Bridge structure damage assessment model construction method, damage assessment method and system
By constructing real tag data sets and filtering data sets, generating adversarial samples, training multiple detection models and adopting adversarial distillation methods, the problem of insufficient data dependence and generalization capabilities in bridge structure damage detection is solved, and the generalization capabilities of the model are improved and recognition effects.
Patent Information
- Application Number
- CN202410764758.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-14
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-06-14
AI Technical Summary
The existing CNN models have problems in the detection of bridge structure damage, excessive dependence on labeling data and insufficient generalization capabilities, especially in complex damage scenarios, which have poor recognition effects.
By constructing real label data sets and filtering data sets, we generate adversarial sample data sets, train multiple detection models, determine pseudo-labels, and train the bridge damage assessment model using adversarial distillation method to reduce dependence on labeled data and improve the model's generalization ability to transform.
The generalization capability of the bridge structure damage assessment model is improved, the dependence on labeled data is reduced, and the identification accuracy and robustness in complex damage scenarios are improved.
Smart Images

Figure CN118918430A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image or video recognition, and in particular to a bridge structure damage assessment model construction method, a damage assessment method and a system. Background Art
[0002] Bridges are key infrastructure in modern transportation systems, and their structural integrity is of great significance to public safety and economic activities. However, once natural disasters such as earthquakes occur, they may cause serious damage to bridge structures, affecting their normal use and even threatening public safety. Traditional bridge structure damage detection methods mainly rely on manual inspection, which is not only time-consuming and laborious, but also easily affected by human factors, resulting in insufficient accuracy and reliability of the detection results.
[0003] In recent years, with the development of computer vision and deep learning technologies, automated bridge damage detection methods based on image processing and machine learning have received widespread attention. These methods usually use convolutional neural networks (CNNs) to analyze bridge images and identify possible damage areas. However, existing CNN models still have some problems when dealing with complex damage scenarios, including difficulties in data collection, dependence on labeled data, and insufficient generalization capabilities of models for different environments. Summary of the invention
[0004] In view of the shortcomings of the existing technology, the present invention proposes a bridge structure damage assessment model construction method, a damage assessment method and a system, which can ensure the generalization ability of the bridge structure damage assessment model while reducing the dependence on labeled data. The specific technical solution is as follows:
[0005] In a first aspect, a method for constructing a bridge structure damage assessment model is provided. In a first implementable manner of the first aspect, the method comprises:
[0006] Constructing a real label data set and a screening data set, wherein the real label data set includes a plurality of different types of building structure damage data with labels, and the screening data set includes a plurality of unlabeled bridge structure damage data;
[0007] Generate multiple adversarial samples based on the real label data set to build an adversarial sample data set;
[0008] Performing detection model training using the real label data set to obtain multiple different detection models;
[0009] The bridge structure damage data in the screening data set are evaluated by each detection model respectively, and the pseudo labels corresponding to each bridge structure damage data are determined, so as to construct a pseudo label data set;
[0010] Based on the pseudo-label dataset, the adversarial sample dataset and the real label dataset, the bridge damage assessment model is trained by using the adversarial distillation method.
[0011] In combination with the first implementable manner of the first aspect, in a second implementable manner of the first aspect, based on the real label data set, a generative adversarial network is used to generate the adversarial sample.
[0012] In combination with the first implementable manner of the first aspect, in a third implementable manner of the first aspect, performing detection model training using the real label data set includes:
[0013] A data enhancement method is used to enhance the real label data set.
[0014] In combination with the first implementable manner of the first aspect, in a fourth implementable manner of the first aspect, the bridge damage assessment model is trained using an adversarial distillation method, including:
[0015] Preliminary training of the bridge damage assessment model using the real label data set;
[0016] Performing secondary training on the initially trained bridge damage assessment model through the pseudo-label data set to transfer the knowledge of each detection model to the bridge damage assessment model;
[0017] Through the adversarial sample dataset and the pseudo-label dataset, an adversarial distillation method is used to perform adversarial training on the secondary trained bridge damage assessment model.
[0018] In combination with the fourth implementable manner of the first aspect, in a fifth implementable manner of the first aspect, KL divergence is used as the loss function of adversarial training, and the specific calculation formula of the KL divergence is as follows:
[0019]
[0020] Among them, P is the output probability distribution of the detection model, Q is the output probability distribution of the bridge damage assessment model, and i is the category index.
[0021] In combination with the first implementable manner of the first aspect, in a sixth implementable manner of the first aspect, further comprising:
[0022] A small amount of labeled bridge structure damage data corresponding to the bridges to be evaluated is obtained to build a data set, and the trained bridge damage assessment model is fine-tuned using the data set.
[0023] In a second aspect, a bridge structure damage assessment method is provided, comprising:
[0024] Using the bridge structure damage assessment model construction method as described in any one of the first to sixth possible implementations of the first aspect, a bridge structure damage assessment model corresponding to the bridge to be assessed is constructed;
[0025] Structural damage data corresponding to the bridge to be evaluated is obtained, and the damage condition of the bridge to be evaluated is evaluated using the bridge structure damage evaluation model.
[0026] In a third aspect, a bridge structure damage assessment model construction system is provided. In a first achievable manner of the third aspect, the system comprises:
[0027] A data preparation module is configured to construct a real label data set and a screening data set, wherein the real label data set includes a plurality of different types of building structure damage data with labels, and the screening data set includes a plurality of unlabeled bridge structure damage data;
[0028] A sample generation module, configured to generate a plurality of adversarial samples based on the real label data set to construct an adversarial sample data set;
[0029] A model training module is configured to perform detection model training using the real label data set to obtain multiple different detection models;
[0030] A label recognition module is configured to evaluate the bridge structure damage data in the screening data set through each detection model, determine the pseudo label corresponding to each bridge structure damage data, and thus construct a pseudo label data set;
[0031] The model building module is configured to train the bridge damage assessment model using an adversarial distillation method based on the pseudo-label dataset, the adversarial sample dataset and the real label dataset.
[0032] In combination with the first implementable manner of the third aspect, in a second implementable manner of the third aspect, the model building module includes:
[0033] A preliminary training unit, configured to perform preliminary training on the bridge damage assessment model using the real label data set;
[0034] A secondary training unit is configured to perform secondary training on the initially trained bridge damage assessment model through the pseudo-label data set to transfer the knowledge of each detection model to the bridge damage assessment model;
[0035] The adversarial training unit is configured to perform adversarial training on the bridge damage assessment model that has undergone secondary training by using the adversarial sample data set and the pseudo-label data set and adopting an adversarial distillation method.
[0036] In a fourth aspect, a bridge structure damage assessment system is provided, comprising:
[0037] An assessment model construction module is configured to adopt the bridge structure damage assessment model construction method described in any one of the first to sixth possible implementations of the first aspect to construct a bridge structure damage assessment model corresponding to the bridge to be assessed;
[0038] The bridge damage assessment module is configured to obtain structural damage data corresponding to the bridge to be assessed, and assess the damage condition of the bridge to be assessed through the bridge structure damage assessment model.
[0039] Beneficial effects: By adopting the bridge structure damage assessment model construction method, damage assessment method and system of the present invention, based on the real label data set composed of structural damage data of various building structures, various different types of detection models for identifying building structure damage can be trained, and then the knowledge of these detection models can be transferred to the bridge structure damage assessment model by using the adversarial distillation method. In this way, while ensuring the generalization ability of the bridge structure damage assessment model, the dependence on the bridge structure damage annotation data can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] In order to more clearly illustrate the specific implementation of the present invention, the following will briefly introduce the drawings required for use in the specific implementation. In all the drawings, each element or part is not necessarily drawn according to the actual scale.
[0041] Figure 1 A flow chart of a method for constructing a bridge structure damage assessment model provided by an embodiment of the present invention;
[0042] Figure 2 A flow chart of a bridge structure damage assessment method provided by an embodiment of the present invention;
[0043] Figure 3 A system block diagram of a bridge structure damage assessment model building system provided by an embodiment of the present invention;
[0044] Figure 4 A system block diagram of a bridge structure damage assessment system provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0045] The following embodiments of the technical solution of the present invention are described in detail in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore only used as examples, and cannot be used to limit the protection scope of the present invention.
[0046] Embodiment 1
[0047] like Figure 1 The flowchart of the bridge structure damage assessment model construction method shown in FIG. 1 includes:
[0048] Step 1: construct a real label data set and a screening data set, wherein the real label data set includes a plurality of different types of building structure damage data with labels, and the screening data set includes a plurality of unlabeled bridge structure damage data;
[0049] Step 2: Generate multiple adversarial samples based on the real label data set to build an adversarial sample data set;
[0050] Step 3: training the detection model using the real label data set to obtain multiple different detection models;
[0051] Step 4: Evaluate the bridge structure damage data in the screening data set through each detection model, determine the pseudo label corresponding to each bridge structure damage data, and thus construct a pseudo label data set;
[0052] Step 5: Based on the pseudo-label dataset, the adversarial sample dataset and the real label dataset, the adversarial distillation method is used to train the bridge damage assessment model.
[0053] Specifically, first, we can obtain multiple labeled different building structures, such as roads, houses or bridges, and the corresponding building structure damage data to build a real label data set, and obtain multiple unlabeled bridge structure damage data to build a screening data set. The building structure damage data can be labeled building structure damage images, and the bridge structure damage data can be bridge structure damage images.
[0054] Since this embodiment adopts the transfer learning method, the real label data set, the screening data set, etc. can be expanded by using open source bridge structure damage data sets, building structure damage data sets, etc. to meet the requirements of large-scale data sets.
[0055] Then, multiple adversarial samples can be generated based on the constructed real label dataset, and an adversarial sample dataset can be constructed based on the generated adversarial samples. In the subsequent adversarial training process, the adversarial samples will produce different outputs through the detection model as the teacher model and the bridge damage assessment model as the student model, thus enhancing the robustness and generalization ability of the bridge damage assessment model.
[0056] Afterwards, a variety of models for identifying structural damage can be trained using real label data sets to obtain a variety of different detection models. In this embodiment, the model used to identify structural damage can select the EfficientNet-B3 network architecture. Compared with other versions, the EfficientNet-B3 network has a greater depth, width, and higher input resolution, thereby improving the expressiveness and accuracy of the model. By adjusting the width, depth, and resolution of the EfficientNet-B3 network, models of different sizes and complexities are generated for training.
[0057] Then, the structural damage in each unlabeled bridge structure damage image in the screening dataset is identified by each trained detection model, so as to determine the pseudo label corresponding to each unlabeled bridge structure damage image. Based on the pseudo label data identified by all detection models, a pseudo label dataset is constructed for subsequent adversarial distillation training.
[0058] Finally, based on the constructed real label dataset, pseudo label dataset and adversarial sample dataset, the adversarial distillation method is used to transfer the knowledge of various different detection models constructed in the early stage to the bridge damage assessment model as the student model. In this way, not only can the generalization ability of the trained bridge damage assessment model be ensured, but also the dependence of model training on bridge structure damage labeled data can be reduced.
[0059] In this embodiment, optionally, in step 2, the adversarial sample can be generated based on the real label data set by using a generative adversarial network. Specifically, the adversarial sample can be generated by using a generative adversarial network according to each real label data in the real label data set, thereby constructing an adversarial sample set.
[0060] The adversarial generative network can generate more realistic and diverse adversarial samples through adversarial training of the generative network and the discriminative network. It does not rely on the gradient information of the target model, making it more flexible in different scenarios and more scalable.
[0061] It should be understood that generating adversarial samples through a generative adversarial network is a prior art method, and its specific method steps are common knowledge in the field and will not be repeated here.
[0062] It should also be understood that this embodiment is only illustrated by using an adversarial generative network, but the present invention is not limited thereto, and other algorithms may also be used to generate adversarial samples, such as the fast gradient method.
[0063] In this embodiment, optionally, in step 3, performing detection model training using the real label data set includes: performing data enhancement on the real label data set using a data enhancement method.
[0064] Specifically, before training the detection model, data enhancement methods such as random flipping, ColorJitter, random rotation, Gaussian blur, noise and / or random perspective transformation can be used to enhance the sample data in the real label data set, thereby improving the robustness of the trained detection model and further enhancing the robustness of the bridge structure damage assessment model obtained through transfer learning.
[0065] It should be understood that data enhancement methods such as random flipping, ColorJitter, random rotation, Gaussian blur, noise and / or random perspective transformation are all existing technical means, and their specific method steps are common knowledge in the field and will not be repeated here.
[0066] In this embodiment, optionally, in step 4, the bridge damage assessment model is trained using an adversarial distillation method, including:
[0067] Preliminary training of the bridge damage assessment model using the real label data set;
[0068] Performing secondary training on the initially trained bridge damage assessment model through the pseudo-label data set to transfer the knowledge of each detection model to the bridge damage assessment model;
[0069] Through the adversarial sample dataset and the pseudo-label dataset, an adversarial distillation method is used to perform adversarial training on the secondary trained bridge damage assessment model.
[0070] Specifically, the training process of the bridge damage assessment model includes three steps: first, the bridge damage assessment model can be preliminarily trained through a real label data set so that the bridge damage assessment model can learn basic features. Then, the bridge damage assessment model that has learned the basic features can be retrained through a pseudo label data set to learn the knowledge of different types of detection models.
[0071] Finally, adversarial training is performed on the bridge damage assessment model that has learned the knowledge of the detection model through adversarial sample datasets and pseudo-label datasets. Adversarial samples can make the bridge damage assessment model face challenges different from the detection model. By calculating the KL divergence between the outputs of the detection model and the bridge damage assessment model, the parameters of the student model can be optimized to gradually approach the performance of the detection model, thereby transferring the knowledge of different detection models to the bridge damage assessment model, ensuring the generalization ability of the bridge damage assessment model, so that the bridge damage assessment model can also achieve good recognition effects when dealing with complex damage scenarios.
[0072] In this embodiment, optionally, during the training process, KL divergence can be used as the loss function of adversarial training, and minimizing KL divergence can be used as the optimization goal, so that the output of the bridge damage assessment model gradually approaches the output of the detection model. The specific calculation formula of the KL divergence is as follows:
[0073]
[0074] Among them, P is the output probability distribution of the detection model, Q is the output probability distribution of the bridge damage assessment model, and i is the category index.
[0075] In this embodiment, optionally, it further includes:
[0076] A small amount of labeled bridge structure damage data corresponding to the bridges to be evaluated is obtained to build a data set, and the trained bridge damage assessment model is fine-tuned using the data set.
[0077] Specifically, in this embodiment, the bridge damage assessment model can use a convolutional neural network. After the bridge damage assessment model is trained by the adversarial distillation method, a small amount of labeled bridge structure damage data of the bridge to be assessed can be obtained to build a data set, and the last convolution layer in the bridge damage assessment model is frozen. The last fully connected layer in the bridge damage assessment model is trained with the constructed data set to further ensure the generalization ability of the bridge damage assessment model.
[0078] Embodiment 2
[0079] like Figure 2 The flow chart of the bridge structure damage assessment method shown in FIG. 1 includes:
[0080] Step S1: using the above-mentioned bridge structure damage assessment model construction method to construct a bridge structure damage assessment model corresponding to the bridge to be assessed;
[0081] Step S2: Acquire structural damage data corresponding to the bridge to be evaluated, and evaluate the damage of the bridge to be evaluated using the bridge structure damage assessment model.
[0082] Specifically, first, the bridge structure damage assessment model construction method described in Example 1 can be used to construct a bridge structure damage assessment model for identifying bridge structure damage. Then, a bridge structure damage image of the bridge to be assessed can be obtained, and the bridge structure damage image can be input into the bridge structure damage assessment model. The bridge damage in the bridge structure damage image can be identified through the bridge structure damage assessment model. Since the bridge damage assessment model has a strong generalization ability, it can also achieve a good recognition effect for complex damage scenes.
[0083] Embodiment 3
[0084] like Figure 3 The system block diagram of the bridge structure damage assessment model construction system shown in FIG. 1 includes:
[0085] A data preparation module is configured to construct a real label data set and a screening data set, wherein the real label data set includes a plurality of different types of building structure damage data with labels, and the screening data set includes a plurality of unlabeled bridge structure damage data;
[0086] A sample generation module, configured to generate a plurality of adversarial samples based on the real label data set to construct an adversarial sample data set;
[0087] A model training module is configured to perform detection model training using the real label data set to obtain multiple different detection models;
[0088] a label recognition module, configured to evaluate the bridge structure damage data in the screening data set through each detection model, determine the pseudo label corresponding to each bridge structure damage data, and thus construct a pseudo label data set;
[0089] The model building module is configured to train the bridge damage assessment model using an adversarial distillation method based on the pseudo-label dataset, the adversarial sample dataset and the real label dataset.
[0090] Specifically, the construction system includes a data preparation module, a sample generation module, a model training module, a label recognition module and a model construction module. The data preparation module can obtain multiple labeled building structure damage data corresponding to different building structures, and use them to construct a real label data set. It can also obtain multiple unlabeled bridge structure damage data to construct a screening data set.
[0091] The sample generation module can generate multiple adversarial samples based on the constructed real label data set, and construct an adversarial sample data set based on the generated adversarial samples. The model training module can train multiple models for identifying structural damage through the real label data set, thereby obtaining multiple different detection models.
[0092] The label recognition module can identify the structural damage in each unlabeled bridge structure damage image in the screening data set through the trained detection models respectively, so as to determine the pseudo-label corresponding to each unlabeled bridge structure damage image, and construct a pseudo-label data set based on the pseudo-label data obtained by all detection models.
[0093] The model building module can use the adversarial distillation method to transfer the knowledge of various detection models constructed in the early stage to the bridge damage assessment model as the student model based on the constructed real label data set, pseudo label data set and adversarial sample data set. In this way, it can not only ensure the generalization ability of the trained bridge damage assessment model, but also reduce the dependence of model training on bridge structure damage labeled data.
[0094] In this embodiment, optionally, the model building module includes:
[0095] A preliminary training unit, configured to perform preliminary training on the bridge damage assessment model using the real label data set;
[0096] A secondary training unit is configured to perform secondary training on the initially trained bridge damage assessment model through the pseudo-label data set to transfer the knowledge of each detection model to the bridge damage assessment model;
[0097] The adversarial training unit is configured to perform adversarial training on the bridge damage assessment model that has undergone secondary training by using the adversarial sample data set and the pseudo-label data set and adopting an adversarial distillation method.
[0098] Specifically, the model building module includes a preliminary training unit, a secondary training unit and an adversarial training unit. The preliminary training unit can perform preliminary training on the bridge damage assessment model through a real label data set so that the bridge damage assessment model learns basic features. The secondary training unit can perform secondary training on the bridge damage assessment model that has learned basic features through a pseudo label data set to learn knowledge of different types of detection models.
[0099] The adversarial training unit can perform adversarial training on the bridge damage assessment model that has learned the detection model knowledge through adversarial sample datasets and pseudo-label datasets, thereby transferring the knowledge of different detection models to the bridge damage assessment model, ensuring the generalization ability of the bridge damage assessment model, and enabling the bridge damage assessment model to achieve good recognition effects when dealing with complex damage scenarios.
[0100] Embodiment 4
[0101] like Figure 4 The system block diagram of the bridge structure damage assessment system shown in FIG. 1 includes:
[0102] An assessment model building module is configured to use the above-mentioned bridge structure damage assessment model building method to build a bridge structure damage assessment model corresponding to the bridge to be assessed;
[0103] The bridge damage assessment module is configured to obtain structural damage data corresponding to the bridge to be assessed, and assess the damage condition of the bridge to be assessed through the bridge structure damage assessment model.
[0104] Specifically, the assessment system includes an assessment model construction module and a bridge damage assessment module. Among them, the assessment model construction module can adopt the bridge structure damage assessment model construction method described in Example 1 to build a bridge structure damage assessment model for identifying bridge structure damage. The bridge damage assessment module can obtain the bridge structure damage image of the bridge to be assessed, and input the bridge structure damage image into the bridge structure damage assessment model. The bridge damage in the bridge structure damage image can be identified through the bridge structure damage assessment model. Since the bridge damage assessment model has a strong generalization ability, it can also achieve a good recognition effect for complex damage scenes.
[0105] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.
Claims
1. A method for constructing a bridge structure damage assessment model, characterized in that: include: Constructing a real label data set and a screening data set, wherein the real label data set includes a plurality of different types of building structure damage data with labels, and the screening data set includes a plurality of unlabeled bridge structure damage data; Generate multiple adversarial samples based on the real label data set to build an adversarial sample data set; Performing detection model training using the real label data set to obtain multiple different detection models; The bridge structure damage data in the screening data set are evaluated by each detection model respectively, and the pseudo labels corresponding to each bridge structure damage data are determined, so as to construct a pseudo label data set; Based on the pseudo-label dataset, the adversarial sample dataset and the real label dataset, the bridge damage assessment model is trained by using the adversarial distillation method.
2. The method for constructing a bridge structure damage assessment model according to claim 1, characterized in that: Based on the real label data set, the adversarial sample is generated by using an adversarial generative network.
3. The method for constructing a bridge structure damage assessment model according to claim 1, characterized in that: The detection model training is performed using the real label data set, including: A data enhancement method is used to enhance the real label data set.
4. The method for constructing a bridge structure damage assessment model according to claim 1, characterized in that: The bridge damage assessment model is trained using an adversarial distillation method, including: Preliminary training of the bridge damage assessment model using the real label data set; Performing secondary training on the initially trained bridge damage assessment model through the pseudo-label data set to transfer the knowledge of each detection model to the bridge damage assessment model; Through the adversarial sample dataset and the pseudo-label dataset, an adversarial distillation method is used to perform adversarial training on the secondary trained bridge damage assessment model.
5. The method for constructing a bridge structure damage assessment model according to claim 4, characterized in that: KL divergence is used as the loss function of adversarial training. The specific calculation formula of the KL divergence is as follows: Among them, P is the output probability distribution of the detection model, Q is the output probability distribution of the bridge damage assessment model, and i is the category index.
6. The bridge structure damage assessment model construction method according to claim 1, characterized in that: Also includes: A small amount of labeled bridge structure damage data corresponding to the bridges to be evaluated is obtained to build a data set, and the trained bridge damage assessment model is fine-tuned using the data set.
7. A bridge structure damage assessment method, characterized in that: include: Using the bridge structure damage assessment model construction method as described in any one of claims 1 to 6, a bridge structure damage assessment model corresponding to the bridge to be assessed is constructed; Structural damage data corresponding to the bridge to be evaluated is obtained, and the damage condition of the bridge to be evaluated is evaluated using the bridge structure damage evaluation model.
8. A bridge structure damage assessment model construction system, characterized in that: include: A data preparation module is configured to construct a real label data set and a screening data set, wherein the real label data set includes a plurality of different types of building structure damage data with labels, and the screening data set includes a plurality of unlabeled bridge structure damage data; A sample generation module, configured to generate a plurality of adversarial samples based on the real label data set to construct an adversarial sample data set; A model training module is configured to perform detection model training using the real label data set to obtain multiple different detection models; A label recognition module is configured to evaluate the bridge structure damage data in the screening data set through each detection model, determine the pseudo label corresponding to each bridge structure damage data, and thus construct a pseudo label data set; The model building module is configured to train the bridge damage assessment model using an adversarial distillation method based on the pseudo-label dataset, the adversarial sample dataset and the real label dataset.
9. The bridge structure damage assessment model construction system according to claim 8, characterized in that: The model building module includes: A preliminary training unit, configured to perform preliminary training on the bridge damage assessment model using the real label data set; A secondary training unit is configured to perform secondary training on the initially trained bridge damage assessment model through the pseudo-label data set to transfer the knowledge of each detection model to the bridge damage assessment model; The adversarial training unit is configured to perform adversarial training on the bridge damage assessment model that has undergone secondary training by using the adversarial sample data set and the pseudo-label data set and adopting an adversarial distillation method.
10. A bridge structure damage assessment system, characterized in that: include: An assessment model construction module is configured to use the bridge structure damage assessment model construction method as described in any one of claims 1 to 6 to construct a bridge structure damage assessment model corresponding to the bridge to be assessed; The bridge damage assessment module is configured to obtain structural damage data corresponding to the bridge to be assessed, and assess the damage condition of the bridge to be assessed through the bridge structure damage assessment model.
Citation Information
Patent Citations
Method and system for training neural network model through adversarial learning and knowledge distillation
CN116171446A
Knowledge distillation-based high-precision lightweight bridge crack identification method
CN116994130A
Road damage inspection cross-domain detection method and system based on semi-supervised federated learning
CN117011563A
Method and system for training a neural network model using adversarial learning and knowledge distillation
US20230222353A1
