A bridge health state detection method and system based on a double-branch mamba network

Abstract

The application discloses a bridge health state detection method and system based on a double-branch mamba network. The method comprises the following steps: collecting bridge vibration field data and processing, generating a vibration field time profile and an instantaneous frequency diagram, obtaining a data set and a training set; taking the data set as input, constructing a double-path lightweight network model for bridge health state identification; using the training set to train and optimize the double-path lightweight network model; inputting a test sample into the trained double-path lightweight network model, outputting a bridge health state classification result, and evaluating the model performance. The system comprises a data acquisition module, a model construction module, a training module and a performance evaluation module. The application uses distributed sensors arranged on both sides of the bridge pavement, and uses vehicle driving as an excitation source, so that the bridge health condition detection can be realized without affecting the normal operation of traffic.

Description

Technical Field

[0001] This invention relates to the field of neural network technology for bridge health detection, and in particular to a method and system for bridge health status detection based on a dual-branch Mamba network. Background Technology

[0002] In recent years, my country's economy has maintained rapid and sustained growth, and various infrastructure projects have been gradually improved, with bridge construction playing a crucial role. Bridges are constantly subjected to dynamic traffic loads and changes in climate, making them prone to structural damage, material aging, and decreased load-bearing capacity. If these problems are not detected and addressed promptly, they may lead to bridge fractures, traffic accidents, or economic losses. Therefore, bridge inspection has received widespread attention. While traditional drilling methods offer high reliability, they are costly and cause some damage to the bridge. Thus, the gradual development and application of non-destructive testing (NDT) technologies has become a major trend. Common NDT techniques for bridge health status include ground-penetrating radar (GPR) and seismic reflection wave analysis. GPR provides high-resolution information by emitting high-frequency electromagnetic waves, but its effectiveness is significantly affected by the electromagnetic properties of the bridge materials. Seismic reflection wave analysis, as a geophysical exploration method, involves deploying multiple seismic wave transmission and reception points at different locations on the bridge for multi-directional detection. When damage exists in the bridge structure, the damaged areas alter the propagation path and reflection characteristics of seismic waves. Although these methods are non-destructive and effective, data analysis and judgment require professional expertise, and the exploration process may require temporary road closures, increasing the complexity of their use.

[0003] With the development of seismic motion measurement technology, using moving vehicles as active vibration sources and sensors as receiving sources, the echo response of the underground medium to seismic waves can be obtained, thereby constructing a time profile of the seismic field. Furthermore, neural networks can be used to model this time profile, enabling intelligent detection of bridge health status. This method does not require damage to the bridge or affect normal road operation, and can continuously monitor traffic flow, improving the level of intelligent operation and maintenance. Therefore, this technology has become a research hotspot in the field of intelligent transportation. However, the following problems still exist in using neural networks for bridge health status detection:

[0004] (1) Bridge health monitoring models suffer from low accuracy and poor generalization ability. Bridge health monitoring requires a large amount of high-quality data to train the model. However, due to the lack of diversity in samples (seismic field time profiles) and the difficulty in labeling, the models exhibit significant deficiencies in accuracy and generalization ability. Furthermore, the acquisition of seismic field time profiles is complex and costly, limiting the scale of the dataset and making it difficult for the model to cope with complex bridge defects and structural changes. In addition, the lack of diverse samples with different sampling frequencies, vibration amplitudes, and propagation paths makes it difficult for the model to comprehensively learn the characteristics of bridge structures under various working conditions, hindering the formation of accurate and widely adaptable models. When faced with new bridge conditions, model predictions are prone to bias, and accuracy is difficult to meet standards.

[0005] (2) Complex bridge health monitoring models and insufficient real-time bridge diagnosis. In the field of bridge health monitoring, existing bridge health assessment models generally suffer from complex structures and deep network layers, thus requiring powerful servers for data processing. However, many older bridges, due to limitations in construction technology at the time, did not have any monitoring sensors installed during construction. For these bridges, the main approach is to install sensors externally via wired connections, followed by data processing for training and testing of the bridge assessment model. This method increases the workload on-site, resulting in low real-time bridge diagnosis, high testing costs, and severely hindering the efficient implementation of bridge health monitoring. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a bridge health status detection method and system based on a dual-branch Mamba network. By deploying distributed sensors on both sides of the bridge road surface and using vehicle traffic as an excitation source, the health status of the bridge can be detected without affecting the normal operation of traffic.

[0007] The embodiments of the present invention are implemented as follows:

[0008] A bridge health status detection method based on a dual-branch Mamba network, comprising:

[0009] Bridge seismic field data are collected and processed to generate seismic field time profiles and instantaneous frequency maps, resulting in datasets and training sets.

[0010] Using the dataset as input, a dual-path lightweight network model is constructed for bridge health status identification.

[0011] The training set is used to train and optimize the dual-path lightweight network model.

[0012] The test samples are input into the trained dual-path lightweight network model, which outputs the bridge health status classification results and evaluates the model performance.

[0013] In a preferred embodiment of the present invention, in the above-described bridge health status detection method based on a dual-branch Mamba network, the step of collecting and processing bridge vibration field data to generate a vibration field time profile and an instantaneous frequency map, and obtaining a dataset and a training set, includes:

[0014] Several nodes are set at equal intervals along the bridge and road, and a vibration sensor is installed at each node.

[0015] The vibration generated by a vehicle traveling at a constant speed is used as the excitation source, and vibration signals are acquired from each of the vibration sensors.

[0016] The array data acquired from the vibration sensor is integrated to obtain a vibration field time profile.

[0017] By using Fourier transform, the time domain of the vibration field time profile is converted into the frequency domain to obtain the instantaneous frequency map.

[0018] A finite element model is established, and the corresponding vibration field time profile and instantaneous frequency diagram are generated on the finite element model using the finite difference method, thus obtaining the dataset and training set.

[0019] In a preferred embodiment of the present invention, in the above-described bridge health status detection method based on a dual-branch Mamba network, the step of using the dataset as input to construct a dual-path lightweight network model for bridge health status identification includes:

[0020] The vibration field time profile and the instantaneous frequency map are used as input data and processed separately through symmetrical dual paths. Each path passes through a block embedding layer in sequence, dividing the input image into 4×4 non-overlapping blocks.

[0021] The segmented image is downsampled and its features extracted through alternating feature extraction modules and convolutional layers.

[0022] The image after feature extraction is input into the feature fusion module, and after channel concatenation is performed by element-wise addition, it enters the depthwise separable convolution.

[0023] Multimodal feature fusion is performed using a multi-head attention module and a channel attention module to generate high-quality fused features.

[0024] The high-quality fused features are input into a classifier to construct a dual-path lightweight network model for bridge health status identification.

[0025] In a preferred embodiment of the present invention, in the above-described bridge health status detection method based on a dual-branch Mamba network, the operations performed by the feature extraction module include:

[0026] The segmented image is divided into sub-feature maps X1 and X2 by channel segmentation, and then input into the convolution branch and the state space model branch, respectively.

[0027] For the sub-feature map X1, normalization processing is performed. Local features are extracted through the first layer 3×3 convolution kernel, and then batch normalization and activation function operations are performed. Then, features are further extracted through the second layer 3×3 convolution kernel, and normalization and activation operations are performed.

[0028] For the sub-feature map X2, normalization is performed, feature dimension transformation is performed through a linear layer, spatial correlation is extracted using depthwise convolution, and activation function is applied to enhance nonlinear expressive power.

[0029] Based on the attention mechanism, the results of the convolutional branch and the state space model branch are fused and output.

[0030] In a preferred embodiment of the present invention, in the above-described bridge health status detection method based on a dual-branch Mamba network, the operations performed by the feature fusion module include:

[0031] The temporal features extracted by the feature extraction module are represented as follows: Frequency domain features are represented as Where i∈{1,2,3,4,5}, i represents the i-th feature extraction module.

[0032] Time-domain feature vectors and frequency domain eigenvectors A concatenation operation is performed along the channel dimension to obtain a bimodal feature vector.

[0033] The time-domain feature vector The frequency domain feature vector and the dual-modal feature vector Layer normalization is performed to accelerate model convergence. Depthwise separable convolutions are then applied sequentially to the normalized feature vectors, followed by SiLU activation to obtain the enhanced temporal feature vector f. i r Enhanced frequency domain eigenvector f i d and enhance the bimodal eigenvector f i z .

[0034] The enhanced temporal feature vector f i r The enhanced frequency domain feature vector f i d and the enhanced bimodal feature vector f i zInput a multi-head attention module to generate attention-weighted features.

[0035] The time-domain feature vector and the frequency domain feature vector Input the channel attention module to generate channel weights ω.

[0036] The features output by the multi-head attention module are multiplied element-wise with the weights generated by the channel attention module to obtain the time-domain weighted feature y. r and frequency domain weighted y d The final fused feature F is generated by adding elements one by one. i .

[0037] In a preferred embodiment of the present invention, in the above-described bridge health status detection method based on a dual-branch Mamba network, the operations performed by the channel attention module include:

[0038] For the time-domain feature vector and the frequency domain feature vector Apply the SiLU activation function respectively.

[0039] Perform global average pooling and global max pooling on the activated features to obtain the temporal pooling vector F. avg and frequency domain pooling vector F max .

[0040] The time-domain pooling vector F avg and the frequency domain pooling vector F max The features are transformed using 1×1 convolution kernels and ReLU activation function is applied. The transformed features are then summed and channel weights ω are generated using the Sigmoid function.

[0041] In a preferred embodiment of the present invention, in the above-described bridge health status detection method based on a dual-branch Mamba network, the operations performed by the multi-head attention module include:

[0042] For the enhanced time-domain feature vector f i r The enhanced frequency domain feature vector f i d and the enhanced bimodal feature vector f i z Perform linear transformations on each element, and map them to different subspaces through matrix multiplication and bias operations to generate query, key, and value vectors.

[0043] Calculate the similarity matrix between the query and the key, adjust it with a scaling factor, and then normalize it using Softmax. Finally, perform a weighted summation of the normalized result and the value vector to generate a single-head attention feature.

[0044] Multiple single-head attention features are concatenated along the channel dimension to form a long vector. The concatenated features are then dimensionality-reduced and fused using a linear transformation layer to generate attention-weighted features.

[0045] In a preferred embodiment of the present invention, the bridge health status detection method based on a dual-branch Mamba network described above, wherein training and optimizing the dual-path lightweight network model using the training set includes:

[0046] The training environment was configured, and the training framework was set up under the Ubuntu 22.04 operating system. The PyTorch 2.1.0 deep learning framework was used, and the model parameters and training data were loaded to the GPU through the torch.cuda module to accelerate training.

[0047] Set the training parameters, define the cross-entropy loss function as the optimization objective for model training, use the AdamW optimizer for parameter updates, set the initial learning rate to 0.001, the weight decay to 0.05, the training batch size to 256, and the total number of iterations to 300.

[0048] The training set is input into the dual-path lightweight network model, and the loss value is calculated by forward propagation in batches. The network parameters are updated by backpropagation, the learning rate is dynamically adjusted, and the convergence efficiency is optimized.

[0049] The vibration field time profile was divided into training samples and test samples at a 7:3 ratio.

[0050] The pre-trained model is fine-tuned using the training samples, and the generalization ability and classification accuracy of the fine-tuned model are verified using the test samples.

[0051] In a preferred embodiment of the present invention, the bridge health status detection method based on a dual-branch Mamba network described above, wherein inputting test samples into the trained dual-path lightweight network model, outputting bridge health status classification results, and evaluating model performance includes:

[0052] The test samples are input into the trained dual-path lightweight network model, which outputs the bridge health status classification results.

[0053] The bridge health status classification results are compared with the true labels of the test samples to calculate performance indicators, including overall accuracy, and to evaluate the model performance.

[0054] A bridge health status detection system based on a dual-branch Mamba network, comprising:

[0055] The data acquisition module is used to collect and process bridge vibration field data, generate vibration field time profiles and instantaneous frequency maps, and obtain datasets and training sets.

[0056] The model building module is used to take the dataset as input and build a dual-path lightweight network model for bridge health status identification.

[0057] The training module is used to train and optimize the dual-path lightweight network model using the training set.

[0058] The performance evaluation module is used to input test samples into the trained dual-path lightweight network model, output the bridge health status classification results, and evaluate the model performance.

[0059] The beneficial effects of the embodiments of the present invention are:

[0060] This invention addresses the problems of low accuracy and poor generalization ability in neural network models for bridge health monitoring by proposing a method combining finite element models with actual data acquisition. First, based on measured data, a dataset of vibration field time profiles and instantaneous frequency maps under different bridge health conditions is generated using the finite element model. This provides rich training samples for the neural network model, improving its generalization ability. Second, the model parameters of the neural network are fine-tuned using actual data acquisition, further enhancing the model's adaptability to real-world scenarios.

[0061] This invention addresses the problems of complex neural network models for bridge health detection and insufficient real-time performance in bridge diagnosis. It proposes a lightweight network model that utilizes efficient two-dimensional scanning, depthwise separable convolution, and multi-head attention modules to efficiently and quickly extract and fuse effective features from time profile maps and instantaneous frequency maps, thereby reducing computational complexity and improving detection efficiency. Attached Figure Description

[0062] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a flowchart of the bridge health status detection method based on a dual-branch Mamba network according to the present invention;

[0064] Figure 2 This is a schematic diagram of the overall network framework structure of the bridge health detection method based on a dual-branch Mamba network according to the present invention.

[0065] Figure 3 This is a schematic diagram of the sensor deployment structure in the bridge health status detection method based on a dual-branch Mamba network of the present invention.

[0066] Figure 4a This is a temporal profile of the vibration field in the bridge health status detection method based on a dual-branch Mamba network of the present invention.

[0067] Figure 4b This is an instantaneous frequency diagram of the bridge health status detection method based on a dual-branch Mamba network in this invention;

[0068] Figure 5 This is a schematic diagram of the dual-branch Mamba network model structure in the bridge health status detection method based on dual-branch Mamba network of the present invention;

[0069] Figure 6 This is a schematic diagram of the feature extraction module in the bridge health status detection method based on a dual-branch Mamba network of the present invention.

[0070] Figure 7 This is a schematic diagram of the feature fusion module structure in the bridge health status detection method based on a dual-branch Mamba network of the present invention.

[0071] Figure 8 This is a schematic diagram of the channel attention module structure in the bridge health status detection method based on dual-branch Mamba network of the present invention;

[0072] Figure 9 This is a schematic diagram of the multi-head attention module structure in the bridge health status detection method based on dual-branch Mamba network of the present invention. Detailed Implementation

[0073] 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 with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0074] Please refer to Figures 1 to 9 The first embodiment of the present invention provides a bridge health status detection method based on a dual-branch Mamba network, which includes: collecting and processing bridge vibration field data to generate a vibration field time profile and an instantaneous frequency map, obtaining a dataset and a training set; using the dataset as input to construct a dual-path lightweight network model for bridge health status identification; using the training set to train and optimize the dual-path lightweight network model; inputting test samples into the trained dual-path lightweight network model, outputting bridge health status classification results, and evaluating model performance.

[0075] like Figure 2 As shown, this invention provides a bridge health status detection method based on a dual-branch Mamba network. By deploying distributed sensors on both sides of the bridge surface and using vehicle traffic as the excitation source, the health status of the bridge can be detected without affecting normal traffic flow. This invention mainly comprises two parts: bridge internal vibration field data acquisition and processing, and a bridge health status detection network. The data acquisition and processing part is used to obtain bridge vibration information and construct the dataset. The bridge health status detection network classifies the health status based on the vibration field time profile and instantaneous frequency diagram.

[0076] In a preferred embodiment of the present invention, in the above-described bridge health status detection method based on a dual-branch Mamba network, the step of collecting and processing bridge vibration field data to generate a vibration field time profile and an instantaneous frequency map, and obtaining a dataset and a training set, includes: Figure 3 As shown, several nodes are set at equal intervals along the bridge road, and a vibration sensor is installed at each node. Specifically, each detection area is 100m long, with a spacing of 0.5m, and a total of 200 vibration sensors are arranged. The vibration generated by a vehicle traveling at a constant speed is used as the excitation source, and vibration signals are acquired from each vibration sensor. The array data acquired from the vibration sensors are integrated to obtain a vibration field time profile, as shown. Figure 4a As shown; by using Fourier transform, the time domain of the vibration field time profile is transformed into the frequency domain, resulting in an instantaneous frequency diagram, as shown. Figure 4b As shown in the "Specifications for Maintenance of Highway Bridges and Culverts", a finite element model is established according to different health states, such as good, relatively good, and poor, and parameter configurations. The corresponding vibration field time profile and instantaneous frequency diagram are generated on the finite element model using the finite difference method, thus obtaining the dataset and training set.

[0077] In a preferred embodiment of the present invention, the bridge health status detection method based on the dual-branch Di-Mamba network described above, wherein the step of using the dataset as input to construct a dual-path lightweight network model for bridge health status identification includes: using the seismic field time profile and the instantaneous frequency map as input data, processing them separately through symmetrical dual paths, with each path sequentially passing through a block embedding layer to divide the input image into 4×4 non-overlapping blocks; the block-based images undergo downsampling and feature extraction through alternating feature extraction modules and convolutional layers; the feature-extracted images are input into a feature fusion module, and after channel concatenation through element-wise addition, they enter a depthwise separable convolution; multimodal feature fusion is performed through a multi-head attention module and a channel attention module to generate high-quality fused features; the high-quality fused features are input into a classifier to construct a dual-path lightweight network model for bridge health status identification. The structure diagram of the dual-path lightweight network model based on the dual-branch Di-Mamba network is shown below. Figure 5 As shown.

[0078] In a preferred embodiment of the present invention, in the bridge health status detection method based on a dual-branch Mamba network, the operation performed by the feature extraction module includes: dividing the segmented image into sub-feature maps X1 and X2 through channel segmentation, and inputting them into the convolutional branch and the state-space model (SSM) branch, respectively; for the sub-feature map X1, normalization processing is performed, local features are extracted through a first-layer 3×3 convolutional kernel, then batch normalization (BatchNorm) and activation function operations are performed, and then features are further extracted through a second-layer 3×3 convolutional kernel, followed by normalization and activation operations; for the sub-feature map X2, further... Normalization is performed, feature dimension transformation is carried out through linear layers, spatial correlation is extracted using depthwise convolution, and activation functions are applied to enhance nonlinear expressive power. Specifically, in the state space model branch, an efficient two-dimensional selective scanning ES2D method is introduced, which skips feature map scanning blocks with a stride p and only scans selected regions, dividing the feature map into multiple smaller blocks, reducing the number of spatial labels to be processed, and preserving key feature information while reducing computational complexity. Based on the attention mechanism concatenation operation, the results of the convolution branch and the state space model branch are fused and output. The network structure of the feature extraction module is as follows: Figure 6 As shown.

[0079] In a preferred embodiment of the present invention, in the above-described bridge health status detection method based on a dual-branch Mamba network, the operation performed by the feature fusion module includes: representing the temporal features extracted by the feature extraction module as follows: Frequency domain features are represented as Where i∈{1,2,3,4,5}, i represents the i-th feature extraction module; the time-domain feature vector and frequency domain eigenvectors A concatenation operation is performed along the channel dimension to obtain a bimodal feature vector. The time-domain feature vector The frequency domain feature vector and the dual-modal feature vector Layer normalization is performed to accelerate model convergence. Depthwise separable convolutions are then applied sequentially to the normalized feature vectors, followed by SiLU activation to obtain the enhanced temporal feature vector f. i r Enhanced frequency domain eigenvector f i d and enhance the bimodal eigenvector f i z The enhanced temporal feature vector f i rThe enhanced frequency domain feature vector f i d and the enhanced bimodal feature vector f i z Input the multi-head attention module to generate attention-weighted features; then input the temporal feature vector. and the frequency domain feature vector Input the channel attention module to generate channel weights ω; multiply the features output by the multi-head attention module with the weights generated by the channel attention module element-wise to obtain the temporal weighted features y. r and frequency domain weighted y d The final fused feature F is generated by adding elements one by one. i The feature fusion module framework is as follows: Figure 7 As shown.

[0080] In a preferred embodiment of the present invention, in the above-described bridge health status detection method based on a dual-branch Mamba network, the operation performed by the channel attention module includes: processing the temporal feature vector. and the frequency domain feature vector The SiLU activation function is applied to each feature; global average pooling and global max pooling are then performed on the activated features to obtain the temporal pooling vector F. avg and frequency domain pooling vector F max ; the time-domain pooling vector F avg and the frequency domain pooling vector F max Feature transformation is performed using 1×1 convolution kernels, and the ReLU activation function is applied. The transformed features are then summed, and the channel weights ω are generated using the Sigmoid function. The structure diagram of the channel attention module is shown below. Figure 8 As shown.

[0081] In a preferred embodiment of the present invention, in the above-described bridge health status detection method based on a dual-branch Mamba network, the operation performed by the multi-head attention module includes: processing the enhanced temporal feature vector f i r The enhanced frequency domain feature vector f i d and the enhanced bimodal feature vector f i z Linear transformations are performed separately, mapping to different subspaces through matrix multiplication and bias operations to generate query, key, and value vectors. The similarity matrix between the query and key is calculated, adjusted by a scaling factor, and normalized using Softmax. The normalized result is then weighted and summed with the value vector to generate single-head attention features. Multiple single-head attention features are concatenated along the channel dimension to form a long vector. A linear transformation layer is then used to reduce the dimensionality and fuse the concatenated features to generate attention-weighted features. The multi-head attention module structure diagram is shown below. Figure 9 As shown.

[0082] In a preferred embodiment of the present invention, the bridge health status detection method based on a dual-branch Mamba network described above, wherein training and optimizing the dual-path lightweight network model using the training set includes: configuring the training environment, building a training framework under the Ubuntu 22.04 operating system, and using PyTorch. The 2.1.0 deep learning framework loads model parameters and training data onto the GPU via the torch.cuda module, leveraging GPU acceleration for training. Training parameters are set, defining the cross-entropy loss function as the optimization objective for model training. The AdamW optimizer is used for parameter updates, with an initial learning rate of 0.001, weight decay of 0.05, a batch size of 256, and a total of 300 iterations. The training set is input into the dual-path lightweight network model, and forward propagation is performed batch by batch to calculate the loss value. Backpropagation updates the network parameters, dynamically adjusting the learning rate to optimize convergence efficiency. The vibration field time profile is divided into training and testing samples in a 7:3 ratio. The pre-trained model is fine-tuned using the training samples, and the generalization ability and classification accuracy of the fine-tuned model are verified using the testing samples.

[0083] In a preferred embodiment of the present invention, the bridge health status detection method based on a dual-branch Mamba network described above, wherein inputting test samples into the trained dual-path lightweight network model, outputting bridge health status classification results, and evaluating model performance includes: inputting test samples into the trained dual-path lightweight network model, outputting bridge health status classification results; comparing the bridge health status classification results with the true labels of the test samples, calculating performance indicators, wherein the performance indicators include overall accuracy, evaluating model performance, specifically, evaluating the model's generalization ability and actual effect on the bridge health status recognition task.

[0084] A second embodiment of the present invention provides a bridge health status detection system based on a dual-branch Mamba network, comprising: a data acquisition module for acquiring and processing bridge vibration field data, generating a vibration field time profile and an instantaneous frequency map, and obtaining a dataset and a training set; a model building module for using the dataset as input to construct a dual-path lightweight network model for bridge health status identification; a training module for training and optimizing the dual-path lightweight network model using the training set; and a performance evaluation module for inputting test samples into the trained dual-path lightweight network model, outputting bridge health status classification results, and evaluating model performance.

[0085] The embodiments of the present invention aim to protect a bridge health status detection method and system based on a dual-branch Mamba network, which has the following effects:

[0086] 1. This invention uses a vibration sensor array to acquire bridge vibration field data and extracts features through vibration field time profiles and instantaneous frequency maps, enabling a comprehensive representation of the data's time and frequency domain information. It utilizes a dual-path lightweight network, combined with multimodal feature fusion (time and frequency domains) and an attention mechanism, to effectively capture key features of the bridge's health status and improve the accuracy of status identification. Generating vibration field data on a finite element model based on the finite difference method helps the model learn various bridge damage modes, improves the model's generalization ability, and makes it suitable for various bridge structures and working environments.

[0087] 2. The dual-branch Mamba network of this invention adopts a symmetrical structure and combines depthwise separable convolution with state space modeling to reduce computation, lower model complexity, and improve real-time detection capabilities. Input data is embedded at the 4×4 block level, combined with block feature extraction to improve the information utilization of the model, while reducing redundant computation and optimizing inference speed. Thanks to its lightweight architecture, the method can be adapted to embedded devices, such as bridge monitoring terminals, to achieve edge computing, reduce the need for high-performance computing equipment, and improve engineering practicality.

[0088] 3. This invention utilizes the PyTorch deep learning framework to improve training speed and shorten model optimization time through GPU parallel computing; it combines the AdamW optimizer with a learning rate decay strategy to improve gradient optimization efficiency, avoid local optima traps, and accelerate model convergence; it divides the vibration field data into a 7:3 ratio and fine-tunes the pre-trained model to improve the model's adaptability to different environments and enhance generalization performance.

[0089] 4. This invention uses a multi-head attention module in the feature fusion stage to improve the model's ability to perceive different damage features, especially to identify minor damage patterns; it uses a channel attention mechanism to adaptively allocate time-domain and frequency-domain feature weights, making the model more focused on key damage features and improving the robustness of anomaly detection; and it combines global average pooling and max pooling strategies to optimize the representation of damage features in spatial distribution, thereby improving the ability to detect damage to different types of bridge structures.

[0090] 5. This invention requires only ordinary vibration sensors and vehicle excitation sources, without the need for expensive specialized equipment such as laser measurement or large-scale sensor network deployment, significantly reducing engineering costs; the vibration sensor placement is flexible and can be adapted to different bridge structures, and the system is compatible with existing bridge monitoring platforms, enabling remote monitoring and data sharing; the entire process is based on deep learning to automatically complete the bridge health status assessment without manual intervention, improving detection efficiency and consistency.

[0091] 6. This invention is applicable to various bridge structures such as long-span bridges, urban viaducts, concrete bridges, and steel structure bridges. It has wide applicability and can identify early damage, fatigue cracks, structural loosening, corrosion, stiffness degradation and other problems in bridges. It provides real-time early warning for maintenance departments to avoid bridge accidents. It has strong anti-interference ability against the external environment and is suitable for bridge monitoring in complex environments.

[0092] The computer program product of the bridge health status detection method and device based on dual-branch Mamba network provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0093] Specifically, the storage medium can be a general-purpose storage medium, such as a portable disk or hard disk. When the computer program on the storage medium is run, it can execute the bridge health status detection method based on the dual-branch Mamba network, thereby enabling bridge health status detection without affecting the normal operation of traffic.

[0094] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0095] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, 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, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A bridge health status detection method based on a dual-branch Mamba network, characterized in that, include: Collect and process bridge vibration field data to generate vibration field time profiles and instantaneous frequency maps, thus obtaining datasets and training sets; Using the dataset as input, a dual-path lightweight network model for bridge health status identification is constructed. The training set is used to train and optimize the dual-path lightweight network model; The test samples are input into the trained dual-path lightweight network model, which outputs the bridge health status classification results and evaluates the model performance. The step of constructing a lightweight dual-path network model for bridge health status identification using the dataset as input includes: taking the seismic field time profile and the instantaneous frequency map as input data, processing them separately through symmetrical dual paths, with each path sequentially passing through a block embedding layer to divide the input image into 4×4 non-overlapping blocks; the block-based images undergo downsampling and feature extraction through alternating feature extraction modules and convolutional layers; the feature-extracted images are input into a feature fusion module, where channel concatenation is performed through element-wise addition, followed by a depthwise separable convolution; multimodal feature fusion is performed through a multi-head attention module and a channel attention module to generate high-quality fused features; and the high-quality fused features are input into a classifier to construct a lightweight dual-path network model for bridge health status identification. The operation performed by the feature fusion module includes: representing the temporal features extracted by the feature extraction module as follows: The frequency domain features are represented as Where i∈{1,2,3,4,5}, i represents the i-th feature extraction module; the temporal feature vector and frequency domain eigenvectors A concatenation operation is performed along the channel dimension to obtain a bimodal feature vector. ; The time-domain feature vector The frequency domain feature vector and the dual-modal feature vector Layer normalization is performed separately to accelerate model convergence. Depthwise separable convolutions are then applied sequentially to the normalized feature vectors, followed by SiLU activation to obtain enhanced temporal feature vectors. Enhanced frequency domain feature vectors and enhanced bimodal eigenvectors ; The enhanced time-domain feature vector The enhanced frequency domain feature vector and the enhanced bimodal feature vector Input a multi-head attention module to generate attention-weighted features; The time-domain feature vector and the frequency domain feature vector Input channel attention module to generate channel weights ω ; The features output by the multi-head attention module are multiplied element-wise with the weights generated by the channel attention module to obtain the time-domain weighted features. and frequency domain weighting The final fused feature is generated by adding elements one by one. .

2. The bridge health status detection method based on a dual-branch Mamba network according to claim 1, characterized in that, The process of collecting and processing bridge seismic field data to generate seismic field time profiles and instantaneous frequency maps results in a dataset and training set, including: Several nodes are set at equal intervals along the bridge and road, and a vibration sensor is installed at each node; The vibration generated by a vehicle moving at a constant speed is used as the excitation source, and vibration signals are acquired from each of the vibration sensors. The array data acquired from the vibration sensor is integrated to obtain a vibration field time profile. By using Fourier transform, the time domain of the vibration field time profile is transformed into the frequency domain to obtain the instantaneous frequency map. A finite element model is established, and the corresponding vibration field time profile and instantaneous frequency diagram are generated on the finite element model using the finite difference method, thus obtaining the dataset and training set.

3. The bridge health status detection method based on a dual-branch Mamba network according to claim 1, characterized in that, The operations performed by the feature extraction module include: The segmented image is divided into sub-feature map X1 and sub-feature map X2 by channel segmentation, and these are input into the convolution branch and the state space model branch, respectively. For the sub-feature map X1, normalization processing is performed. Local features are extracted through the first layer 3×3 convolution kernel, and then batch normalization and activation function operations are performed. Then, features are further extracted through the second layer 3×3 convolution kernel, and normalization and activation operations are performed. For the sub-feature map X2, normalization is performed, feature dimension transformation is performed through a linear layer, spatial correlation is extracted by deep convolution, and activation function is applied to enhance nonlinear expression capability. Based on the attention mechanism, the results of the convolutional branch and the state space model branch are fused and output.

4. The bridge health status detection method based on a dual-branch Mamba network according to claim 1, characterized in that, The operations performed by the channel attention module include: For the time-domain feature vector and the frequency domain feature vector Apply the SiLU activation function respectively; Perform global average pooling and global max pooling on the activated features to obtain the temporal pooling vector. and frequency domain pooling vector ; The time-domain pooling vector and the frequency domain pooling vector Feature transformations are performed using 1×1 convolution kernels, and the ReLU activation function is applied. The transformed features are then summed, and channel weights are generated using the Sigmoid function. ω .

5. The bridge health status detection method based on a dual-branch Mamba network according to claim 1, characterized in that, The operations performed by the multi-head attention module include: For the enhanced time-domain feature vector The enhanced frequency domain feature vector and the enhanced bimodal feature vector Perform linear transformations on each element, and map them to different subspaces through matrix multiplication and bias operations to generate query, key, and value vectors. Calculate the similarity matrix between the query and the key, adjust it with a scaling factor, and then normalize it using Softmax. Finally, perform a weighted summation of the normalized result and the value vector to generate a single-head attention feature. Multiple single-head attention features are concatenated along the channel dimension to form a long vector. The concatenated features are then dimensionality-reduced and fused using a linear transformation layer to generate attention-weighted features.

6. The bridge health status detection method based on a dual-branch Mamba network according to claim 1, characterized in that, The step of training and optimizing the dual-path lightweight network model using the training set includes: The training environment was configured, and the training framework was set up under the Ubuntu 22.04 operating system. The PyTorch 2.1.0 deep learning framework was used, and the model parameters and training data were loaded to the GPU through the torch.cuda module to accelerate the training. Set training parameters, define the cross-entropy loss function as the optimization objective for model training, use the AdamW optimizer for parameter updates, set the initial learning rate to 0.001, the weight decay to 0.05, the training batch size to 256, and the total number of iterations to 300. The training set is input into the dual-path lightweight network model, and the loss value is calculated by forward propagation in batches. The network parameters are updated by back propagation, the learning rate is dynamically adjusted, and the convergence efficiency is optimized. The vibration field time profile was divided into training samples and test samples at a 7:3 ratio. The pre-trained model is fine-tuned using the training samples, and the generalization ability and classification accuracy of the fine-tuned model are verified using the test samples.

7. The bridge health status detection method based on a dual-branch Mamba network according to claim 6, characterized in that, The process of inputting test samples into the trained dual-path lightweight network model, outputting bridge health status classification results, and evaluating model performance includes: Input the test samples into the trained dual-path lightweight network model and output the bridge health status classification results; The bridge health status classification results are compared with the true labels of the test samples to calculate performance indicators, including overall accuracy, and to evaluate the model performance.

8. A bridge health status detection system based on a dual-branch Mamba network, characterized in that, include: The data acquisition module is used to collect and process bridge vibration field data, generate vibration field time profile and instantaneous frequency map, and obtain dataset and training set; The model building module is used to take the dataset as input and build a dual-path lightweight network model for bridge health status identification. The training module is used to train and optimize the dual-path lightweight network model using the training set. The performance evaluation module is used to input test samples into the trained dual-path lightweight network model, output the bridge health status classification results, and evaluate the model performance. The bridge health status detection system based on dual-branch Mamba network is implemented based on the bridge health status detection method based on dual-branch Mamba network as described in any one of claims 1-7.

Images