature tensor is dimensionally transformed to obtain dimensionally transformed input features. The dimensionally transformed input features are then subjected to depthwise separable convolution, and the result of the convolution is added to the dimensionally transformed input features to obtain local features. The local features are then mapped through the first convolutional layer to obtain query vectors, key vectors, and value vectors. Attention weights are calculated using the query vectors and key vectors. The global features are obtained by multiplying the attention weights by the value vectors. The global features are restored to channel dimensions through linear projection to obtain the projected global feature map. The projected global feature map is added to the result of the local location enhancement function to obtain the enhanced feature map. The enhanced feature map is processed by the second convolutional layer and then passed through the first activation function to obtain the gated weight tensor. The gated weight tensor is multiplied with the enhanced feature map to obtain the gated feature map. The gated feature map is input into the second channel attention module to obtain the channel-weighted feature map. The channel-weighted feature map is passed through the multilayer perceptron and the global response normalization layer to obtain the multilayer perceptron output feature. The multilayer perceptron output feature is added to the channel-weighted feature map and then subjected to dimension restoration to obtain a feature map with the same shape as the input, which is the final output feature of the LAG-Block module. The model training module is used to train the risk identification model; The risk identification module is used to identify the location of bridge flood risks using a trained risk identification model.

8. A bridge flood risk location identification system according to claim 7, characterized in that, The first channel attention module and the second channel attention module have the same structure. The processing procedure of the second channel attention module is as follows: The gated feature map is subjected to a second global average pooling to obtain the channel description vector. The channel description vector is then processed by a third convolutional layer and a second activation function, and then by a fourth convolutional layer and a third activation function to obtain the channel attention weight vector. The channel attention weight vector is multiplied pixel-by-pixel with the gated feature map to obtain the channel-weighted feature map.

Images