US-12626029-B2 - Physics-enhanced data-driven method and device for intelligent structural design of shear wall building

US12626029B2US 12626029 B2US12626029 B2US 12626029B2US-12626029-B2

Abstract

The present application provides a physics-enhanced data-driven method and device for intelligent structural design of shear wall building, and the method includes: obtaining an architectural design image and a basic design condition text to be processed; inputting the architectural design image and the basic design condition text into a structural design model, and obtaining a structural design image; the structural design model being obtained by performing a structural design image generation capability training and a physical performance optimization training for a physics-enhanced data-driven generative adversarial network; and vectorizing the structural design image and the architectural design image to obtain a structural design result of shear wall building. The physics-enhanced data-driven method and device for intelligent structural design of shear wall building provided by the present application improve the efficiency and the reliability of structural design.

Inventors

XINZHENG LU
Wenjie LIAO
Zhe Zheng
Yuan Tian

Assignees

TSINGHUA UNIVERSITY

Dates

Publication Date: 20260512
Application Date: 20220314
Priority Date: 20210902

Claims (8)

1 . A physics-enhanced data-driven method for intelligent structural design of a shear wall building that meets requirements, comprising: obtaining an architectural design image and a basic design condition text to be processed; inputting the architectural design image and the basic design condition text into a structural design model, and obtaining a structural design image output by the structural design model; wherein, the structural design model is obtained by performing a structural design image generation capability training and a physical performance optimization training for a physics-enhanced data-driven generative adversarial network, based on sample data of architectural design image, sample data of basic design condition text corresponding to the sample data of architectural design image and sample data of structural design image corresponding to the sample data of architectural design image; and extracting vector data of shear wall components from the structural design image and vector data of building floor outline from the architectural design image, and then obtaining a structural design result of the shear wall building based on the vector data of the shear wall components and the vector data of building floor outline to check strength of shear wall components.
2 . The physics-enhanced data-driven method for intelligent structural design of the shear wall building according to claim 1 , wherein, a training process for the structural design model comprises: building the physics-enhanced data-driven generative adversarial network; inputting sample data with complete labels into the physics-enhanced data-driven generative adversarial network, and performing the structural design image generation capability training for the physics-enhanced data-driven generative adversarial network; wherein, the sample data with complete labels comprises sample data of architectural design image, sample data of basic design condition text corresponding to the sample data of architectural design image and sample data of structural design image corresponding to the sample data of architectural design image; performing a calculation of physical performance for the structural design image generated during a process of the structural design image generation capability training, and obtaining a calculation result of physical performance; building a structural physical performance prediction neural network, and performing a structural physical performance prediction capability training for the structural physical performance prediction neural network with training data of the structural design image generated during the process of structural design image generation capability training and the calculation result of physical performance, and obtaining a structural physical performance prediction model; calculating image data loss during the process of structural design image generation capability training, calculating corresponding physics loss by the structural physical performance prediction model, fusing the image data loss and the physics loss, to obtain a loss function of the physics-enhanced data-driven generative adversarial network; and performing, based on the loss function, an optimization training for the physics-enhanced data-driven generative adversarial network after the structural design image generation capability training, and obtaining the structural design model.
3 . The physics-enhanced data-driven method for intelligent structural design of the shear wall building according to claim 2 , wherein, the physics-enhanced data-driven generative adversarial network comprises: an image generator, which is configured to perform encoding and feature extraction for the sample data of architectural design image and the sample data of basic design condition text, respectively, to obtain an image feature and a text feature, fuse the image feature and the text feature and decode a feature fused by the image feature and the text feature, and generate the structural design image; and an image discriminator, which is configured to perform feature extraction and authenticity discrimination for the generated structural design image.
4 . The physics-enhanced data-driven method for intelligent structural design of the shear wall building according to claim 2 , wherein, the performing the calculation of physical performance for the structural design image generated during the process of the structural design image generation capability training, and obtaining the calculation result of physical performance, comprises: extracting the vector data of the shear wall components from the generated structural design image and the vector data of building floor outline from the sample data of architectural design image; calculating, based on the vector data of shear wall components from the generated structural design image and the vector data of the floor outline from the sample data of architectural design image, floor mass and floor stiffness, creating a mass matrix and a stiffness matrix of the shear wall building, and obtaining a multi-degree-of-freedom mechanics calculation model of the shear wall building; and performing, based on the multi-degree-of-freedom mechanics calculation model, a calculation of structural mechanical response and a performance analysis for the shear wall building, and obtaining the calculation result of physical performance.
5 . The physics-enhanced data-driven method for intelligent structural design of the shear wall building according to claim 2 , wherein, a process of building, training, and prediction of the structural physical performance prediction model comprises: building, based on a residual network structure, a structural physical performance prediction neural network configured for image feature extraction and physical performance prediction; performing a physical performance prediction capability training for the structural physical performance prediction neural network based on the structural design image generated during the process of the structural design image generation capability training and a corresponding calculation result of physical performance, and obtaining a structural physical performance prediction model; and performing, by the structural physical performance prediction model, a physical performance prediction for the structural design image generated during the process of the structural design image generation capability training, to obtain a corresponding physics loss.
6 . The physics-enhanced data-driven method for intelligent structural design of the shear wall building according to claim 2 , wherein, after performing, based on the loss function, the optimization training for the physics-enhanced data-driven generative adversarial network after the structural design image generation capability training, the method further comprises: performing a semi-supervised optimization training for the physics-enhanced data-driven generative adversarial network, based on sample data without complete labels and sample data with complete labels; wherein, the sample data without complete labels comprises the sample data of architectural design image and the sample data of basic design condition text.
7 . A non-transient computer-readable storage medium, having a computer program stored therein, wherein, when the program is executed by a processor, it causes the processor to implement the steps of the physics-enhanced data-driven method for intelligent structural design of the shear wall building according to claim 1 .
8 . A physics-enhanced data-driven device for intelligent structural design of a shear wall building that meets requirements, comprising a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein, when the program is executed by the processor, causes the processor to: obtain an architectural design image and a basic design condition text to be processed; input the architectural design image and the basic design condition text into a structural design model, and obtain a structural design image output by the structural design model; wherein, the structural design model is obtained by performing a structural design image generation capability training and a physical performance optimization training for a physics-enhanced data-driven generative adversarial network, based on sample data of architectural design image, sample data of basic design condition text corresponding to the sample data of architectural design image and sample data of structural design image corresponding to the sample data of architectural design image; and extract vector data of shear wall components from the structural design image and vector data of building floor outline from the architectural design image, and obtain a structural design result of the shear wall building based on the vector data of the shear wall components and the vector data of building floor outline to check strength of shear wall components.

Description

CROSS-REFERENCE TO RELATED APPLICATION The present application is a national stage application of International Patent Application No. PCT/CN2022/080626 filed on Mar. 14, 2022, entitled “PHYSICS-ENHANCED DATA-DRIVEN METHOD AND DEVICE FOR INTELLIGENT STRUCTURAL DESIGN OF SHEAR WALL BUILDING,” which claims priority to Chinese patent application, No. 202111028475.8, filed on Sep. 2, 2021, entitled “Physics-Enhanced Data-Driven Method and Device for Intelligent Structural Design of Shear Wall Building”, which is hereby incorporated by reference in its entirety. TECHNICAL FIELD The present application relates to the technical field of architecture structure design and machine learning, in particular to a physics-enhanced data-driven method and device for intelligent structural design of shear wall building. BACKGROUND At present, the structural scheme design of shear wall building is generally completed by experienced engineers. Due to excessive dependence on professional experience, artificial design is difficult to be effectively popularized and applied. In addition, artificial design is time-consuming and labor-intensive, lacks intelligence, and has low efficiency. Hence, intelligent structural design methods emerge as the times require to match the rapid development of intelligent construction. However, physical mechanisms of structural design are generally described by formula, text, and other complex forms, which are significantly different from the description of image data. And the neural network can only effectively learn the characteristics of image data and can hardly learn the physical mechanisms of structural design directly. As a result, the traditional intelligent structural design methods are only driven by data, and intelligent algorithms cannot learn the physical mechanisms and empirical knowledge of structural design. Furthermore, the final structural design results are challenging to meet the actual physical constraints and requirements of structural design and have a low practical value. Therefore, there is an urgent need for an intelligent structural design method of shear wall building that can learn physical mechanisms and data raws simultaneously. SUMMARY The present application provides a physics-enhanced data-driven method and device for the intelligent structural design of shear wall building, to solve the deficiencies in the related art in which a design method for shear wall building structure cannot learn both physical mechanisms and data raws through intelligent algorithms, resulting in a low practical value of a structure design result. In a first aspect, a physics-enhanced data-driven method for intelligent structural design of shear wall building provided by the present application, including: obtaining an architectural design image and a basic design condition text to be processed;inputting the architectural design image and the basic design condition text into a structural design model, and obtaining a structural design image output by the structural design model; where, the structural design model is obtained by performing a structural design image generation capability training and a physical performance optimization training for a physics-enhanced data-driven generative adversarial network, based on sample data of architectural design image, sample data of basic design condition text corresponding to the sample data of architectural design image and sample data of structural design image corresponding to the sample data of architectural design image; andvectorizing the structural design image and the architectural design image; extracting vector data of shear wall components from the structural design image and vector data of building floor outline from the architectural design image, and subsequently obtaining a structural design result of shear wall building based on the extracted vector data of shear wall components and the vector data of building floor outline. According to the physics-enhanced data-driven method for intelligent structural design of shear wall building provided by the present application, a training process for the structural design model includes: building the physics-enhanced data-driven generative adversarial network;inputting sample data with complete labels into the physics-enhanced data-driven generative adversarial network, and performing the structural design image generation capability training for the physics-enhanced data-driven generative adversarial network; where, the sample data with complete labels includes sample data of architectural design image, sample data of basic design condition text corresponding to the sample data of architectural design image and sample data of structural design image corresponding to the sample data of architectural design image;performing a calculation of physical performance for the structural design image generated during a process of the structural design image generation capability training, and obtaini