Structural damage detection method based on bridge digital twin group

Abstract

The invention discloses a structural damage detection method based on a bridge digital twinning group, and the method comprises the steps: building a general damage detection model through a multi-dimensional digital twinning technology: building a reference bridge model through Abaqus, and generating a digital twinning group containing a single / multiple damage scene; training the ethnic group by improving a WGAN-GP architecture, outputting a marked damage image by a generator, introducing a multi-scale attention module into a discriminator, and constructing an ethnic group characteristic spectrum (EFM) matrix through characteristic distillation; then collecting feature parameters to train a CNN model, and fusing features through SENet channel attention; through Gaussian white noise verification and n-fold cross optimization, the F1 score of the model is greater than or equal to 0.89 in a 40% noise environment, and the average accuracy rate reaches 95% or above. According to the technology, the limitation of single structure detection is broken through, the model generalization ability is improved by 300% and the detection efficiency is improved by 50% through a dynamic twin-driven convolutional network self-evolution mechanism, and complex scenes with + / -20% size change and 0-100% damage degree can be accurately identified; and an efficient anti-interference solution is provided for health monitoring of structures such as bridges, high-rise buildings and the like.

Description

Technical Field

[0001] This invention relates to a structural damage detection method based on bridge digital twin families, which enhances the compatibility of CNN models and thus obtains a more robust and widely applicable intelligent model for damage detection, belonging to the field of infrastructure health monitoring. Background Technology

[0002] Structural damage detection is a popular research area in structural health monitoring, aiming to identify and assess damage in structures and eliminate potential hazards to ensure reliable structural use. Traditional structural damage detection methods typically rely on a large number of sensors and acquisition systems, making data processing and analysis complex. In recent years, structural damage detection methods based on convolutional neural networks (CNNs) have received widespread attention and research because they can achieve accurate damage detection by extracting structural feature information without requiring a large number of sensors. However, existing CNN-based structural damage detection methods still face some challenges.

[0003] Most existing methods are designed for single-structure models. When the structure's geometry (e.g., changes in length and width), damage scenario (e.g., single damage evolving into multiple damage), or external excitation conditions change, the detection accuracy drops significantly. For example, some methods rely solely on data from a single accelerometer, failing to consider the multi-dimensional dynamic characteristics of the structure; others construct numerical models but fail to introduce random variables, making the models unable to adapt to complex scenarios such as geometric disturbances and noise interference commonly found in real-world engineering. Furthermore, traditional methods lack the ability to identify multiple damage scenarios and lack systematic verification of the model's noise robustness, making it difficult to meet the practical needs of industrial-grade health monitoring. Therefore, to overcome the limitations of traditional CNNs that only detect single, specific structures, a novel structural damage detection method is needed. This method should overcome the limitations and challenges of traditional CNNs, possess strong generalization capabilities, and comprehensively and effectively detect damage in complex structures, providing more accurate and effective data support for health monitoring and maintenance services in the industrial field.

[0004] A typical prior art example is CN116579214A, which discloses a three-dimensional visualization bridge pier monitoring system and method based on digital twins, relating to the field of bridge pier monitoring technology. This system includes: a sensing subsystem, a twin subsystem, a visualization subsystem, and an early warning feedback subsystem. The sensing subsystem includes a static information module and a dynamic information module, transmitting the sensed static and dynamic information to the twin subsystem. The twin subsystem is used to establish a finite element simulation model based on the real-time sensing information sent by the sensing subsystem and the solution generated by the early warning feedback subsystem. The visualization subsystem is used for visualization display operations. The early warning feedback subsystem is used to generate solutions and provide bridge pier safety warnings. This invention establishes a three-dimensional visualization model of the bridge pier based on digital twin technology, performs finite element simulation combined with real-time data and solutions, and provides early warning feedback on the simulation results, achieving multi-objective, all-round visualized real-time bridge pier monitoring.

[0005] For example, CN117387559A discloses a concrete bridge monitoring system and method based on digital twins. This system updates the model in real time by collecting bridge structural parameters from various monitoring nodes on the concrete bridge, and monitors the deformation of the concrete bridge by collecting site monitoring parameters. Simulation models are established separately and then linked to the digital twin model. Utilizing the information interaction function of the digital twin model, the simulation model is corrected more promptly and accurately based on the feedback deformation data, achieving the function of artificial intelligence-assisted safety prediction and providing quality assurance for construction. This invention links the bridge's information model to the digital twin, uses algorithms to predict the model's simulation of the bridge's internal force effects, and the digital twin model provides real-time feedback and control throughout the construction process, improving the information utilization rate in construction operations, ensuring project quality to a certain extent, and providing convenience for bridge construction management.

[0006] However, the above methods have problems such as insufficient accuracy, insufficient performance, and low computational efficiency. Based on these problems, the inventors proposed a structural damage detection method based on bridge digital twin populations. Summary of the Invention

[0007] To address the limitation of traditional CNNs in detecting only a single specific shape and structure, this invention provides a structural damage detection method based on bridge digital twin families. This method achieves cross-structure and interference-resistant damage detection through a dynamic twin-driven convolutional network self-evolution mechanism, specifically including two innovative modules: a twin family feature transfer algorithm and a dynamic convolutional kernel self-adaptation algorithm.

[0008] In the first aspect, this invention proposes a structural damage detection method based on bridge digital twin populations, comprising the following steps:

[0009] Step 1: Establish a numerical model of the benchmark bridge structure and define geometric parameters, material properties, and boundary conditions;

[0010] Furthermore, a numerical model of the benchmark bridge structure was established using Abaqus, and the structural geometric parameters (such as length L0, width B0, and height H0), material properties (elastic modulus E and density ρ) and boundary conditions were clearly defined as the benchmark template for the population model.

[0011] Step 2 introduces four-dimensional random variables: random scaling of geometric dimensions, random assignment of damage degree, noise disturbance of excitation force, and random selection of excitation point location, to generate a digital twin population containing N similar structures;

[0012] Furthermore, the geometric dimensions are randomly scaled by ±20%, the damage degree is randomly assigned a value of 0-100%, the excitation force is ±30%, the noise disturbance and the excitation point location are randomly selected as four-dimensional random variables to generate a digital twin population containing N similar structures.

[0013] Furthermore, the excitation point location is randomly selected from 10-15 four-dimensional random variables, preferably 13.

[0014] Step 3: Train the digital twin population by inputting bridge geometric parameters and damage parameters, and outputting an labeled damage image; introduce a multi-scale attention module and construct the population feature map EFM matrix through feature distillation technology;

[0015] Furthermore, the digital twin population is trained using an improved WGAN-GP architecture. The generator takes bridge geometric parameters (length / width / height) and damage parameters (location / degree) as input and outputs an labeled damage image. The discriminator introduces a multi-scale attention module to force the network to focus on key damage features such as crack edges and stress concentration. The population features are compressed into a matrix using feature distillation technology and used as pre-training weights for CNN initialization, thereby achieving cross-size damage knowledge transfer.

[0016] The loss function is:

[0017]

[0018] L GAN The total loss function of the generator consists of two parts: adversarial loss and gradient penalty term.

[0019] The adversarial loss for real samples is calculated by taking a sample x from the real data distribution Pdata and calculating the log-expected value of the discriminator D(x).

[0020] Generative sample adversarial loss: the loss on samples generated by the generator. (following the generating distribution Pgen), calculate the log-expected probability of the discriminator's misclassification.

[0021] The gradient penalty term ensures that the generator does not only output a few damage scenarios that are "easy to fool the discriminator" (such as generating only a single crack), but also covers diverse scenarios such as ±20% geometric size changes of the bridge and combinations of single / multiple damages. By smoothing the gradient, the problem of "violent oscillation and difficulty in convergence" in traditional GAN ​​training is solved, improving the generation efficiency of digital twin populations by 40%.

[0022] Where λ = 10 is the gradient penalty coefficient to ensure the diversity of generated damage features; the population features are visualized by TSNE dimensionality reduction to verify the cross-size transfer capability of the EFM matrix.

[0023] Furthermore, in the digital twin family, one or more scenarios of single damage, double damage, and triple damage are set, and damage is simulated by modifying the elastic modulus of the component. The model is then distributed to the CNN training set, validation set, and test set in an 8:1:1 ratio.

[0024] Step 4: Train a CNN model based on multi-dimensional feature parameters. The CNN model includes a parameterized convolutional kernel generation function: based on the population features in the EFM matrix, the kernel size (3×3 to 15×15) and weights are dynamically adjusted using a lightweight MLP network, where the formula is as follows: Where θ i，j σ is a learnable parameter, and σ is the Swish activation function. Three parallel convolutional branches are used: a 3×3 detail branch, a 7×7 semantic branch, and a 1×1 residual branch. Features are weighted and fused through the SENet channel attention mechanism.

[0025] Furthermore, K i，j The weights in the i-th row and j-th column of the generated convolution kernel form a convolution kernel matrix of variable size (from 3×3 to 15×15).

[0026] σ: Swish activation function, used to enhance the ability to represent nonlinear features and alleviate the gradient vanishing problem.

[0027] MLP: Lightweight Multilayer Perceptron (2 hidden layers, number of hidden units 64→32), responsible for mapping input features to convolutional kernel weights.

[0028] M EFM [k,:]: The k-th row feature vector (256-dimensional) extracted from the EFM matrix, containing the cross-size damage common features of the k-th bridge model in the digital twin population (such as crack propagation path, stress concentration factor, etc.).

[0029] θ i，j The learnable parameter vector (256-dimensional) is bound to the convolution kernel position (i, j) to inject personalized adjustment information into the weights at different positions, thereby improving the flexibility of feature extraction.

[0030] The vector concatenation operation concatenates the population feature vector M_{EFM}[k,:]MEFM[k,:] with the learnable parameters θi,j to form a 512-dimensional input vector, which is used as the input to the MLP.

[0031] It employs three parallel convolutional branches: a detail branch of 3×3, a semantic branch of 7×7, and a residual branch of 1×1, and uses the SENet channel attention mechanism to weighted fuse features.

[0032] Furthermore, the training process of the CNN model includes: taking one or more multi-dimensional features such as acceleration, displacement, mode shape, strain mode, and modal strain energy as input, using damage type as label, optimizing network parameters through k-fold cross-validation, and introducing a dropout layer to prevent overfitting.

[0033] Furthermore, the damage detection model can detect ±20% geometric size changes, 0-100% damage degree and single / multiple damage scenarios, and has an F1 score ≥0.89 in a 40% Gaussian white noise environment.

[0034] Step 5: The noise resistance of the proposed method is verified by adding Gaussian white noise of different intensities, and the robustness of the CNN is evaluated by using k-fold cross-validation. In one case, k = 10 in the k-fold cross-validation evaluation of the CNN.

[0035] A second aspect of the present invention is to provide a structural damage detection system based on a bridge digital twin family, wherein the system performs detection based on a structural damage detection method based on a bridge digital twin family as described in the first aspect.

[0036] Furthermore, the system is used for bridge health monitoring and maintenance services.

[0037] The beneficial effects of this invention are as follows:

[0038] This invention significantly improves the generalization ability and anti-interference performance of CNN models by extending the detection capability of CNNs from a single specific structure to the detection of similar structures.

[0039] By using twin population feature transfer and dynamic convolution kernel self-adaptation, it breaks through the traditional CNN's dependence on a single structure, which is different from the conventional technical route of "single digital twin + fixed finite element simulation" in existing technologies.

[0040] In tests on 500 bridge models with dimensional deviations of ±20%, the detection accuracy reached 92.3%, which is 13.8% higher than the existing single digital twin + fixed finite element simulation scheme; the F1 score remained at 0.89 at 40% noise intensity, which is significantly better than the traditional method's 0.65.

[0041] The number of model parameters is reduced by 62%, and the inference speed is increased by 2.3 times, solving the defect of "poor real-time performance of finite element simulation" in existing technologies.

[0042] This technology provides an effective method for structural damage detection and analysis, which can improve the efficiency of structural health monitoring and maintenance services in the industrial sector in practical applications. Attached Figure Description

[0043] Figure 1 Numerical model of the benchmark bridge structure;

[0044] Figure 2 The amplitude curve is random.

[0045] Figure 3 Distribution of excitation points and sampling points;

[0046] Figure 4 The locations of 13 observation flat steel bars;

[0047] Figure 5 This is a schematic diagram of k-fold cross-validation;

[0048] Figure 6 This is a schematic diagram of a structural damage detection method based on bridge digital twin populations. Detailed Implementation

[0049] The specific embodiments of the present invention are described in detail below with reference to the technical solutions and accompanying drawings.

[0050] Example 1

[0051] Reference Appendix Figure 1-6 A structural damage detection method based on bridge digital twin populations includes the following implementation steps:

[0052] Step 1: Establish a numerical model of the benchmark bridge structure using Abaqus.

[0053] The structural geometric parameters (length L0, width B0, height), material properties (elastic modulus E, density p), and boundary conditions are clearly defined as the baseline template for the population model.

[0054] A numerical model of the benchmark bridge structure was established using Abaqus (e.g., Figure 1 );

[0055] Step 2: Implement dynamic twin population generation and feature extraction using Python:

[0056] (1) Generate 1000 bridge models spanning different dimensions through the Abaqus secondary development interface, with geometric dimensions randomly scaled by ±20%;

[0057] (2) An improved WGAN-GP model is used to train the population model, and the generator loss function is:

[0058]

[0059] Where λ = 10 is the gradient penalty coefficient, ensuring the diversity of generated damage features;

[0060] (3) Visualize the population characteristics by using TSNE dimensionality reduction to verify the cross-size transfer capability of the EFM matrix;

[0061] Step 3: In these random numerical models (bridge population) generated in Step 2, different damage scenarios are set up. Some of these models are used to train the CNN, while others are used to evaluate the effectiveness of the CNN.

[0062] In this embodiment, based on the secondary development interface provided by the Abaqus platform, N random models are automatically built using Python code, and numerical simulations are performed on various damage scenarios (single damage scenario and multi-damage scenario) (automatically modifying the elastic modulus of the damaged component).

[0063] For a single damage scenario, each damaged flat steel bar represents a damage scenario. Under each damage scenario, the following parameters are collected: acceleration, displacement, and mode shape, as well as their derivatives: the first derivative of displacement, the second derivative of displacement, the first derivative of mode shape, and the second derivative of mode shape. The modal strain energy and strain modes of the flat steel bar corresponding to the sampling point are also collected.

[0064] Among N random models, the models are allocated according to a certain ratio. In this embodiment, the allocation ratio is 8:1:1, where 80% of the models are used to train the CNN, 10% are used to validate the CNN, and 10% are used to test the CNN. The main scenario is a single-damage scenario, where only one flat steel bar is damaged (in this embodiment, the location of the damaged flat steel bar is set as follows). Figure 4 Repeat the above process to simulate multi-damage scenarios. In this example, the multi-damage scenarios mainly include dual-damage and triple-damage scenarios. Compare the parameters collected under single, dual, and triple-damage scenarios to analyze the impact of the number of damages on the response.

[0065] Step 4, CNN model training

[0066] Using parameters acquired during the modeling and acquisition phases as CNN input and labels, a CNN model for recognizing different damage scenarios is trained. The Adam optimizer is used, with cross-entropy as the loss function, and the training period is set to 50 epochs. 10-fold cross-validation (e.g., ...) is employed. Figure 5 To evaluate the robustness of the model, test and validate the performance of CNN on new models, continuously improve and optimize the CNN model, and increase the recognition accuracy.

[0067] Step 5: The noise resistance of the proposed method is verified by adding Gaussian white noise of different intensities, and the robustness of the CNN is evaluated by using k-fold cross-validation.

[0068] Vibration signals from actual bridges often contain a certain amount of noise. To verify the noise resistance of the damage detection method based on digital twin populations, this embodiment also applies noise of different intensities to the test samples. Artificial noise is generated using the MATLAB built-in function wgn (Gaussian white noise) and added to the normal data. In this embodiment, noise intensities ranging from 5% to 40% (in increments of 5%) are used to illustrate the impact on the detection results.

[0069] Then k-fold cross-validation was used (e.g.) Figure 5 To evaluate network performance, k-fold cross-validation is often used to assess the robustness and robustness of network models.

[0070] The specific implementation process is as follows: all samples are divided into k parts, and one part is used as the validation sample each time, while the remaining k-1 parts are used as the training samples. Therefore, a complete k-fold cross-validation process requires k trials. In this example, k is set to 10.

[0071] In some embodiments, the present invention also claims protection for a structural damage detection system based on bridge digital twin populations, including a road defect intelligent identification module, the module being configured to assess road defects using a road defect intelligent identification and assessment method based on multimodal large model and instance segmentation algorithm in some embodiments.

[0072] In some embodiments, the present invention also claims protection for an electronic device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a structural damage detection method based on bridge digital twin populations in some embodiments.

[0073] In some embodiments, the present invention also claims protection for a computer-readable storage medium having stored thereon computer instructions that, when executed, perform a structural damage detection method based on a bridge digital twin population in some embodiments to detect structural damage to bridges.

[0074] The above embodiments are merely preferred exemplary embodiments of this application, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any development or substitution that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention shall fall within the protection and disclosure scope of the present invention.

Claims

1. A structural damage detection method based on bridge digital twin populations, characterized in that: Includes the following steps: Step 1: Establish a numerical model of the benchmark bridge structure and define geometric parameters, material properties, and boundary conditions; Step 2: Introduce four-dimensional random variables: random scaling of geometric dimensions, random assignment of damage degree, noise disturbance of excitation force, and random selection of excitation point location, to generate a digital twin population containing N similar structures. Step 3: Train the digital twin population by inputting bridge geometric parameters and damage parameters, and outputting an labeled damage image; introduce a multi-scale attention module and construct the population feature map EFM matrix through feature distillation technology; In the digital twin family, one or more scenarios of single damage, double damage, and triple damage are set, and damage is simulated by modifying the elastic modulus of the component. The model is allocated to the CNN training set, validation set, and test set in a ratio of 8:1:

1. Step 4: Train the CNN model based on multi-dimensional feature parameters; Step 5: Introduce Gaussian white noise to verify the anti-interference capability, and use k-fold cross-validation to evaluate the robustness of the model.

2. The structural damage detection method based on bridge digital twin populations according to claim 1, characterized in that: The generation process of the digital twin population includes: establishing a numerical model of the benchmark bridge structure through Abaqus, calling the Abaqus API interface through Python, and applying uniformly distributed random perturbations to the length, width, and height of the benchmark model respectively: the random perturbation range is ±20%; Apply a uniform, random assignment of values ​​from 0 to 100% to the degree of damage; A Gaussian white noise perturbation of ±30% is applied to the excitation force amplitude, and the excitation point is randomly selected from n observation flat steel bar positions.

3. The structural damage detection method based on bridge digital twin populations according to claim 1, characterized in that: A CNN model is trained based on one or more multi-dimensional feature parameters: acceleration, displacement, mode shape and its derived strain modes, and modal strain energy. The CNN model includes a parameterized convolutional kernel generation function: based on the population features in the EFM matrix, the kernel size (3×3 to 15×15) and weights are dynamically adjusted through a lightweight MLP network, as shown in the formula. This is the core of the dynamic convolution kernel self-adaptation algorithm in this invention, which is used to dynamically generate convolution kernel parameters based on the characteristics of the bridge digital twin family, so as to achieve adaptive detection across sizes and damage scenarios; K i，j The weights in the i-th row and j-th column of the generated convolutional kernel form a convolutional kernel matrix of variable size (from 3×3 to 15×15). σ: Swish activation function, used to enhance the ability to represent nonlinear features and alleviate the gradient vanishing problem; MLP: Lightweight multilayer perceptron with 2 hidden layers, the number of hidden units changes from 64 to 32, responsible for mapping input features to convolutional kernel weights; M EFM [k,:]: The 256-dimensional feature vector extracted from the kk-th row of the EMF matrix, containing the cross-size damage common features of the kk-th bridge model in the digital twin population. Crack propagation path, stress concentration factor, or one or more; θ i，j The learnable parameter vector has a dimension of 256 and is bound to the convolution kernel position (i, j). It is used to inject personalized adjustment information into the weights at different positions, improving the flexibility of feature extraction. The vector concatenation operation concatenates the population feature vector MEFM[k,:] with the learnable parameters θi,j into a 512-dimensional input vector, which is used as the input to the MLP. It employs three parallel convolutional branches: a detail branch of 3×3, a semantic branch of 7×7, and a residual branch of 1×1, and uses the SENet channel attention mechanism to weighted fuse features.

4. The structural damage detection method based on bridge digital twin populations according to claim 1, characterized in that: The damage detection model can detect ±20% geometric size changes, 0-100% damage degree and single / multiple damage scenarios, and has an F1 score ≥0.89 in a 40% Gaussian white noise environment.

5. The structural damage detection method based on bridge digital twin populations according to claim 1, characterized in that: The k-fold cross-validation is used to evaluate the robustness of the CNN, where k = 10.

6. The structural damage detection method based on bridge digital twin populations according to claim 1, characterized in that: Step 2 involves using Python to dynamically generate and extract features from twin populations: an improved WGAN-GP model is used to train the population model, and the generator loss function is: L GAN The generator's total loss function consists of two parts: adversarial loss and gradient penalty term. The adversarial loss for real samples is calculated by taking the log-expected value of the discriminator D(x) for a sample x in the real data distribution Pdata. Generative sample adversarial loss: the loss on samples generated by the generator. Given a generative distribution Pgen, calculate the expected logarithm of the discriminant's misclassification probability. The gradient penalty term ensures that the generator does not only output a few types of damage scenarios that are "easy to fool the discriminator"; Where λ = 10 is the gradient penalty coefficient to ensure the diversity of generated damage features; the population features are visualized by TSNE dimensionality reduction to verify the cross-size transfer capability of the EFM matrix.

7. The structural damage detection method based on bridge digital twin populations according to claim 1, characterized in that: In step 5, the intensity of the Gaussian white noise is 5-40%.

8. The structural damage detection method based on bridge digital twin populations according to claim 1, characterized in that: The excitation point location is randomly selected from 10-15 four-dimensional random variables.

9. A structural damage detection system based on bridge digital twin populations, characterized in that... The structural damage detection method based on bridge digital twin populations, as described in any one of claims 1-8, is used for detection.

10. A structural damage detection system based on bridge digital twin populations according to claim 9, characterized in that: Used for health monitoring or maintenance services for bridges.

Images