Intelligent Assessment Method for Cable Damage in Long-Span Bridges

By combining a self-attention neural network with a multi-scale finite element model, the TractGraphFormer hybrid model is constructed, which solves the problems of accuracy and applicability in the assessment of sling stiffness loss in existing technologies, and realizes real-time and accurate assessment of sling damage and improves operation and maintenance efficiency.

CN123069760APending Publication Date: 2026-11-10CHONGQING UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611282909.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-24
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies cannot effectively integrate multi-source features, cannot achieve real-time and accurate assessment of sling stiffness loss, are difficult to balance structural safety and operation and maintenance efficiency, and rely on preset damage conditions, which are difficult to adapt to the uncertainty of damage parameters in actual engineering.

Method used

By employing a self-attention neural network and combining multi-source feature fusion with advanced neural network modeling, a TractGraphFormer hybrid model is constructed. This model, along with multi-scale finite element modeling and multi-dimensional damage feature extraction, enables real-time assessment of suspension cable stiffness loss.

Benefits of technology

It achieves high-precision, adaptive sling damage assessment, and can output stiffness loss and risk level in real time, improving the scientific nature and efficiency of bridge operation and maintenance. It is applicable to multiple bridge types and reduces labor costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069760A_ABST
    Figure CN123069760A_ABST
Patent Text Reader

Abstract

This application proposes an intelligent assessment method for cable damage in long-span bridges, belonging to the field of bridge structural health monitoring and safety assessment. Based on a self-attention neural network, this application achieves high-precision and adaptive assessment, providing technical support for intelligent operation and maintenance and full life-cycle safety management of bridges. The method includes collecting image and sensor data from multiple long-span cable-stayed bridges in a bridge group; after preprocessing such as denoising and standardization, constructing a "global-local" multi-scale finite element model; extracting multi-dimensional damage features, including static feature vectors based on Gaussian distribution function fitting, dynamic feature vectors based on FFT transform, and local damage features based on CNN; generating multiple damage conditions based on hyper-Latin cube sampling; and constructing a multi-source heterogeneous damage database by fusing measured and simulated data; training a TractGraphFormer hybrid model incorporating a self-attention mechanism; and deploying the model in the operation and maintenance system to achieve real-time output of cable stiffness loss and risk level.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application proposes a novel method based on a self-attention neural network, which can be widely applied to the stiffness degradation monitoring and damage assessment of suspension cables in long-span bridges such as suspension bridges, cable-stayed bridges, and basket arch bridges. This method belongs to the field of bridge structural health monitoring and safety assessment. Background Technology

[0002] In long-span bridges such as suspension bridges and cable-stayed bridges, the suspenders are the core load-bearing components, primarily responsible for the force transmission between the main cable and the stiffening girder. Their health directly determines the structural safety and service life of the bridge. Currently, with the increase in service life, suspenders are susceptible to damage such as corrosion, fatigue, and wire breakage due to factors such as vehicle loads, environmental erosion, and material aging. This can lead to a loss of stiffness, and if not assessed in a timely and accurate manner, it may cause major safety accidents such as bridge deck subsidence and structural collapse.

[0003] With the application of existing artificial intelligence technology, some research results have adopted neural networks for damage identification, but there are still the following obvious limitations: First, static and dynamic data are processed separately, making it impossible to take into account both local accuracy and global representation; second, relying on preset damage conditions and using classification algorithms to make predictions is difficult to adapt to the uncertainty of damage parameters in actual engineering; third, the relevant evaluation models have insufficient noise resistance and interpretability, making it difficult to meet the needs of engineering applications.

[0004] Existing technologies lack a systematic feature fusion and intelligent modeling framework, making it impossible to achieve real-time and accurate assessment of cable stiffness loss, and thus difficult to balance structural safety and operational efficiency. Therefore, there is an urgent need to develop an intelligent model and assessment method that integrates multi-source features, adaptive learning, and high-precision prediction to overcome the bottlenecks of existing technologies. In view of this, this application is hereby submitted. Summary of the Invention

[0005] The intelligent assessment method for cable damage in long-span bridges proposed in this application aims to address the problems existing in the prior art by using a self-attention neural network. Through multi-source feature fusion and advanced neural network modeling, it achieves high-precision and adaptive assessment, providing reliable technical support for intelligent operation and maintenance and full life-cycle safety management of bridges.

[0006] Therefore, the intelligent assessment method for cable damage in long-span bridges described in this application firstly collects image and sensor data from multiple long-span cable-stayed bridges in a bridge group. After preprocessing such as denoising and standardization, a multi-scale finite element model of "global-local" structure is constructed. Then, multi-dimensional damage features are extracted, including static feature vectors based on Gaussian distribution function fitting, dynamic feature vectors based on FFT transform, and local damage features based on CNN. Multiple damage conditions are generated based on hyper-Latin cube sampling, and a multi-source heterogeneous damage database is constructed by fusing measured and simulated data. Furthermore, a TractGraphFormer hybrid model with a self-attention mechanism is trained. Finally, the model is deployed in the operation and maintenance system to achieve real-time output of cable stiffness loss and risk level.

[0007] The method includes the following steps: Step (1): Data acquisition and preprocessing; Structural data of multiple long-span cable-stayed bridges in the bridge group were collected; The image data is processed by denoising, grayscale enhancement, and size normalization. Noise is removed using a combination of median filtering and Gaussian filtering, grayscale enhancement is achieved through histogram equalization, and the image size is normalized to a fixed resolution; let the original input image be... ,in Two-dimensional coordinates representing image pixels, , , The resolution of the original image is [resolution value], and the output image after preprocessing is [resolution value]. ; Step (2): Construction of multi-scale finite element model; Based on bridge design drawings, material testing reports and other data, the geometric dimensions, material properties and node connection details of the suspension cables are extracted; finite element simulation software such as MidasCivil and ANSYS are used to construct a "global-local" multi-scale finite element model, that is, to establish a global multi-scale finite element model of the structure and a local sub-model of the suspension cables. Step (3): Extraction of multi-dimensional damage features; This includes static feature vector extraction, dynamic feature vector extraction, and local damage feature extraction. Static feature vectors are obtained by calculating the difference in stiffening girder deflection between the damaged and healthy states of the suspension cables, and then extracting core parameters after fitting with an nth-order Gaussian distribution function. Dynamic feature vectors are obtained by performing FFT transformation on bridge deck vibration data and calculating the negative natural logarithm of the first three natural periods and the corresponding stiffness change rate. Local damage features are obtained by segmenting the suspension cable images using a CNN model and extracting shape and texture features. Step (4): Damage database construction; Based on hyper-Latin cube sampling, a three-dimensional damage condition is generated that covers the location, quantity, and degree of damage. The sampling process ensures that the samples are uniformly distributed in the parameter space. The mechanical response is simulated by substituting a multi-scale finite element model, and the simulation-extracted features and measured features are integrated as input parameters. The damage condition values ​​are used as output parameters. A database is constructed and divided into training and testing sets in an 8:2 ratio. Substitute each damage condition into the overall multi-scale finite element model and the local sub-model of the suspension cable constructed in step (2) to simulate the mechanical response of the bridge under different damage conditions, and extract the static feature vector, dynamic feature vector and local damage features of the corresponding conditions; integrate the three types of features obtained by actual measurement and the three types of features extracted by simulation to form the input feature matrix; Step (5), Intelligent Model Training; Construct a TractGraphFormer hybrid model that incorporates a self-attention mechanism and train it on the dataset; Step (6), Model Application and Operation and Maintenance Guidance; The trained TractGraphFormer hybrid model is deployed in a real bridge operation and maintenance system. The three types of features are input after real-time preprocessing, and the output is the evaluation result of the stiffness loss of the suspension cable.

[0008] Step (1) includes: (1a) Denoising by combining median filtering and Gaussian filtering; First, median filtering is applied to the original image to suppress impulse noise. The median filtering operator is denoted as... The filter window size is ( (For odd numbers), the image after median filtering is The formula is: (1)

[0009] Then, Gaussian filtering is applied to the median-filtered image to smooth out Gaussian noise. The Gaussian filtering operator is denoted as... The Gaussian kernel function is ( (where the standard deviation is Gaussian), the denoised image after filtering is The formula is: (2) (3)

[0010] in, This represents a two-dimensional convolution operation, enabling smoothing of the convolution between the filtering operator and the image; (1b) Histogram equalization enhances grayscale; Let the denoised image be The grayscale range is ( Gray levels, typically The execution steps are as follows: First, calculate the image grayscale histogram. grayscale Number of pixels appearing: (4)

[0011] Then, calculate the gray-level cumulative distribution function. : (5) Secondly, the histogram equalization mapping function The original grayscale level Mapped to enhanced grayscale levels : (6) in, This indicates rounding to the nearest integer, ensuring that the grayscale level is an integer. By normalizing the cumulative distribution function, the following image after grayscale enhancement is obtained. : ,when Time; (7) (1c) Image size normalization; Image after grayscale enhancement Normalized to a fixed resolution ( Coordinate mapping is achieved using bilinear interpolation. Let the normalized image coordinates be... , , The coordinate mapping relationship is as follows: (8) (9) in, Indicates rounding down. , Given the coordinate mapping values ​​between the original image and the normalized image, bilinear interpolation is used to calculate the normalized pixel values. : (10) in, , which are the interpolation weighting coefficients ( ),when or When taking the boundary pixel value ; (1d) The overall image preprocessing can be represented as a composite mapping relationship: (11) in, This represents the noise reduction operation combining median filtering and Gaussian filtering. This indicates a histogram equalization grayscale enhancement operation. This indicates the size normalization operation.

[0012] Step (1) involves outlier removal, smoothing filtering, and standardization of the sensor data; outliers are removed using the 3σ criterion, smoothed using moving average filtering, and data units are standardized using the Z-score method; this includes... (2a) The 3σ criterion is used to remove outliers; First, calculate the data sequence of each sensor. mean and standard deviation : (12) (13)

[0013] Then, according to the 3σ criterion, when the data sample satisfies When an outlier is detected, it is identified and removed, resulting in a valid data sequence after outlier removal. ( (for valid sample indexes), that is: (14) (2b) Smoothing process using moving average filtering; Let the size of the sliding window be... ( It is a positive integer, and ), for valid data sequences Perform a moving average filter to obtain a smoothed data sequence. The formula is: (15) in, When the window exceeds the data range, the boundary extension method (filling with boundary values) is used to ensure the integrity of the filtering. (2c) Z-score standardization and unified dimension; For the smoothed data sequence The Z-score standardization method is used to eliminate dimensional differences and obtain a standardized data sequence. The formula is: (16) in, For smoothed data The mean, Its standard deviation is calculated as follows: (17) (18)

[0014] (2d) Global mapping; The newly added data processing flow can be represented as a composite operation relationship: (19) in, This represents the outlier removal operation based on the 3σ criterion. This represents the moving average filtering operation. This represents the Z-score normalization operation.

[0015] Step (3) includes: Step (3.1), Static feature vector extraction; Based on the structural mechanics theory and modal analysis method of suspension cables, firstly, the reference value of the stiffening beam deflection under the healthy state of the suspension cable is calculated using a finite element model. Then, the measured or simulated values ​​of the deflection under different damage conditions are calculated, and the difference in stiffening beam deflection between the damaged state and the healthy state is obtained as follows: Δw=w_d-w0 (20) in, w_d For the deflection under damage conditions, w 0 The deflection under the most unfavorable load condition in a healthy state; An nth-order Gaussian distribution function is used to fit the discrete deflection difference data. The fitting method is adapted to the damage condition: a first-order Gaussian function is used for single damage conditions, and a multi-order superposition form is used for multiple damage conditions. y = a·exp[-(xb)] 2 / (2c 2 )] (twenty one) y=Σa i ·exp[-(xb i ) 2 / (2c i 2 )] (twenty two) Where x is the coordinate along the length of the bridge, and y is the deflection difference; Extracting the core parameters of the Gaussian function, including magnitude. a mean b Standard deviation c This forms a three-dimensional static feature vector, enabling accurate representation of local damage features; Step (3.2), dynamic feature vector extraction; The preprocessed bridge deck vibration data were subjected to a Fast Fourier Transform to convert the time-domain signal into a frequency-domain signal, separating environmental interference noise from the effective structural vibration signal and identifying the first three natural frequencies of the structure. f 1 、f 2 、f 3 Using the relationship between frequency and period f=1 / T, calculate the first three natural periods. T 1 、T 2 、T 3 ; Based on the differential equations of vertical motion of suspension bridges, a correlation model between frequency and stiffness is established using the Galerkin method. An innovative quantitative relationship between frequency and stiffness changes before and after damage is proposed. This quantitative relationship possesses both sensitivity when the damage is small and stability when the damage is large, as shown below: (twenty three) Calculate the rate of change of stiffness and take the negative natural logarithm, then construct 3D dynamic eigenvectors corresponding to the first three modes. This enables a quantitative characterization of overall stiffness degradation. Step (3.3): Extraction of local damage features; For the preprocessed local image data of the sling, GraphCNN is used to locate and segment the sling region. The CNN model automatically learns the feature differences between the sling and the background through multi-layer convolution and pooling operations to achieve accurate extraction of the sling region. Fully connected layers and Softmax activation function are used to complete the damage recognition of the segmented region. Two types of local damage features were extracted from the segmented sling region.

[0016] Step (5) describes the TractGraphFormer hybrid model, which consists of an input layer, a CNN convolutional layer, a max pooling layer, a Transformer layer, an adaptive attention layer, a fully connected layer, and an output layer. The Adam optimization algorithm is used, and hyperparameters are set to train the model. RMSE, MAE, and R are then evaluated. 2 The accuracy of the model is evaluated by four indicators: RPD. Among them, the TractGraphFormer hybrid model aggregates features of non-Euclidean space topological relationships through two layers of EdgeConv operators, integrates spatial features by dimensionality reduction through one layer of Conv1d operator, captures temporal evolution patterns through multiple sets of TransformerLayers, enhances key damage information through residual Conv1d attention module, and finally outputs the prediction results by MLP network.

[0017] In step (5), the model training suppresses overfitting through Dropout and L2 regularization; the adjacency matrix AdjacentMatrix is ​​constructed using the cable damage feature vector as the node feature and the spatial positional relationship of the cable along the bridge length; through two layers of EdgeConv operators, the non-Euclidean spatial topological relationship of the cable image is directly aggregated to accurately capture the stiffness loss transmission effect between adjacent cables; then, the spatial features are reduced and integrated through one layer of Conv1d operator to output N×96-dimensional graph structure features.

[0018] This application also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: when the processor executes the program, it implements the above-mentioned intelligent assessment method for damage to suspension cables of long-span bridges.

[0019] This application also proposes a computer-readable storage medium storing a computer program, characterized in that: when the computer program is executed, it can realize the above-mentioned intelligent assessment method for cable damage of long-span bridges.

[0020] In summary, this application has the following advantages and beneficial effects compared with the prior art: 1. This application can perform a full-process analysis of time-varying stiffness degradation, multi-source characteristic response and intelligent prediction of suspension cables, thereby obtaining high-precision damage assessment results, providing scientific guidance for bridge operation and maintenance, and effectively making up for the technical defects of existing suspension cable stiffness loss assessment methods, such as single-dimensional analysis, reliance on preset working conditions, insufficient accuracy and lack of systematic modeling.

[0021] 2. This application has the advantages of wide adaptability and compatibility with multiple bridge types. It can evaluate the stiffness loss of the suspenders of different types of long-span cable-stayed bridges such as suspension bridges, basket arch bridges, and cable-stayed bridges without having to redesign the model for a single bridge type, which greatly improves the engineering applicability of the method.

[0022] 3. This application establishes a multi-source fusion and multi-dimensional feature system, which makes the evaluation more comprehensive and accurate. The constructed "static-dynamic-local" three-dimensional feature extraction system can accurately locate local damage. Dynamic features reflect the overall stiffness degradation, and local features capture surface damage details. The three types of features complement and fuse, which solves the defect of single feature information in the existing technology. Combined with the multi-scenario damage database constructed by hyper-Latin cube sampling, it can ensure the comprehensiveness of model training.

[0023] 4. This application establishes a multi-scale modeling and advanced network architecture, resulting in stronger generalization capabilities: It adopts a "global-local" multi-scale finite element model, taking into account both the overall bridge mechanical response and local damage details. It integrates the TractGraphFormer hybrid model, combining spatial feature extraction, temporal pattern capture, and key information enhancement capabilities, enabling rapid transfer to different bridge scenarios and further improving the method's engineering applicability.

[0024] 5. This application features full-process automation and standardization, resulting in higher operational efficiency. It can automate the entire process from data collection, preprocessing, feature extraction, model prediction to operational guidance without human intervention, significantly reducing labor costs.

[0025] 6. This application establishes a standardized definition of damage conditions, feature extraction process, and model evaluation system, which can be applied to the unified operation and maintenance management at the bridge group scale. It provides accurate data support for the safe operation and maintenance of bridges throughout their entire life cycle, avoids resource waste and safety risks caused by over-maintenance or untimely maintenance, and has significant economic and social benefits. Attached Figure Description

[0026] The implementation process of the present application will now be further explained and illustrated with reference to the following figures.

[0027] Figure 1 This is a flowchart of the intelligent assessment method for cable damage in long-span bridges described in this application; Figure 2 A schematic diagram of an existing long-span cable-stayed bridge and its detailed model; Figure 3 A comparison of deflection difference distribution curves under three damage conditions; Figure 4 The graph shows the Gaussian fitting effect of the deflection difference. Figure 5 This is a schematic diagram of the TractGraphFormer hybrid model structure; Figure 6 This is a schematic diagram of a dataset of multi-source heterogeneous bridge cable damage. Figure 7 Spatial distribution map of damage conditions in hyper-Latin cube sampling; Figure 8 Training graph for the surrogate model; Figures 9-1 to 9-3 These are schematic diagrams comparing the predicted results of the number, location, and extent of sling damage. Detailed Implementation

[0028] To better understand the above-mentioned objectives, features, and advantages of this application, the application will be further described below in conjunction with the accompanying drawings and embodiments. Many specific details are set forth in the following description to provide a thorough understanding of this application; however, this application may be implemented in other ways than those described herein, and therefore, this application is not limited to the specific embodiments disclosed below.

[0029] Example 1, such as Figures 1 to 9-3 As shown, this application proposes an intelligent assessment method for cable damage in long-span bridges. First, image and sensor data from multiple long-span cable-stayed bridges in a bridge group are collected. After preprocessing such as denoising and standardization, a multi-scale finite element model of "global-local" structure is constructed. Then, multi-dimensional damage features are extracted, including static feature vectors based on Gaussian distribution function fitting, dynamic feature vectors based on FFT transform, and local damage features based on CNN. Multiple damage conditions are generated based on hyper-Latin cube sampling, and a multi-source heterogeneous damage database is constructed by fusing measured and simulated data. Next, a TractGraphFormer hybrid model with a self-attention mechanism is trained, which can accurately capture the spatiotemporal correlation of features. Finally, the model is deployed in the operation and maintenance system to achieve real-time output of cable stiffness loss and risk level.

[0030] The method includes the following implementation steps: Step (1): Data acquisition and preprocessing; Structural data was collected from multiple long-span cable-stayed bridges within the bridge group; specifically, the bridge types included suspension bridges, basket arch bridges, and cable-stayed bridges. The structural data encompassed both image data and sensor data. Specifically, sensor data included static displacement data of the bridge deck and the first three natural frequencies of the bridge obtained from deck vibration; image data was acquired through high-definition cameras, drone inspections, and other methods to obtain local images of the suspension cables, focusing on key areas such as cable surfaces and connection nodes, capturing signs of damage such as broken wires and corrosion; sensor data was collected through the deployment of displacement sensors, strain sensors, and acceleration sensors to gather displacement, strain, and deck vibration data during bridge operation, reflecting the structural mechanical response state. The image data is processed by denoising, grayscale enhancement, and size normalization. Specifically, noise is removed by combining median filtering and Gaussian filtering, grayscale enhancement is achieved through histogram equalization, and the image size is normalized to a fixed resolution (e.g., 256×256 pixels), laying the foundation for subsequent feature extraction. Let the original input image be ,in Two-dimensional coordinates representing image pixels, , , The resolution of the original image is [resolution value], and the output image after preprocessing is [resolution value]. The entire preprocessing process is implemented through the following step-by-step formulas, laying the foundation for subsequent feature extraction; including: (1a) Denoising by combining median filtering and Gaussian filtering; First, median filtering is applied to the original image to suppress impulse noise. The median filtering operator is denoted as... The filter window size is ( (For odd numbers), the image after median filtering is The formula is: (1)

[0031] Then, Gaussian filtering is applied to the median-filtered image to smooth out Gaussian noise. The Gaussian filtering operator is denoted as... The Gaussian kernel function is ( (where the standard deviation is Gaussian), the denoised image after filtering is The formula is: (2) (3)

[0032] in, This represents a two-dimensional convolution operation, enabling smoothing of the convolution between the filtering operator and the image; (1b) Histogram equalization enhances grayscale; Let the denoised image be The grayscale range is ( Gray levels, typically The execution steps are as follows: First, calculate the image grayscale histogram. grayscale Number of pixels appearing: (4)

[0033] Then, calculate the gray-level cumulative distribution function. : (5) Secondly, the histogram equalization mapping function The original grayscale level Mapped to enhanced grayscale levels : (6) in, This indicates rounding to the nearest integer, ensuring that the grayscale level is an integer. By normalizing the cumulative distribution function, the following image after grayscale enhancement is obtained. : ,when Time; (7) (1c) Image size normalization; Image after grayscale enhancement Normalized to a fixed resolution ( Coordinate mapping is achieved using bilinear interpolation. Let the normalized image coordinates be... , , The coordinate mapping relationship is as follows: (8) (9) in, Indicates rounding down. , Given the coordinate mapping values ​​between the original image and the normalized image, bilinear interpolation is used to calculate the normalized pixel values. : (10) in, , which are the interpolation weighting coefficients ( ),when or When taking the boundary pixel value ; (1d) The overall image preprocessing can be represented as a composite mapping relationship: (11) in, This represents the noise reduction operation combining median filtering and Gaussian filtering. This indicates a histogram equalization grayscale enhancement operation. This indicates a size normalization operation; The sensor data is processed by outlier removal, smoothing filtering, and standardization. Outliers were removed using the 3σ criterion, smoothed using moving average filtering, and data was standardized using the Z-score method to eliminate the influence of differences between different sensors and different bridges. (2a) The 3σ criterion is used to remove outliers; First, calculate the data sequence of each sensor. mean and standard deviation : (12) (13)

[0034] Then, according to the 3σ criterion, when the data sample satisfies When an outlier is detected, it is identified and removed, resulting in a valid data sequence after outlier removal. ( (for valid sample indexes), that is: (14) (2b) Smoothing process using moving average filtering; Let the size of the sliding window be... ( It is a positive integer, and ), for valid data sequences Perform a moving average filter to obtain a smoothed data sequence. The formula is: (15) in, When the window exceeds the data range, the boundary extension method (filling with boundary values) is used to ensure the integrity of the filtering. (2c) Z-score standardization and unified dimension; For the smoothed data sequence The Z-score standardization method is used to eliminate dimensional differences and obtain a standardized data sequence. The formula is: (16) in, For smoothed data The mean, Its standard deviation is calculated as follows: (17) (18)

[0035] (2d) Global mapping; The newly added data processing flow can be represented as a composite operation relationship: (19) in, This represents the outlier removal operation based on the 3σ criterion. This represents the moving average filtering operation. This represents the Z-score normalization operation; Step (2): Construction of multi-scale finite element model; Based on bridge design drawings, material testing reports, and other data, core parameters such as the geometric dimensions (diameter, length), material properties (elastic modulus, tensile strength), and node connection details (connection method, constraint conditions) of the suspension cables are extracted. Finite element simulation software such as MidasCivil and ANSYS were used to construct a "global-local" multi-scale finite element model, that is, to establish a global multi-scale finite element model of the structure and a local sub-model of the sling. The local sub-model of the sling includes a healthy sling and a damaged sling model with stiffness loss levels of 0-4. Among them, the overall multi-scale finite element model takes into account both the structural integrity of the entire bridge and computational efficiency, and performs reasonable meshing of the main components of the bridge such as the main beam, main cable, and suspension cable to simulate the overall mechanical response of the entire bridge under load. The local sub-model of the slings provides a detailed model of the slings and connection nodes. It adopts shell-solid and beam-solid sub-model techniques to extract local boundary conditions from the overall model and apply them to the sub-models to ensure the consistency of local and overall mechanical responses. For example, the local sub-model of the slings includes healthy slings and damaged slings with stiffness loss levels of 0-4, which can accurately simulate the structural response characteristics at different stiffness degradation stages. like Figure 2 As shown, based on the design drawings of three bridges, parameters such as the geometric dimensions, material properties, and node connection details of the suspension cables were extracted: the suspension bridge cable diameter is 50mm and the elastic modulus is 206GPa; the basket arch bridge cable diameter is 45mm and the elastic modulus is 206GPa; and the cable-stayed bridge cable diameter is 55mm and the elastic modulus is 206GPa. A comprehensive multi-scale finite element model was established using MidasCivil, and the local sub-models of the suspension cables were refined using ANSYS. Local boundary conditions were extracted using beam-solid sub-model technology to construct cable models in healthy and 0-4 level damage states. The overall model's natural frequency error with the actual bridge is ≤5%, verifying the model's effectiveness.

[0036] Step (3): Extraction of multi-dimensional damage features; This includes static feature vector extraction, dynamic feature vector extraction, and local damage feature extraction. Static feature vectors are obtained by calculating the difference in deflection between the damaged and healthy stiffening beams, and then fitting the core parameters using an nth-order Gaussian distribution function. Dynamic feature vectors are obtained by performing an FFT transform on the bridge deck vibration data and calculating the first three natural periods and the corresponding negative natural logarithm of the stiffness change rate. Local damage features are obtained by segmenting the cable images using a CNN model and then extracting shape and texture features. Specifically, this includes: Step (3.1), Static feature vector extraction; Based on the structural mechanics theory and modal analysis method of suspension cables, firstly, the reference value of the stiffening beam deflection under the healthy state of the suspension cable is calculated using a finite element model. Then, the measured or simulated values ​​of the deflection under different damage conditions are calculated, and the difference in stiffening beam deflection between the damaged state and the healthy state is obtained as follows: Δw=w_d-w0 (20) in, w_d For the deflection under damage conditions, w 0 The deflection under the most unfavorable load condition in a healthy state; An nth-order Gaussian distribution function is used to fit the discrete deflection difference data. The fitting method is adapted to the damage condition: a first-order Gaussian function is used for single damage conditions, and a multi-order superposition form is used for multiple damage conditions. y = a·exp[-(xb)] 2 / (2c 2 )] (twenty one) y=Σa i ·exp[-(xb i ) 2 / (2c i 2 )] (twenty two) Where x is the coordinate along the length of the bridge, and y is the deflection difference; Extracting the core parameter from the Gaussian function: magnitude a (Positively correlated with the degree of damage), mean b (Corresponding damage location coordinates), standard deviation c (Characterizing the extent of damage), forming a 3D static feature vector to achieve accurate characterization of local damage features, with a goodness-of-fit R. 2 ≥0.95 to ensure feature reliability; like Figure 3 As shown, the deflection difference of the stiffening girders in the damaged and healthy states of the three bridges is calculated. Static feature vectors are extracted using first-order or multi-order Gaussian function fitting. a,b,c ];like Figure 4 As shown, the goodness of fit R 2 All ≥0.96; Step (3.2), dynamic feature vector extraction; The preprocessed bridge deck vibration data were subjected to Fast Fourier Transform (FFT) to convert the time-domain signal into a frequency-domain signal, separating environmental interference noise from the effective structural vibration signal and identifying the first three natural frequencies of the structure. f 1 、f 2 、f 3 Using the relationship between frequency and period, f=1 / T, the first three natural periods are calculated. T 1 、T 2 、T 3 ; Based on the differential equations of vertical motion of suspension bridges, a correlation model between frequency and stiffness is established using the Galerkin method, clarifying the quantitative relationship between frequency changes and stiffness changes before and after damage as follows: (twenty three) Calculate the rate of change of stiffness and take the negative natural logarithm, then construct 3D dynamic eigenvectors corresponding to the first three modes. This enables a quantitative characterization of overall stiffness degradation. Step (3.3): Extraction of local damage features; For the preprocessed local image data of the sling, a graph convolutional neural network (GraphCNN) is used to locate and segment the sling region. The CNN model automatically learns the feature differences between the sling and the background through multi-layer convolution and pooling operations to achieve accurate extraction of the sling region. Fully connected layers and the Softmax activation function are used to complete the damage identification of the segmented region. Two types of local damage features were extracted from the segmented sling region: Among them, shape features include the number of broken wires, the length of broken wires, the corrosion area, and the perimeter of the corrosion region, which are obtained through morphological processing and contour extraction algorithms; texture features include the gray-level co-occurrence matrix parameters (energy, entropy, contrast) of the corrosion texture, surface roughness, etc., which are calculated through texture analysis algorithms to form a multi-dimensional local damage feature set, comprehensively capturing the surface damage state of the sling. Step (4): Damage database construction; Based on hyper-Latin cube sampling, a three-dimensional damage condition is generated that covers the location, quantity, and degree of damage. The sampling process ensures that the samples are uniformly distributed in the parameter space. The mechanical response is simulated by substituting a multi-scale finite element model, and the simulation-extracted features and measured features are integrated as input parameters. The damage condition values ​​are used as output parameters. A database is constructed and divided into training and testing sets in an 8:2 ratio. The location of the damage is expressed as a percentage of the total bridge length (1%-100%), covering the entire bridge cable arrangement area; The number of damages corresponds to the number of slings (1-10), covering scenarios involving damage to single or multiple slings; The degree of damage is divided into 5 levels using the tensile stiffness (EA) reduction factor: Level 0 (broken wire, reduction factor 0%), Level 1 (severe pitting of the cable, reduction factor 25%), Level 2 (moderate random pitting, reduction factor 50%), Level 3 (slight pitting, light to moderate corrosion, reduction factor 75%), and Level 4 (slight corrosion, reduction factor 100%), covering all scale damage scenarios from slight damage to complete failure. Substitute each damage condition into the overall multi-scale finite element model and the local sub-model of the suspension cable constructed in step (2) to simulate the mechanical response of the bridge under different damage conditions, and extract the static feature vector, dynamic feature vector and local damage features of the corresponding conditions. The three types of features obtained from actual measurements and the three types of features extracted from simulation are combined to form an input feature matrix; Using damage condition values ​​(damage quantity, location, and severity) generated by hyper-Latin cube sampling as output labels, a multi-source heterogeneous bridge cable damage database is constructed; for example... Figure 5 As shown, 100 damage conditions (damage location 1%-100%, number 1-10, severity 0-4) were generated based on hyper-Latin cube sampling. These conditions were then substituted into the multi-scale finite element models of the three bridges to extract 100 sets of simulation feature data. By integrating 20 sets of measured health data and 30 sets of measured damage data from the three bridges, a damage database containing 150 sets of samples was constructed. This database was divided into a training set (120 sets) and a test set (30 sets) in an 8:2 ratio. The specific database structure is as follows: Figure 6 As shown, the database sample size is no less than 120 sets (20 sets of measured data and 100 sets of finite element simulation damage condition data), which are divided into training set (for model parameter optimization) and test set (for model performance verification) in an 8:2 ratio to ensure the representativeness and diversity of the database and improve the generalization ability of the model. Step (5), Intelligent Model Training; Construct a TractGraphFormer hybrid model that incorporates a self-attention mechanism and train it on the dataset; like Figure 6 As shown, the TractGraphFormer hybrid model consists of an input layer, a CNN convolutional layer, a max pooling layer, a Transformer layer, an adaptive attention layer, a fully connected layer, and an output layer. It employs the Adam optimization algorithm, sets hyperparameters, and trains the model using RMSE, MAE, and R-squared values. 2 The accuracy of the model is evaluated using four indicators: RPD (Real-Time Performance). The TractGraphFormer hybrid model aggregates features of non-Euclidean spatial topological relationships through two layers of EdgeConv operators, integrates spatial features through one layer of Conv1d operators for dimensionality reduction, captures temporal evolution patterns through multiple sets of TransformerLayers, enhances key damage information through residual Conv1d attention modules, and finally outputs prediction results through an MLP network. The hyperparameters for model training are set to a maximum of 1000 iterations, an initial learning rate of 0.01, an adjustment of the learning rate by a decrease factor of 0.5 after 700 iterations, Dropout_rate=0.2, and L2 regularization λ=0.001. Model training uses Dropout and L2 regularization to suppress overfitting and ensure the model's generalization ability; Specifically, using the cable damage feature vector as node features, an adjacency matrix is ​​constructed based on the spatial positional relationships of the cables along the bridge length, breaking through the limitations of traditional CNN's "Euclidean space feature extraction". Through two layers of EdgeConv operators (each layer configured with a 64-dimensional multilayer perceptron mlp), the non-Euclidean space topological relationships of the cable images are directly aggregated to accurately capture the stiffness loss transmission effect between adjacent cables. Subsequently, a single layer of Conv1d operator (configured with a 96-dimensional mlp) is used to reduce the dimensionality of the spatial features and integrate them, outputting an N×96-dimensional graph structure feature. Based on global temporal evolution feature capture of Transformer layers, a 6-dimensional damage feature vector input to FeatureTokenizer can be converted into an N×96-dimensional feature sequence. Through multiple TransformerLayers (including multi-head attention mechanism, Add&Norm layer, and feedforward network), and by using the multi-head attention mechanism to simultaneously capture the stiffness loss evolution law at different time scales, combined with the feedforward network to enhance the nonlinear expression capability of features, the N×96-dimensional temporal correlation features are output, realizing global modeling of cable damage from "instantaneous state" to "full-cycle evolution". The adaptive attention layer enhances key damage information. Specifically, after the spatial features (assigned weight w1) output by GraphCNN and the temporal features (assigned weight w2) output by Transformer are weighted and fused in the feature fusion layer, an innovative "residual Conv1d attention module" is introduced: the fused features are mapped through two layers of Conv1d operators (each layer is configured with 32-dimensional MLP), the original feature information is preserved by combining residual connections, and dynamic attention weights are generated by the Sigmoid activation function to enhance the fused features. This automatically identifies key information such as "stiffness mutation points and multi-sling collaborative damage areas" and suppresses the interference of environmental noise and irrelevant features. Finally, the attention-weighted features are input into the multilayer perceptron (MLP) network, and the prediction results of the number, location, and degree of sling damage are output through the Softmax activation function, thus completing the reasoning process of the intelligent model. The Adam optimization algorithm was used to train the TractGraphFormer mixture model. The maximum number of iterations was set to 1000, with an initial learning rate of 0.01. When the number of iterations reached 700, the learning rate was adjusted by a decrease factor of 0.5. Overfitting was suppressed using Dropout (dropout_rate=0.2), and training continued until the model loss converged and stabilized. RMSE (Root Mean Square Error), MAE (Mean Absolute Error), and R-squared values ​​were used as the benchmarks. 2The model accuracy was evaluated using the coefficient of determination (COD) and relative prediction deviation (RPD) metrics. The effectiveness of each innovative module was verified using ablation experiments: removing the GraphCNN submodule increased the RMSE for lesion location prediction by 92.4%; removing the Transformer submodule increased the RPD for lesion severity prediction by [missing data]. 2 The overall predicted RPD decreased by 25.1% after removing the attention module, a decrease of 18.7%, fully demonstrating the necessity and synergistic effect of each innovative module. The trained model was applied to actual bridges to achieve real-time assessment and risk warning of cable stiffness loss, forming a standardized intelligent assessment method and technical system to guide bridge operation and maintenance. Step (6), Model Application and Operation and Maintenance Guidance; The trained TractGraphFormer hybrid model is deployed in a real bridge operation and maintenance system. The three types of features are input after real-time preprocessing, and the output is the evaluation result of the stiffness loss of the suspension cable.

[0037] Specifically, during the bridge's service life, only the data acquisition equipment is needed to acquire cable image data and sensor data in real time. After preprocessing in step (1), static feature vectors, dynamic feature vectors, and local damage features are extracted and uploaded to the proxy prediction model. The model quickly outputs the cable stiffness loss assessment results (including the number, location, degree of damage, and corresponding stiffness loss value) and risk levels (low risk, medium risk, high risk, and extremely high risk), realizing real-time assessment and risk warning of cable stiffness loss. Based on the assessment results, targeted operation and maintenance suggestions are provided to the bridge operation and maintenance unit: maintain routine inspections under low-risk conditions, increase monitoring frequency under medium-risk conditions, carry out special inspections under high-risk conditions, and immediately take reinforcement or replacement measures under extremely high-risk conditions to guide the bridge's preventive maintenance and achieve a dynamic balance between structural safety and operation and maintenance efficiency.

[0038] Figure 9 shows the evaluation results of the model on the test set: Damage quantity: RMSE=0.72, MAE=0.58, R 2 =0.92, RPD=3.56, 88% of the error is concentrated within ±0.5 roots. Damage location: RMSE=2.35%, MAE=1.42%, R 2 =0.99, RPD=9.38, with errors concentrated within ±2%. Damage severity: RMSE=0.28, MAE=0.21, R 2 =0.91, RPD=3.25, classification accuracy >92%. All indicators met the engineering requirements, and the model demonstrated good performance on three different types of bridges, validating its generalization ability. The trained model was deployed on the operation and maintenance systems of the three bridges, and data was collected and preprocessed in real time to extract three types of features for input into the model.

[0039] This application also proposes a novel electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned intelligent assessment method for cable damage in long-span bridges.

[0040] This application also proposes a novel computer-readable storage medium storing a computer program that, when executed, enables the implementation of the aforementioned intelligent assessment method for cable damage in long-span bridges.

[0041] As described above, similar technical solutions can be derived from the solutions presented in the accompanying drawings and description, and all of them still fall within the scope of the claims of this application.

Claims

1. A smart assessment method for cable damage in long-span bridges, characterized in that: First, image and sensor data from multiple long-span cable-stayed bridges in the bridge group were collected. After preprocessing such as denoising and standardization, a multi-scale finite element model of "global-local" structure was constructed. The sensor data included static data: displacement data from bridge deck displacement sensors; dynamic data: natural vibration period data obtained through bridge vibration calculations; and local data: actual local damage characteristics of the suspension cables captured in photographs. Then, multi-dimensional damage features are extracted, including static feature vectors based on Gaussian distribution function fitting, dynamic feature vectors based on FFT transformation, and local damage features based on CNN. Multiple damage conditions are generated based on hyper-Latin cube sampling, and a multi-source heterogeneous damage database is constructed by integrating measured and simulated data. Then, a TractGraphFormer hybrid model incorporating a self-attention mechanism is trained; Finally, the model is deployed in the operation and maintenance system to achieve real-time output of sling stiffness loss and risk level.

2. The intelligent assessment method for cable damage in long-span bridges according to claim 1, characterized in that: Including, Step (1): Data acquisition and preprocessing; Structural data of multiple long-span cable-stayed bridges in the bridge group were collected; The image data is processed by denoising, grayscale enhancement, and size normalization. Noise is removed using a combination of median filtering and Gaussian filtering, grayscale enhancement is achieved through histogram equalization, and the image size is normalized to a fixed resolution; let the original input image be... ,in Two-dimensional coordinates representing image pixels, , , The resolution of the original image is [resolution value], and the output image after preprocessing is [resolution value]. ; Step (2): Construction of multi-scale finite element model; Based on bridge design drawings, material testing reports and other data, the geometric dimensions, material properties and node connection details of the suspension cables are extracted; finite element simulation software such as MidasCivil and ANSYS are used to construct a "global-local" multi-scale finite element model, that is, to establish a global multi-scale finite element model of the structure and a local sub-model of the suspension cables. Step (3): Extraction of multi-dimensional damage features; This includes static feature vector extraction, dynamic feature vector extraction, and local damage feature extraction; The static feature vector is obtained by calculating the difference in stiffening girder deflection between the damaged and healthy states of the suspension cable, and then extracting the core parameters after fitting with an nth-order Gaussian distribution function; the dynamic feature vector is obtained by performing an FFT transformation on the bridge deck vibration data and calculating the negative natural logarithm of the first three natural periods and the corresponding stiffness change rate; the local damage features are obtained by segmenting the suspension cable image using a CNN model and extracting shape and texture features. Step (4): Damage database construction; Based on hyper-Latin cube sampling, a three-dimensional damage condition is generated that covers the location, quantity, and degree of damage. The sampling process ensures that the samples are uniformly distributed in the parameter space. The mechanical response is simulated by substituting a multi-scale finite element model, and the simulation-extracted features and measured features are integrated as input parameters. The damage condition values ​​are used as output parameters. A database is constructed and divided into training and testing sets in an 8:2 ratio. Substitute each damage condition into the overall multi-scale finite element model and the local sub-model of the suspension cable constructed in step (2) to simulate the mechanical response of the bridge under different damage conditions, and extract the static feature vector, dynamic feature vector and local damage features of the corresponding conditions; integrate the three types of features obtained by actual measurement and the three types of features extracted by simulation to form the input feature matrix; Step (5), Intelligent Model Training; Build, model This includes training on the dataset; A multi-task hybrid recognition framework for cable-stayed bridge cable damage is constructed, integrating a self-attention mechanism and TractGraphFormer, to achieve joint intelligent assessment of cable damage location, quantity, and severity. The model employs a dual-branch parallel feature extraction structure. On one hand, global structural data of the cable-stayed bridge is input into the Transformer module, utilizing its self-attention mechanism to mine long-distance dependencies and temporal correlations between overall structural units, extracting a global mechanical state representation of the bridge. On the other hand, for local image data of cable components, a Graph Convolutional Neural Network (GCN) is introduced to mine local spatial texture features of surface corrosion, cracks, and other damage, achieving deep extraction of local visual information of components. Subsequently, global structural features and local visual features are fused to achieve complementarity of multi-source heterogeneous information. An attention module is introduced to adaptively weight and enhance the fused features, strengthening damage-sensitive features, suppressing redundant noise, and improving the model's ability to focus on key damage information. Finally, the attention-enhanced features are input into a multilayer perceptron, and through a multi-task learning strategy, three assessment results—cable damage location, quantity, and severity—are simultaneously output, achieving refined damage detection and health status assessment of cable-stayed bridge cables. Step (6), Model Application and Operation and Maintenance Guidance; The converged TractGraphFormer hybrid model was deployed in a real-world bridge operation and maintenance system. Inputting three types of feature data—bridge structure, component visual data, etc.—after real-time preprocessing, the model outputs quantitative assessment results of cable stiffness loss, accurately representing three key indicators: cable damage location, damage degree (stiffness degradation degree), and damage range. The constructed model exhibits good transfer and generalization capabilities, adaptable to cable health monitoring scenarios for similar cable-stayed bridges, providing reliable data support for bridge operation and maintenance decisions.

3. The intelligent assessment method for cable damage in long-span bridges according to claim 2, characterized in that: Step (1) includes, (1a) Denoising by combining median filtering and Gaussian filtering; First, median filtering is applied to the original image to suppress impulse noise. The median filtering operator is denoted as... The filter window size is ( (For odd numbers), the image after median filtering is The formula is: (1) Then, Gaussian filtering is applied to the median-filtered image to smooth out Gaussian noise. The Gaussian filtering operator is denoted as... The Gaussian kernel function is ( (where the standard deviation is Gaussian), the denoised image after filtering is The formula is: (2) (3) in, This represents a two-dimensional convolution operation, enabling smoothing of the convolution between the filtering operator and the image; (1b) Histogram equalization enhances grayscale; Let the denoised image be The grayscale range is ( Gray levels, typically The execution steps are as follows: First, calculate the image grayscale histogram. grayscale Number of pixels appearing: (4) Then, calculate the gray-level cumulative distribution function. : (5) Secondly, the histogram equalization mapping function The original grayscale level Mapped to enhanced grayscale levels : (6) in, This indicates rounding to the nearest integer, ensuring that the grayscale level is an integer. By normalizing the cumulative distribution function, the following image after grayscale enhancement is obtained. : ,when Time; (7) (1c) Image size normalization; Image after grayscale enhancement Normalized to a fixed resolution ( Coordinate mapping is achieved using bilinear interpolation. Let the normalized image coordinates be... , , The coordinate mapping relationship is as follows: (8) (9) in, Indicates rounding down. , Given the coordinate mapping values ​​between the original image and the normalized image, bilinear interpolation is used to calculate the normalized pixel values. : (10) in, , which are the interpolation weighting coefficients ( ),when or When taking the boundary pixel value ; (1d) The overall image preprocessing can be represented as a composite mapping relationship: (11) in, This represents the noise reduction operation combining median filtering and Gaussian filtering. This indicates a histogram equalization grayscale enhancement operation. This indicates the size normalization operation.

4. The intelligent assessment method for cable damage in long-span bridges according to claim 2, characterized in that: Step (1) involves outlier removal, smoothing filtering, and standardization of the sensor data; outliers are removed using the 3σ criterion, smoothed using moving average filtering, and data units are standardized using the Z-score method; this includes... (2a) The 3σ criterion is used to remove outliers; First, calculate the data sequence of each sensor. mean and standard deviation : (12) (13) Then, according to the 3σ criterion, when the data sample satisfies When an outlier is detected, it is identified and removed, resulting in a valid data sequence after outlier removal. ( (for valid sample indexes), that is: (14) (2b) Smoothing process using moving average filtering; Let the size of the sliding window be... ( It is a positive integer, and ), for valid data sequences Perform a moving average filter to obtain a smoothed data sequence. The formula is: (15) in, When the window exceeds the data range, the boundary extension method (filling with boundary values) is used to ensure the integrity of the filtering. (2c) Z-score standardization and unified dimension; For the smoothed data sequence The Z-score standardization method is used to eliminate dimensional differences and obtain a standardized data sequence. The formula is: (16) in, For smoothed data The mean, Its standard deviation is calculated as follows: (17) (18) (2d) Global mapping; The newly added data processing flow can be represented as a composite operation relationship: (19) in, This represents the outlier removal operation based on the 3σ criterion. This represents the moving average filtering operation. This represents the Z-score normalization operation.

5. The intelligent assessment method for cable damage in long-span bridges according to claim 2, characterized in that: Step (3) includes, Step (3.1), Static feature vector extraction; Based on the structural mechanics theory and modal analysis method of suspension cables, firstly, the reference value of the stiffening beam deflection under the healthy state of the suspension cable is calculated using a finite element model. Then, the measured or simulated values ​​of the deflection under different damage conditions are calculated, and the difference in stiffening beam deflection between the damaged state and the healthy state is obtained as follows: Δw=w_d-w0 (20) in, w_d For the deflection under damage conditions, w 0 The deflection under the most unfavorable load condition in a healthy state; An nth-order Gaussian distribution function is used to fit the discrete deflection difference data. The fitting method is adapted to the damage condition: a first-order Gaussian function is used for single damage conditions, and a multi-order superposition form is used for multiple damage conditions. y=a·exp[-(x-b) 2 / (2c 2 )](21) y=Σa i ·exp[-(x-b i ) 2 / (2c i 2 )] (22) Where x is the coordinate along the length of the bridge, and y is the deflection difference; Extracting the core parameters of the Gaussian function, including magnitude. a mean b Standard deviation c This forms a three-dimensional static feature vector, enabling accurate representation of local damage features; Step (3.2), dynamic feature vector extraction; The preprocessed bridge deck vibration data were subjected to a Fast Fourier Transform to convert the time-domain signal into a frequency-domain signal, separating environmental interference noise from the effective structural vibration signal and identifying the first three natural frequencies of the structure. f 1 、f 2 、f 3 Using the relationship between frequency and period f=1 / T, calculate the first three natural periods. T 1 、T 2 、T 3 ; Based on the differential equations of vertical motion of suspension bridges, a correlation model between frequency and stiffness is established using the Galerkin method, clarifying the quantitative relationship between frequency changes and stiffness changes before and after damage as follows: (23) Calculate the rate of change of stiffness and take the negative natural logarithm, then construct 3D dynamic eigenvectors corresponding to the first three modes. This enables a quantitative characterization of overall stiffness degradation. Step (3.3): Extraction of local damage features; For the preprocessed local image data of the sling, GraphCNN is used to locate and segment the sling region. The CNN model automatically learns the feature differences between the sling and the background through multi-layer convolution and pooling operations to achieve accurate extraction of the sling region. Fully connected layers and Softmax activation function are used to complete the damage recognition of the segmented region. Two types of local damage features were extracted from the segmented sling region.

6. The intelligent assessment method for cable damage in long-span bridges according to claim 2, characterized in that: Step (5) describes the TractGraphFormer hybrid model, which consists of an input layer, a CNN convolutional layer, a max pooling layer, a Transformer layer, an adaptive attention layer, a fully connected layer, and an output layer. The Adam optimization algorithm is used, and hyperparameters are set to train the model. RMSE, MAE, and R are then evaluated. 2 The accuracy of the model is evaluated using four indicators: RPD (Real-Time Performance). The TractGraphFormer hybrid model aggregates features of non-Euclidean spatial topological relationships through two layers of EdgeConv operators, integrates spatial features by reducing dimensionality through one layer of Conv1d operators, captures temporal evolution patterns through multiple sets of TransformerLayers, enhances key damage information through residual Conv1d attention modules, and finally outputs prediction results through an MLP network.

7. The intelligent assessment method for cable damage in long-span bridges according to claim 2, characterized in that: In step (5), model training suppresses overfitting through Dropout and L2 regularization; Using the cable damage feature vector as node features, an adjacency matrix AdjacentMatrix is ​​constructed based on the spatial positional relationship of the cables along the bridge length. Through two layers of EdgeConv operators, the non-Euclidean spatial topological relationship of the cable image is directly aggregated to accurately capture the stiffness loss transmission effect between adjacent cables. Then, the spatial features are reduced and integrated through one layer of Conv1d operator to output an N×96-dimensional graph structure feature.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the above-mentioned intelligent assessment method for cable damage in long-span bridges.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed, it can realize the above-mentioned intelligent assessment method for cable damage in long-span bridges.