A design method and system for bridge RC pile reinforcement based on a two-stage diffusion model of mask information constraint
By using a two-stage diffusion model based on mask information constraints, the steel reinforcement layout diagram of bridge RC piles is automatically generated, solving the problem of low efficiency in existing technologies and realizing efficient and intelligent steel reinforcement design.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2026-03-18
- Publication Date
- 2026-07-10
AI Technical Summary
The design process for the reinforcement layout of RC piles in the substructure of bridges is inefficient, with insufficient automation and intelligence, resulting in high manpower consumption and low design efficiency.
A two-stage diffusion model based on mask information constraints is adopted. By obtaining design requirements, a mask matrix is generated and input into a pre-trained diffusion network model for iterative optimization. Combined with a variational autoencoder, a rebar layout drawing is generated and converted into a standard format through a CAD secondary development plugin.
The intelligent and automated arrangement of RC pile reinforcement has been achieved, which improves design efficiency, ensures that the generated results meet the structural geometric design constraints, and improves the quality and stability of the generated results.
Smart Images
Figure CN122365635A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of civil engineering and computer deep learning applications, specifically to the application of artificial intelligence in civil engineering structural design, and more specifically to a design method and system for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information. Background Technology
[0002] Currently, in the design phase of bridge substructures, the reinforcement of pile foundations includes various types of longitudinal bars and stirrups. The design process requires comprehensive consideration of design requirements and different combinations of reinforcement arrangements to ensure that the reinforcement arrangement method meets the various verification requirements of the design specifications. The design phase requires manual iterative calculations and repeated adjustments, with insufficient automation and intelligence. Meanwhile, the drafting phase is labor-intensive and inefficient.
[0003] Therefore, there is significant room for improvement in the efficiency of reinforcement design and drafting of RC piles (reinforced concrete piles), and there is an urgent need for an intelligent and efficient design method for RC pile reinforcement layout that takes into account design requirements and constraints. Summary of the Invention
[0004] To address the technical problems existing in the prior art, the purpose of this invention is to provide a design method and system for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information, in order to solve the problem of low efficiency in the reinforcement design stage of existing bridge substructure piles.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] A design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information includes the following steps:
[0007] S1: Obtain the reinforcement design requirements for the RC pile structure to be processed;
[0008] S2: Extract key information from the reinforcement design requirements of RC pile structures and perform matrix processing to generate a pile foundation design mask matrix to be input;
[0009] S3: Sample from random Gaussian noise to obtain the initial noise tensor. Input the pile foundation design mask matrix and the initial noise tensor into the pre-trained RC pile reinforcement diffusion network model constrained by mask information. Repeat the steps of the pre-trained RC pile reinforcement diffusion network model constrained by mask information until the iteration time step reaches the preset value to obtain the latent feature tensor of RC pile reinforcement design.
[0010] S4: Input the latent feature tensor of the RC pile reinforcement design in low-dimensional space into a pre-trained variational autoencoder to obtain the RC pile reinforcement layout design drawing in pixel space.
[0011] In step S3, the RC pile reinforcement diffusion network model constrained by mask information is formed by coupling the potential diffusion network model and the mask control network, and is trained based on the mask matrix sample data and the RC pile reinforcement design feature tensor label data.
[0012] As a preferred option, in step S1, the reinforcement design requirements for the RC pile structure include: design pile length, design pile diameter, design value of axial force combination, and design value of bending moment combination.
[0013] As a preferred embodiment, step S2 includes:
[0014] S21: Initialize a mask matrix where all elements are zero;
[0015] S22: Convert key information in the design requirements, including the design pile length and design pile diameter, into a length-to-pixel scale.
[0016] S23: Based on the geometric position of the pile in the two-dimensional coordinate system, the mask matrix is indexed and located, and the pixel element at the corresponding position is assigned a value of one to represent the pile design area, thereby obtaining the pile foundation design mask matrix.
[0017] As a preferred embodiment, step S3 includes:
[0018] S31: The initial noise tensor Z sampled randomly T , and pile foundation design mask matrix C i The predicted noise tensor N is obtained by inputting the RC pile reinforcement diffusion network model constrained by the mask information. T-1 The predicted noise tensor N T-1 For the initial noise tensor Z T Reconstruction and correction are performed to obtain the noise tensor Z from the previous time step. T-1 ;
[0019] S32: Repeat the above process T times, continuously changing the noise tensor Z. t Pile foundation design mask matrix C i The RC pile reinforcement diffusion network model, which is constrained by the input mask information, outputs the predicted noise tensor N. t-1 Continuously change the prediction noise tensor N t-1 For the initial noise tensor Z t Reconstruction and correction are performed; finally, the latent characteristic tensor of RC pile reinforcement design in low-dimensional space is obtained. .
[0020] As a preferred embodiment, step S4 includes: processing the latent characteristic tensor of the RC pile reinforcement design. The input is fed into the decoder module of the pre-trained variational autoencoder to obtain the RC pile reinforcement design feature tensor X0 in pixel space, and then converted into an RC pile reinforcement layout design drawing.
[0021] As a preferred embodiment, in step S3, the RC pile reinforcement diffusion network model constrained by mask information is formed by coupling the potential diffusion network model and the mask control network. The mask control network is connected to the potential diffusion network model in the intermediate feature layer through a zero convolution module.
[0022] The potential diffusion network model is a potential diffusion generation model pre-trained on the RC pile reinforcement diagram dataset.
[0023] The mask control network includes multiple convolutional layers initialized with zero convolutional modules, used to receive a mask matrix generated from the pile foundation design parameters and extract constraint features.
[0024] As a preferred option, in step S3, the RC pile reinforcement diffusion network model constrained by mask information is trained based on mask matrix sample data and RC pile reinforcement design feature tensor label data, specifically including:
[0025] For any given iteration, a time step t is randomly selected. The label data Z0 of the RC pile reinforcement design feature tensor is input into the noise generator. The noise level at time step t is sampled and added to obtain the noisy RC pile reinforcement design feature tensor Z at time step t. t and the real noise tensor The characteristic tensor Z of the reinforcement design for noisy RC piles t and pile foundation design mask matrix C i The RC pile reinforcement diffusion network model is input to the mask information constraint and outputs the predicted noise tensor N. t By calculating the gradient value of the loss function, the parameters of the RC pile reinforcement diffusion network model constrained by mask information are adjusted to minimize the prediction noise tensor N. t and the real noise tensor The differences.
[0026] As a preferred option, a design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information further includes step S5: converting the RC pile reinforcement layout design drawing in pixel space into a standard CAD reinforcement drawing in DWG format.
[0027] As a preferred option, step S5 includes calling the API interface of the CAD program and loading the CAD secondary development plugin.
[0028] A design system for bridge RC pile reinforcement based on a two-stage diffusion model with mask information constraints is provided, which implements a design method for bridge RC pile reinforcement based on a two-stage diffusion model with mask information constraints, including:
[0029] The mask matrix conversion unit is used to extract key information from the reinforcement design requirements of RC pile structures and perform matrix processing to generate the pile foundation design mask matrix to be input.
[0030] The diffusion network generation unit is used to input the pile foundation design mask matrix and the initial noise tensor into the pre-trained mask information constrained RC pile reinforcement diffusion network model, and output the predicted noise tensor.
[0031] The iterative calculation unit is used to repeatedly iterate the steps of generating the diffusion network unit until the iteration time step reaches a preset value, and obtain the potential characteristic tensor of RC pile reinforcement design.
[0032] The latent decoding unit is used to encode and decode the latent feature tensor of RC pile reinforcement design in low-dimensional space to the RC pile reinforcement design feature tensor in pixel space.
[0033] The CAD conversion unit is used to convert the pixel drawing of RC pile reinforcement design into a standard CAD reinforcement drawing in dwg format using CAD secondary development plugin commands.
[0034] The present invention has the following advantages:
[0035] This invention utilizes artificial intelligence to optimize the operating system and middleware, combined with emerging software and information technology services, to achieve intelligent management and automated generation of RC pile reinforcement layout. Through computer vision software, function libraries, and information system integration services, it can provide efficient data processing and condition control capabilities during the engineering design process, realizing the automated, intelligent, and systematic application of reinforcement design.
[0036] This invention constructs a pile foundation design mask matrix by matrixing the geometric parameters of the designed pile length and diameter, and encodes the design load as a feature vector. Both are embedded into the diffusion generation process, effectively simplifying the expression of complex physical information and achieving deep integration of prior engineering knowledge with the generative model, enhancing the rationality and controllability of the results. A generative framework coupling a latent diffusion network model and a mask control network is employed. The latent diffusion network model's denoising mechanism gradually restores the high-resolution RC pile reinforcement layout design drawing from initial Gaussian noise. The mask control network introduces the pile foundation design mask matrix for conditional adjustment during the inverse denoising process, ensuring that the generated results strictly meet the structural geometric design constraints, improving generation quality and stability. The RC pile reinforcement layout design drawing is input into CAD software via a CAD secondary development plugin, automatically converting the generated reinforcement layout results into the engineering standard format. This invention solves the problem of low reinforcement layout efficiency in existing bridge RC pile structures, achieving efficient and reliable intelligent reinforcement design. Attached Figure Description
[0037] Figure 1 This is a flowchart illustrating the design process of the method of this invention.
[0038] Figure 2This is a schematic diagram of converting the design requirements of step S2 into a pile foundation design mask matrix.
[0039] Figure 3 This is a schematic diagram of the RC pile reinforcement diffusion network model architecture constrained by mask information.
[0040] Figure 4 This is a schematic diagram of a module of a bridge RC pile reinforcement design system based on a two-stage diffusion model constrained by mask information.
[0041] Figure 5 This is a schematic diagram of the training data preprocessing for the RC pile reinforcement diffusion network model constrained by mask information.
[0042] Figure 6 This is a schematic diagram of the training of a one-stage potential diffusion network model.
[0043] Figure 7 This is a schematic diagram of the training of a two-stage mask information constrained RC pile reinforcement diffusion network model.
[0044] Figure 8 This is a schematic diagram showing the conversion of CAD drawings for RC pile reinforcement design. Detailed Implementation
[0045] The present invention will now be described in further detail with reference to specific embodiments.
[0046] Example 1
[0047] Figure 1 As shown, a design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information includes the following steps:
[0048] S1: Obtain the reinforcement design requirements for the RC pile structure to be processed;
[0049] S2: Extract key information from the reinforcement design requirements of RC pile structures and perform matrix processing to generate a pile foundation design mask matrix to be input;
[0050] S3: Sample from random Gaussian noise to obtain the initial noise tensor. Input the pile foundation design mask matrix and the initial noise tensor into the pre-trained RC pile reinforcement diffusion network model constrained by mask information. Repeat the steps of the pre-trained RC pile reinforcement diffusion network model constrained by mask information until the iteration time step reaches the preset value to obtain the latent feature tensor of RC pile reinforcement design.
[0051] S4: Input the latent feature tensor of the RC pile reinforcement design in low-dimensional space into a pre-trained variational autoencoder to obtain the RC pile reinforcement layout design drawing in pixel space.
[0052] In step S3, the RC pile reinforcement diffusion network model constrained by mask information is formed by coupling the potential diffusion network model and the mask control network, and is trained based on the mask matrix sample data and the RC pile reinforcement design feature tensor label data.
[0053] In this invention, by inputting the pile foundation design mask matrix and the initial noise tensor into a pre-trained RC pile reinforcement diffusion network model constrained by mask information, the latent feature tensor of RC pile reinforcement design is quickly generated, which greatly improves the design efficiency of the reinforcement layout stage of bridge pile foundation structure.
[0054] In step S1, the reinforcement design requirements for RC pile structures include: design pile length, design pile diameter, design value of axial force combination, and design value of bending moment combination.
[0055] In this invention, according to national bridge design specifications, the compressive and flexural bearing capacities of RC piles must be greater than the design values for axial force combination and bending moment combination, respectively. The design value for bending moment combination can generally be satisfied by the structural requirements specified in the specifications, while the magnitude of the axial force is related to the selection of the design pile diameter; the design pile diameter can indirectly represent the design value for axial force combination. Therefore, the key design requirements are the design pile length and the design pile diameter.
[0056] Figure 2 As shown, step S2 includes:
[0057] S21: Initialize a mask matrix where all elements are zero;
[0058] S22: Convert key information in the design requirements, including the design pile length and design pile diameter, into a length-to-pixel scale.
[0059] S23: Based on the geometric position of the pile in the two-dimensional coordinate system, the mask matrix is indexed and located, and the pixel element at the corresponding position is assigned a value of one to represent the pile design area, thereby obtaining the pile foundation design mask matrix.
[0060] Figure 3 As shown, step S3 includes:
[0061] S31: The initial noise tensor Z sampled randomly T and pile foundation design mask matrix C i The predicted noise tensor N is obtained by inputting the RC pile reinforcement diffusion network model constrained by the mask information. T-1 The predicted noise tensor N T-1 For the initial noise tensor Z T Reconstruction and correction are performed to obtain the noise tensor Z from the previous time step. T-1 ;
[0062] S32: Repeat the above process T times, continuously changing the noise tensor Z. tPile foundation design mask matrix C i The RC pile reinforcement diffusion network model, which is constrained by the input mask information, outputs the predicted noise tensor N. t-1 Continuously change the prediction noise tensor N t-1 For the initial noise tensor Z t Reconstruction and correction are performed; finally, the latent characteristic tensor of RC pile reinforcement design in low-dimensional space is obtained. .
[0063] Figure 1 As shown, step S4 includes: converting the latent characteristic tensor of the RC pile reinforcement design. The input is fed into the decoder module of the pre-trained variational autoencoder to obtain the RC pile reinforcement design feature tensor X0 in pixel space, and then converted into an RC pile reinforcement layout design drawing.
[0064] Figure 8 As shown, a design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information further includes step S5: converting the RC pile reinforcement layout design drawing in pixel space into a standard CAD reinforcement drawing in DWG format.
[0065] Figure 8 As shown, step S5 includes:
[0066] S51: Calls the API interface of the CAD program and loads the CAD secondary development plugin;
[0067] S52: Run the plugin conversion command to read the RC pile reinforcement pixel map and convert it into tensor elements. Scan the feature tensor element by element to identify different types of pixels.
[0068] S53: Convert pixel coordinates to actual size coordinates, perform geometric vectorization on the graphic, generate corresponding vector elements, and write all generated CAD entities into DWG file format.
[0069] Figure 3 The model framework is shown. The RC pile reinforcement diffusion network model constrained by mask information is composed of a potential diffusion network model and a mask control network. The mask control network is connected to the potential diffusion network model in the intermediate feature layer through a zero convolution module.
[0070] The potential diffusion network model is a potential diffusion generation model pre-trained on the RC pile reinforcement diagram dataset.
[0071] The mask control network includes multiple convolutional layers initialized with zero convolutional modules, used to receive a mask matrix generated from the pile foundation design parameters and extract constraint features.
[0072] Figure 5-7The diagram illustrates the model training process. The mask-constrained RC pile reinforcement diffusion network model is trained based on mask matrix sample data and RC pile reinforcement design feature tensor label data, specifically including:
[0073] For any given iteration, a time step t is randomly selected. The label data Z0 of the RC pile reinforcement design feature tensor is input into the noise generator. The noise level at time step t is sampled and added to obtain the noisy RC pile reinforcement design feature tensor Z at time step t. t and the real noise tensor The characteristic tensor Z of the reinforcement design for noisy RC piles t and pile foundation design mask matrix C i The RC pile reinforcement diffusion network model is input to the mask information constraint and outputs the predicted noise tensor N. t By calculating the gradient value of the loss function, the parameters of the RC pile reinforcement diffusion network model constrained by mask information are adjusted to minimize the prediction noise tensor N. t and the real noise tensor The differences.
[0074] The specific training process will be described below.
[0075] Figure 5 The image shows the preprocessing process of the dataset. Key image information includes: the pile body, the location and arrangement of various longitudinal bars and stirrups extracted from the RC pile reinforcement design drawings; key design requirements include: the design pile length and design pile diameter required by the engineer.
[0076] For each of the key image information points mentioned above, the drawing is scaled down to a tensor of 3*1536*768. Pixel elements containing key image information are assigned RGB values based on their attributes: elements with the attribute of rebar are assigned red (255, 0, 0), elements with the attribute of pile body are assigned gray (132, 132, 132), and elements without key image information are assigned white (255, 255, 255). This yields the RC pile reinforcement design feature tensor.
[0077] For each of the above key design requirements (design pile length, design pile diameter), the scale is converted, and a 1536*768 matrix is initialized. Pixels containing piles in the two-dimensional coordinate system are assigned a value of one, and all other elements are assigned a value of zero. This yields the pile foundation design mask matrix.
[0078] In this invention, the training process of the mask information-constrained RC pile reinforcement diffusion network model is divided into two stages. The first stage is the training of the potential diffusion generation model based on the RC pile reinforcement design feature tensor label data. The second stage is the training of the mask information-constrained RC pile reinforcement diffusion network model based on the mask matrix sample data and the RC pile reinforcement design feature tensor label data.
[0079] This implementation example Figure 6 In the first stage of training, the latent diffusion generative model is trained based on the label data of the RC pile reinforcement design feature tensor, which specifically includes:
[0080] For any iteration, a time step t is randomly selected (1≤t≤1000). The label data Z0 of the RC pile reinforcement design feature tensor is input into the noise generator. The noise level at time step t is sampled and noise is added to obtain the noisy RC pile reinforcement design feature tensor Z at time step t. t and the real noise tensor ;
[0081] The characteristic tensor Z of the reinforcement design of noisy RC piles t Input to the latent diffusion network model, output the predicted noise tensor N t ;
[0082] By calculating the gradient of the loss function, the parameters of the latent diffusion network model are adjusted to minimize the prediction noise matrix N. t and the real noise matrix The differences.
[0083] Repeat the above training process 10,000 times until the predicted noise matrix N is reached. t and the real noise matrix The differences converged.
[0084] like Figure 7 The second-stage training process involves training the mask-information-constrained RC pile reinforcement diffusion network model based on mask matrix sample data and RC pile reinforcement design feature tensor label data. Specifically, this includes:
[0085] The pre-trained latent diffusion network model is coupled with the mask control network, which is connected to the latent diffusion network model in the intermediate feature layer through several layers of zero convolutional modules.
[0086] For any iteration, a time step t (1≤t≤1000) is randomly selected. The label data Z0 of the RC pile reinforcement design feature tensor is input into the noise generator. The noise level of the sampled time step t is added to obtain the noisy RC pile reinforcement design feature tensor Z at time step t. t and the real noise tensor ;
[0087] The characteristic tensor Z of the reinforcement design of noisy RC piles t and pile foundation design mask matrix C i The RC pile reinforcement diffusion network model is input to the mask information constraint and outputs the predicted noise tensor N. t ;
[0088] By calculating the gradient value of the loss function, the parameters of the RC pile reinforcement diffusion network model constrained by mask information are adjusted to minimize the prediction noise matrix N. t and the real noise matrix The differences.
[0089] Repeat the above training process for 3000 rounds until the predicted noise matrix N is reached. t and the real noise matrix The differences converged.
[0090] The noise level of the noise generator is set to change linearly from 0.0015 to 0.0195, and the time step is set to 1000 steps.
[0091] Example 2
[0092] Figure 4 As shown, a design system for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information is presented, implementing the design method of Embodiment 1. The system includes:
[0093] The mask matrix conversion unit is used to extract key information from the reinforcement design requirements of RC pile structures and perform matrix processing to generate the pile foundation design mask matrix to be input.
[0094] The diffusion network generation unit is used to input the pile foundation design mask matrix and the initial noise tensor into the pre-trained mask information constrained RC pile reinforcement diffusion network model, and output the predicted noise tensor.
[0095] The iterative calculation unit is used to repeatedly iterate the steps of generating the diffusion network unit until the iteration time step reaches a preset value, and obtain the potential characteristic tensor of RC pile reinforcement design.
[0096] The latent decoding unit is used to encode and decode the latent feature tensor of RC pile reinforcement design in low-dimensional space to the RC pile reinforcement design feature tensor in pixel space.
[0097] The CAD conversion unit is used to convert the pixel drawing of RC pile reinforcement design into a standard CAD reinforcement drawing in dwg format using CAD secondary development plugin commands.
[0098] The detailed descriptions of each component of the system are as follows.
[0099] The mask matrix transformation unit initializes a mask matrix with all elements set to zero; it performs a length-to-pixel scale conversion on key information in the design requirements, including the design pile length and design pile diameter; it indexes and locates the mask matrix based on the geometric position of the pile in the two-dimensional coordinate system, assigning a value of one to the corresponding pixel element to represent the pile design area, thereby obtaining the pile foundation design mask matrix.
[0100] The diffusion network generation unit inputs the pile foundation design mask matrix and the initial noise tensor into a pre-trained RC pile reinforcement diffusion network model constrained by the mask information to obtain the latent feature tensor of RC pile reinforcement design. Specifically, this includes: inputting the randomly sampled initial noise tensor Z... T and pile foundation design mask matrix C i The predicted noise tensor N is obtained by inputting the RC pile reinforcement diffusion network model constrained by the mask information. T-1 The initial noise tensor Z T With the prediction noise tensor N T-1 By performing element-wise subtraction, we obtain the noise tensor Z from the previous time step. T-1 .
[0101] The iterative computation unit repeats the above process T times, continuously converting the noise tensor Z... t Pile foundation design mask matrix C i The noise tensor N of the RC pile reinforcement diffusion network model, which is constrained by the input mask information, is output. t-1 Continuously change the prediction noise tensor N t-1 For the initial noise tensor Z t Reconstruction and correction are performed. Finally, the latent characteristic tensor for RC pile reinforcement design in low-dimensional space is obtained. .
[0102] Latent decoding unit, used to decode the latent characteristic tensor of RC pile reinforcement design. The input is fed into the decoder module of the pre-trained variational autoencoder to obtain the RC pile reinforcement design feature tensor X0 in pixel space, and then converted into an RC pile reinforcement layout design drawing.
[0103] The CAD conversion unit calls the CAD API interface, loads the CAD secondary development plugin, runs the plugin conversion command, reads the RC pile reinforcement pixel map and converts it into tensor elements, scans the feature tensor element by element to identify different types of pixels, converts the pixel coordinates into actual size coordinates, performs geometric vectorization on the graphics, generates corresponding vector elements, and writes all generated CAD entities into DWG file format.
[0104] The bridge RC pile reinforcement design system based on a two-stage diffusion model constrained by mask information provided by this invention inputs the pile foundation design mask matrix and the initial noise tensor into a pre-trained RC pile reinforcement diffusion network model constrained by mask information, and quickly generates the latent feature tensor of RC pile reinforcement design, which greatly improves the design efficiency of the steel reinforcement layout stage of bridge pile foundation structure.
[0105] Install required library dependencies: Install relevant deep learning libraries in your computer environment, such as PyTorch and Transformers libraries, which include some pre-trained models.
[0106] Loading pre-trained language models: Use the Transformers library to load pre-trained models, such as GPT-Neo.
[0107] Process design requirements input by users / engineers, extract keywords such as pile length and pile diameter, save them in JSON format suitable for fine-tuning, and then use them to convert them into a mask matrix.
[0108] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information, characterized in that, Includes the following steps: S1: Obtain the reinforcement design requirements for the RC pile structure to be processed; S2: Extract key information from the reinforcement design requirements of RC pile structures and perform matrix processing to generate a pile foundation design mask matrix to be input; S3: Sample from random Gaussian noise to obtain the initial noise tensor. Input the pile foundation design mask matrix and the initial noise tensor into the pre-trained RC pile reinforcement diffusion network model constrained by mask information. Repeat the steps of the pre-trained RC pile reinforcement diffusion network model constrained by mask information until the iteration time step reaches the preset value to obtain the latent feature tensor of RC pile reinforcement design. S4: Input the latent feature tensor of the RC pile reinforcement design in low-dimensional space into a pre-trained variational autoencoder to obtain the RC pile reinforcement layout design drawing in pixel space. In step S3, the RC pile reinforcement diffusion network model constrained by mask information is formed by coupling the potential diffusion network model and the mask control network, and is trained based on the mask matrix sample data and the RC pile reinforcement design feature tensor label data.
2. The design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information as described in claim 1, characterized in that, In step S1, the reinforcement design requirements for RC pile structures include: design pile length, design pile diameter, design value of axial force combination, and design value of bending moment combination.
3. The design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information as described in claim 1, characterized in that, Step S2 includes: S21: Initialize a mask matrix where all elements are zero; S22: Convert key information in the design requirements, including the design pile length and design pile diameter, into a length-to-pixel scale. S23: Based on the geometric position of the pile in the two-dimensional coordinate system, the mask matrix is indexed and located, and the pixel element at the corresponding position is assigned a value of one to represent the pile design area, thereby obtaining the pile foundation design mask matrix.
4. The design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information as described in claim 1, characterized in that, Step S3 includes: S31: The initial noise tensor Z sampled randomly T , and pile foundation design mask matrix C i The predicted noise tensor N is obtained by inputting the RC pile reinforcement diffusion network model constrained by the mask information. T-1 The predicted noise tensor N T-1 For the initial noise tensor Z T Reconstruction and correction are performed to obtain the noise tensor Z from the previous time step. T-1 ; S32: Repeat the above process T times, continuously changing the noise tensor Z. t Pile foundation design mask matrix C i The RC pile reinforcement diffusion network model, which is constrained by the input mask information, outputs the predicted noise tensor N. t-1 Continuously change the prediction noise tensor N t-1 For the initial noise tensor Z t Reconstruction and correction are performed; finally, the latent characteristic tensor of RC pile reinforcement design in low-dimensional space is obtained. .
5. The design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information as described in claim 1, characterized in that, Step S4 includes: calculating the latent characteristic tensor of the RC pile reinforcement design. The input is fed into the decoder module of the pre-trained variational autoencoder to obtain the RC pile reinforcement design feature tensor X0 in pixel space, and then converted into an RC pile reinforcement layout design drawing.
6. The design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information as described in claim 1, characterized in that, In step S3, the RC pile reinforcement diffusion network model constrained by mask information is formed by coupling the potential diffusion network model and the mask control network. The mask control network is connected to the potential diffusion network model in the intermediate feature layer through a zero convolution module. The potential diffusion network model is a potential diffusion generation model pre-trained on the RC pile reinforcement diagram dataset. The mask control network includes multiple convolutional layers initialized with zero convolutional modules, used to receive a mask matrix generated from the pile foundation design parameters and extract constraint features.
7. The design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information as described in claim 4, characterized in that, In step S3, the RC pile reinforcement diffusion network model constrained by mask information is trained based on mask matrix sample data and RC pile reinforcement design feature tensor label data, specifically including: For any given iteration, a time step t is randomly selected. The label data Z0 of the RC pile reinforcement design feature tensor is input into the noise generator. The noise level at time step t is sampled and added to obtain the noisy RC pile reinforcement design feature tensor Z at time step t. t and the real noise tensor ; The characteristic tensor Z of the reinforcement design of noisy RC piles t and pile foundation design mask matrix C i The RC pile reinforcement diffusion network model is input to the mask information constraint and outputs the predicted noise tensor N. t ; By calculating the gradient value of the loss function, the parameters of the RC pile reinforcement diffusion network model constrained by mask information are adjusted to minimize the prediction noise tensor N. t and the real noise tensor The differences.
8. The design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information as described in claim 1, characterized in that, It also includes step S5: converting the RC pile reinforcement layout design drawing in pixel space into a standard CAD reinforcement drawing in DWG format.
9. The design method for bridge RC pile reinforcement based on a two-stage diffusion model constrained by mask information as described in claim 8, characterized in that, Step S5 includes calling the API interface of the CAD program and loading the CAD secondary development plugin.
10. A design system for bridge RC pile reinforcement based on a two-stage diffusion model with mask information constraints, implementing the design method for bridge RC pile reinforcement based on a two-stage diffusion model with mask information constraints as described in any one of claims 1 to 9, characterized in that, include: The mask matrix conversion unit is used to extract key information from the reinforcement design requirements of RC pile structures and perform matrix processing to generate the pile foundation design mask matrix to be input. The diffusion network generation unit is used to input the pile foundation design mask matrix and the initial noise tensor into the pre-trained mask information constrained RC pile reinforcement diffusion network model, and output the predicted noise tensor. The iterative calculation unit is used to repeatedly iterate the steps of generating the diffusion network unit until the iteration time step reaches a preset value, and obtain the potential characteristic tensor of RC pile reinforcement design. The latent decoding unit is used to encode and decode the latent feature tensor of RC pile reinforcement design in low-dimensional space to the RC pile reinforcement design feature tensor in pixel space. The CAD conversion unit is used to convert the pixel drawing of RC pile reinforcement design into a standard CAD reinforcement drawing in dwg format using CAD secondary development plugin commands.