Intelligent design method for high-rise building structure

Abstract

The application discloses a high-rise building structure intelligent design method, which comprises intelligent generation of a high-rise building plan layout, intelligent modeling and intelligent optimization.The intelligent generation algorithm is an intelligent generation method of the high-rise building plan layout, which integrates high-dimensional feature extraction of the plan layout, multi-suite intelligent layout and multi-suite layout refinement.The intelligent modeling algorithm is an intelligent modeling method of the high-rise building structure, which integrates intelligent segmentation and identification of the suite, intelligent layout of the stress component and parameterized modeling.The intelligent optimization algorithm is an intelligent optimization method of the high-rise building structure, which integrates a deep reinforcement learning architecture, an efficient training strategy of the deep reinforcement learning and a transfer learning architecture.The application forms the intelligent design method of the high-rise building structure through the intersection of the two disciplines of structural engineering and artificial intelligence, and solves the problems of low efficiency, long cycle and strong subjectivity in the traditional high-rise building structure design.

Description

Technical Field

[0001] This invention relates to the field of architectural design, and more specifically to an intelligent design method for high-rise building structures. Background Technology

[0002] Traditional high-rise building structural design suffers from problems such as low efficiency, long cycle and strong subjectivity. Specifically, it manifests as follows: (1) In the preliminary design stage, the client's needs cannot be clearly defined, resulting in repeated iterations of the design work; (2) In the construction drawing design stage, structural engineers need to interact repeatedly with architectural engineers, and structural engineers need to repeatedly adjust and verify the calculation model; (3) For the same building, different structural engineers usually give different design results.

[0003] Current intelligent design methods focus on the intelligent generation of high-rise building floor plans for single apartment types, lacking research on multiple apartment types. Furthermore, the intelligent generation of high-rise building floor plans has failed to integrate with industrialization. Current automatic generation methods for building BIM models are limited to recognizing low-dimensional features such as walls and doorways, lacking the recognition of high-dimensional features such as apartment types (residential and office) and public areas. However, intelligent structural modeling relies on prior knowledge of apartment types and public areas. Current intelligent structural optimization methods all employ heuristic algorithms, which are suitable for single-step optimization of structural calculation models, but cannot guide intelligent optimization of similar buildings due to their inability to learn from optimization experience. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent design method for high-rise building structures to solve the problems existing in the prior art.

[0005] The technical solution adopted to achieve the purpose of this invention is as follows: a smart design method for high-rise building structures, comprising the following steps:

[0006] 1) Intelligent generation of floor plans for high-rise buildings;

[0007] 2) Intelligent modeling of high-rise building structures;

[0008] 3) Intelligent optimization of high-rise building structures.

[0009] Furthermore, step 1) includes the following sub-steps:

[0010] 1-1) Collect existing high-rise building floor plans, extract apartment types from the high-rise building floor plans, and form an apartment type library;

[0011] 1-2) Using residential unit types, office unit types, public passages, elevators and stairs as basic elements, semantics and outer contour pixel extraction are performed on the floor plan of high-rise buildings. Batch processing of the floor plan of high-rise buildings is carried out to form a high-dimensional feature library of floor plan of high-rise buildings.

[0012] 1-3) Based on the text and pixel fusion method, a neural network architecture with text information and outer contour pixels as input is built. Samples are randomly extracted from the high-dimensional feature library to train the neural network architecture and perform intelligent deployment of multiple sets.

[0013] 1-4) A heuristic algorithm is used to refine the layout of multiple sets; among them, the corner coordinates of each building unit are optimization variables, the input outer contour pixels and the non-overlapping of building units are constraints, and the difference between the total area of ​​the building units and the input total area is the objective function.

[0014] Furthermore, step 2) includes the following sub-steps:

[0015] 2-1) Collect the floor plans of existing high-rise buildings and the floor plans of high-rise buildings intelligently generated in steps 1-4), and perform manual preprocessing on the floor plans to form a drawing database;

[0016] 2-2) Based on the drawing database, the naming rules of the layers are summarized, and then a high-rise building component extraction algorithm is developed based on the layer analysis method to realize the automatic extraction and positioning of walls, windows, ordinary doors and elevator doors;

[0017] 2-3) The automatically extracted high-rise building components are converted into point clouds and binarized into images. The binarized images are then processed using a connected component analysis algorithm to perform intelligent room segmentation.

[0018] 2-4) Establish the room graph structure through the connectivity of doors, clean up, merge, cluster and segment the graph structure to perform intelligent segmentation of apartment types; use Protodyakonov analysis to match the corner points of apartment types to complete the intelligent identification of the same apartment type.

[0019] 2-5) Based on the drawing database, current structural design codes and expert experience, summarize the arrangement rules of anti-side members and load-bearing members to form prior knowledge of the arrangement of load-bearing members;

[0020] 2-6) By writing the corresponding software data format, the initial design scheme of high-rise building structure can be automatically generated.

[0021] Furthermore, steps 2-4) include the following sub-steps:

[0022] 2-4-1) Construct a graph structure of rooms using the connectivity of doors, where a wall, represented by l, is partitioned if it has at least one door. If there are no doors, it is divided into The walls surrounding room i# converged into It can be expressed using formula (1):

[0023]

[0024] In the formula, Mij=1 This indicates that room i# is connected to room j#, by removing M. i For rooms with a value of 0, clean up the diagram structure;

[0025] 2-4-2) Perform important node analysis on the room plan structure to obtain the public area, living room, and apartment layout structure with the living room as the core of the entire floor plan;

[0026] 2-4-3) Identify similar apartment types; among them, a set of walls The intersection point represents the housing unit and another housing unit. The indicated housing units are matched, and Protodyakonov analysis is used to match the corner points of the housing units. The calculation method is formula (2) to (6):

[0027]

[0028] Where R is the rotation matrix and t is the translation matrix; the diagonal matrix Σ, the left singular matrix U, and the right singular matrix V are all obtained by singular value decomposition of matrix W; μ t and μ s Let S be the mean coordinates of point sets T and S, respectively; S i and T i These are control point pairs determined by the nearest neighbor algorithm.

[0029] Furthermore, step 3) includes the following sub-steps:

[0030] 3-1) Based on current structural design codes and expert experience, we propose specific forms of the objective function and penalty function for intelligent optimization of high-rise building structures. Combining the parameter value range and the inherent constraint relationship between parameters, we form a mathematical model for intelligent optimization of high-rise building structures.

[0031] 3-2) For the mathematical model of intelligent optimization of high-rise building structure, a deep reinforcement learning architecture for intelligent optimization of high-rise building structure is built.

[0032] 3-3) Use heuristic algorithms to obtain the structure of high-rise buildings and perform intelligent optimization to form a database of <initial state, feasible solution>; collect samples from the database to complete the training of the deep reinforcement learning architecture and form a deep reinforcement learning training strategy.

[0033] 3-4) Intelligent design is performed on high-rise building structures that do not meet the set requirements to form a database of <structural plan layout and deep reinforcement learning parameters>.

[0034] 3-5) Construct a network architecture for transfer learning, propose a method for evaluating the similarity of structural planar layout diagrams, and then establish the loss function form for transfer learning;

[0035] 3-6) Several <structural plan layout diagrams and deep reinforcement learning parameters> samples are used to train the transfer learning and complete the intelligent design of the new high-rise building.

[0036] Furthermore, in step 3-2), when building the deep reinforcement learning architecture, such as Figure 9 As shown; the objective function is calculated using the following formula:

[0037]

[0038] In the formula, C is the target value. Indicates the period ratio, Indicates the displacement ratio. Indicates the maximum inter-story drift angle. and Indicates the length of the wall segment in two directions in the plane. The value represents the state S at time t. t Take action below The expected return, where π represents the strategy of the action. This indicates the time discount rate; the superscript * represents the optimal strategy. Indicates the next time step (t+1), superscript To represent different neural networks, the parameters that need to be adjusted throughout the reinforcement learning process are... express.

[0039] Furthermore, in steps 3-5), the similarity evaluation method uses the following formula for calculation:

[0040]

[0041] Where x and y represent two images being compared, and μ x μ y σ x σ y σ xy Let x and y represent the mean gray level of image x, the mean gray level of image y, the gray level variance of image x, the gray level variance of image y, and the gray level covariance of images x and y, respectively. L is the dynamic range of pixel values. The final structural planar layout diagram has a similarity range of [0, 1], and the larger the value, the more similar the images are.

[0042] The loss function for transfer learning is constructed based on the maximum mean difference:

[0043]

[0044] Where H indicates that this distance is caused by The data is mapped to a regenerated Hilbert space for measurement.

[0045] The technical effects of this invention are undeniable. This invention combines image processing technology, evolutionary algorithms, adversarial neural networks, deep reinforcement learning, transfer learning and other technologies and algorithms. It is a complete intelligent design method for high-rise building structures, providing basic algorithmic support for intelligent design technology of high-rise building structures, promoting the integration of artificial intelligence and high-rise structure design, and improving the technological level, work efficiency and design quality of high-rise building structure design. Attached Figure Description

[0046] Figure 1 This is a block diagram of the method of the present invention;

[0047] Figure 2 This is a flowchart illustrating the intelligent generation process of the high-rise building floor plan layout according to the present invention.

[0048] Figure 3 This is a schematic diagram of the adversarial generative network for the floor plan layout of a high-rise building according to the present invention.

[0049] Figure 4 This is a flowchart of the intelligent modeling process for high-rise building structures according to the present invention;

[0050] Figure 5 This is a schematic diagram of the intelligent segmentation and recognition of high-rise building structures according to the present invention;

[0051] Figure 6 This is a schematic diagram of intelligent modeling of high-rise building structures according to the present invention;

[0052] Figure 7 This is a flowchart illustrating the intelligent optimization process of high-rise building structures based on evolutionary algorithms according to the present invention.

[0053] Figure 8 This is a flowchart of the intelligent optimization process for high-rise building structures based on neural networks, as described in this invention.

[0054] Figure 9 This is a flowchart of the intelligent optimization process for high-rise building structures based on deep reinforcement learning, as described in this invention.

[0055] Figure 10 Design architecture for adversarial generative networks and discriminative networks. Detailed Implementation

[0056] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0057] Example 1:

[0058] See Figure 1 This embodiment discloses an intelligent design method for high-rise building structures, comprising three main steps: intelligent generation, intelligent modeling, and intelligent optimization. See also... Figure 2 Intelligent generation includes the following steps:

[0059] 1) Collect floor plans of existing high-rise buildings with a high degree of industrialization, extract apartment types from the floor plans of high-rise buildings, and form an apartment type library oriented towards industrialization;

[0060] 2) Using residential unit types, office unit types, public passages, elevators, and staircases as basic elements, high-rise building floor plans are processed in batches using rapid semanticization and outer contour pixel extraction algorithms to form a high-dimensional feature library of high-rise building floor plans; the high-dimensional feature database is a collection of text information and outer contour pixels;

[0061] 3) Based on the text and pixel fusion method, a neural network architecture is built with text information and outer contour pixels as input and a set layout diagram as output; an adversarial neural network is used to train the neural network architecture by randomly extracting samples from a high-dimensional feature library, thereby realizing intelligent arrangement of multiple sets. Figure 3 This is a schematic diagram of the adversarial generative network for the floor plan of a high-rise building as described in this embodiment. This network can quickly generate the layout of apartment types.

[0062] 4) Since the results of multi-unit intelligent layout are usually coarse, there are gaps or overlaps between the generated building units, and the bounding boxes of all building units do not completely match the input outer contour pixels; therefore, a heuristic algorithm is used to refine the multi-unit layout. The corner coordinates of each building unit are optimization variables, the input outer contour pixels and the non-overlapping of building units are constraints, and the difference between the total area of ​​the building units and the total area of ​​the input is the objective function.

[0063] See Figure 4 The intelligent modeling includes the following steps:

[0064] 5) By collecting a large number of existing and intelligently generated high-rise building floor plans, the building floor plans are manually preprocessed to form a drawing database;

[0065] 6) Based on the drawing database, the naming rules of the layers are summarized, and then a high-rise building component extraction algorithm is developed based on the layer analysis method to realize the automatic extraction and positioning of walls, windows, ordinary doors and elevator doors;

[0066] 7) The automatically extracted high-rise building components are converted into point clouds and binarized; the binarized image is processed using a connected component analysis algorithm to achieve intelligent room segmentation;

[0067] 8) Construct a graphical structure of the rooms using the connectivity of doors, where... A wall, if it has at least one door, can be divided into... If there are no doors, it can be divided into The walls surrounding room i# can be gathered together. See formula (1):

[0068]

[0069] M ij = 1 indicates that room i# is connected to room j#. This is achieved by removing M. i For rooms with a value of 0, clean up the diagram structure.

[0070] Then, the important nodes of the room diagram structure are analyzed to obtain the public area of ​​the entire plane, the living room, and the apartment layout structure with the living room as the core. In formula (2), s represents the starting point of the path, which is randomly obtained in each round; e represents the vertex of the graph structure, which is traversed in each round; v represents the vertex passed through by this path; Let represent the shortest path between s and e; after multiple rounds of repeated operations, obtain the numerical value C of the path traversing all vertices in the graph over multiple rounds. v Then, KMeans clustering was used to separate the public area and the living room.

[0071]

[0072] Identify similar apartment layouts; among them, a set of walls The intersection point represents the housing unit and another housing unit. The represented housing units are matched, and Protodyakonov analysis is used to match the corner points of the housing units, as shown in formulas (3)-(7). Where R is the rotation matrix, t is the translation matrix; the diagonal matrix Σ, the left singular matrix U, and the right singular matrix V are all obtained from the singular value decomposition of matrix W; μ t and μ s Let S be the mean coordinates of point sets T and S, respectively; S i and T i These are control point pairs determined by the nearest neighbor algorithm.

[0073]

[0074] Figure 5 (a) is a schematic diagram of room segmentation based on the connected component analysis algorithm; Figure 5 (b) is a schematic diagram of nest type identification based on graph structure algorithm and Protodyakonov analysis;

[0075] 9) Based on the drawing database, current structural design specifications and expert experience, summarize the rules for the arrangement of anti-slip and load-bearing components to form prior knowledge for the arrangement of load-bearing components. For example, wall settings should avoid straight walls as much as possible and should be designed as T-shaped or L-shaped.

[0076] 10) See Figure 6 By writing data into the public data interfaces of structural analysis software (PKPM, YJK, ETABS, SAP2000), the initial design scheme of high-rise building structure can be automatically generated.

[0077] The intelligent optimization includes the following steps:

[0078] 11) Based on current structural design codes and expert experience, propose specific forms of the objective function and penalty function for intelligent optimization of high-rise building structures. Combined with the parameter range and the inherent constraint relationship between parameters in Part II, form a mathematical model for intelligent optimization of high-rise building structures.

[0079] 12) For the mathematical model of intelligent optimization of high-rise building structures, build a deep reinforcement learning architecture for intelligent optimization of high-rise building structures, such as... Figure 9 As shown; the objective function is calculated using formulas (8)(9)(10)(11)(12)(13)(14).

[0080]

[0081] C is the target value. Indicates the period ratio, Indicates the displacement ratio. L represents the maximum inter-story drift angle. i1 and L i2 This represents the length of the wall segment in both directions of the plane. The Q value represents the state S at time t. t Take action below The expected return, where π represents the strategy of the action. This indicates the time discount rate; the superscript * represents the optimal strategy. Indicates the next time step (t+1), superscript Different neural networks are represented by , and the parameters that need to be adjusted throughout the reinforcement learning process are . To represent this.

[0082] 13) Use heuristic algorithms to obtain the structure of high-rise buildings for intelligent optimization, forming a database of <initial state, feasible solution>; collect samples from the database to complete the training of the deep reinforcement learning architecture, thereby forming an efficient and reliable deep reinforcement learning training strategy;

[0083] 14) Intelligent design of high-rise building structures with significant differences in floor plan layout is carried out to form a database of <structural floor plan layout and deep reinforcement learning parameters>.

[0084] 15) Construct a network architecture for transfer learning; propose a structural planar layout similarity evaluation method based on image structural similarity, using the following formula for calculation:

[0085]

[0086]

[0087] Where x and y represent two images being compared, and μ x μ y σ x σ y σ xy Let x and y represent the mean gray level of image x, the mean gray level of image y, the gray level variance of image x, the gray level variance of image y, and the gray level covariance of images x and y, respectively. L is the dynamic range of pixel values. The final structural planar layout diagram has a similarity range of [0, 1], and the larger the value, the more similar the images are.

[0088] The loss function for transfer learning is constructed based on the maximum mean difference:

[0089]

[0090] Where H indicates that this distance is caused by The data is mapped to a regenerated Hilbert space for measurement.

[0091] 16) A large number of <structural plan layout diagrams, deep reinforcement learning parameters> samples are used to train transfer learning, thereby realizing rapid intelligent design of new high-rise buildings.

[0092] This embodiment can employ intelligent optimization methods based on prior knowledge and evolutionary algorithms (such as...). Figure 7 As shown in the figure, the system can automatically generate and arrange the initial structural design schemes for high-rise buildings; in addition, an optimization algorithm based on neural networks can be used, the process of which is as follows: Figure 8 As shown.

[0093] In summary, this embodiment proposes an intelligent algorithm for generating floor plans of high-rise buildings, forming an intelligent method for generating floor plans of high-rise buildings that integrates high-dimensional feature extraction, intelligent layout of multiple apartment types, and refinement of the layout of multiple apartment types. This embodiment also proposes an intelligent modeling algorithm for high-rise building structures, forming an intelligent modeling method for high-rise building structures that integrates intelligent segmentation and recognition of apartment types, intelligent layout of load-bearing components, and parametric modeling. Based on prior knowledge, computer graphics algorithms, and an improved intelligent optimization algorithm, this embodiment can complete parametric modeling of high-rise building structures and automatic design and optimization of high-rise reinforced concrete shear wall structures using only architectural design drawings. Finally, this embodiment proposes an intelligent optimization algorithm for high-rise building structures, forming an intelligent optimization method for high-rise building structures that integrates a deep reinforcement learning architecture, an efficient training strategy for deep reinforcement learning, and a transfer learning architecture.

[0094] It is worth noting that the intelligent design method for high-rise building structures proposed in this embodiment combines image processing technology, evolutionary algorithms, adversarial neural networks, deep reinforcement learning, transfer learning, and other technologies or algorithms. Specifically, it includes an intelligent method for generating high-rise building floor plans, an intelligent modeling method for high-rise building structures, and an intelligent optimization method for high-rise building structures. The system algorithm for intelligent design of high-rise building structures in this embodiment is a complete intelligent design method for high-rise building structures, providing fundamental algorithmic support for intelligent design technology of high-rise building structures, promoting the integration of artificial intelligence and high-rise structure design, and improving the technological level, work efficiency, and design quality of high-rise building structure design.

[0095] Example 2:

[0096] This embodiment discloses an intelligent design method for high-rise building structures, including the following steps:

[0097] 1) Intelligent generation of floor plans for high-rise buildings;

[0098] 2) Intelligent modeling of high-rise building structures;

[0099] 3) Intelligent optimization of high-rise building structures.

[0100] Example 3:

[0101] The main steps of this embodiment are the same as those of embodiment 2. Furthermore, step 1) includes the following sub-steps:

[0102] 1-1) Collect existing high-rise building floor plans, extract apartment types from the high-rise building floor plans, and form an apartment type library;

[0103] 1-2) Using residential unit types, office unit types, public passages, elevators and stairs as basic elements, semantics and outer contour pixel extraction are performed on the floor plan of high-rise buildings. Batch processing of the floor plan of high-rise buildings is carried out to form a high-dimensional feature library of floor plan of high-rise buildings.

[0104] 1-3) Based on the text and pixel fusion method, a neural network architecture with text information and outer contour pixels as input is built. Samples are randomly extracted from the high-dimensional feature library to train the neural network architecture and perform intelligent deployment of multiple sets.

[0105] 1-4) A heuristic algorithm is used to refine the layout of multiple sets; among them, the corner coordinates of each building unit are optimization variables, the input outer contour pixels and the non-overlapping of building units are constraints, and the difference between the total area of ​​the building units and the input total area is the objective function.

[0106] Example 4:

[0107] The main steps of this embodiment are the same as those of embodiment 3. Furthermore, step 2) includes the following sub-steps:

[0108] 2-1) Collect the floor plans of existing high-rise buildings and the floor plans of high-rise buildings intelligently generated in steps 1-4), and perform manual preprocessing on the floor plans to form a drawing database;

[0109] 2-2) Based on the drawing database, the naming rules of the layers are summarized, and then a high-rise building component extraction algorithm is developed based on the layer analysis method to realize the automatic extraction and positioning of walls, windows, ordinary doors and elevator doors;

[0110] 2-3) The automatically extracted high-rise building components are converted into point clouds and binarized into images. The binarized images are then processed using a connected component analysis algorithm to perform intelligent room segmentation.

[0111] 2-4) Establish the room graph structure through the connectivity of doors, clean up, merge, cluster and segment the graph structure to perform intelligent segmentation of apartment types; use Protodyakonov analysis to match the corner points of apartment types to complete the intelligent identification of the same apartment type.

[0112] 2-5) Based on the drawing database, current structural design codes and expert experience, summarize the arrangement rules of anti-side members and load-bearing members to form prior knowledge of the arrangement of load-bearing members;

[0113] 2-6) By writing the corresponding software data format, the initial design scheme of high-rise building structure can be automatically generated.

[0114] Example 5:

[0115] The main steps in this embodiment are the same as in embodiment 4. Furthermore, steps 2-4) include the following sub-steps:

[0116] 2-4-1) Construct a graph structure of rooms using the connectivity of doors, where a wall, represented by l, is partitioned if it has at least one door. If there are no doors, it is divided into The walls surrounding room i# converged into It can be expressed using formula (1):

[0117]

[0118] In the formula, Mij=1 This indicates that room i# is connected to room j#, by removing M. i For rooms with a value of 0, clean up the diagram structure;

[0119] 2-4-2) Perform important node analysis on the room plan structure to obtain the public area, living room, and apartment layout structure with the living room as the core of the entire floor plan;

[0120] 2-4-3) Identify similar apartment types; among them, a set of walls The intersection point represents the housing unit and another housing unit. The indicated housing units are matched, and Protodyakonov analysis is used to match the corner points of the housing units. The calculation method is formula (2) to (6):

[0121]

[0122]

[0123] Where R is the rotation matrix and t is the translation matrix; the diagonal matrix Σ, the left singular matrix U, and the right singular matrix V are all obtained by singular value decomposition of matrix W; μ t and μ s Let S be the mean coordinates of point sets T and S, respectively; S i and T i These are control point pairs determined by the nearest neighbor algorithm.

[0124] Example 6:

[0125] The main steps in this embodiment are the same as in embodiment 4. Furthermore, step 3) includes the following sub-steps:

[0126] 3-1) Based on current structural design codes and expert experience, we propose specific forms of the objective function and penalty function for intelligent optimization of high-rise building structures. Combining the parameter value range and the inherent constraint relationship between parameters, we form a mathematical model for intelligent optimization of high-rise building structures.

[0127] 3-2) For the mathematical model of intelligent optimization of high-rise building structure, a deep reinforcement learning architecture for intelligent optimization of high-rise building structure is built.

[0128] 3-3) Use heuristic algorithms to obtain the structure of high-rise buildings and perform intelligent optimization to form a database of <initial state, feasible solution>; collect samples from the database to complete the training of the deep reinforcement learning architecture and form a deep reinforcement learning training strategy.

[0129] 3-4) Intelligent design is performed on high-rise building structures that do not meet the set requirements to form a database of <structural plan layout and deep reinforcement learning parameters>.

[0130] 3-5) Construct a network architecture for transfer learning, propose a method for evaluating the similarity of structural planar layout diagrams, and then establish the loss function form for transfer learning;

[0131] 3-6) Several <structural plan layout diagrams and deep reinforcement learning parameters> samples are used to train the transfer learning and complete the intelligent design of the new high-rise building.

[0132] Example 7:

[0133] The main steps in this embodiment are the same as in embodiment 6. Further, in step 3-2), when building the deep reinforcement learning architecture, as follows... Figure 9 As shown; the objective function is calculated using the following formula:

[0134]

[0135]

[0136] In the formula, C is the target value. Indicates the period ratio, Indicates the displacement ratio. L represents the maximum inter-story drift angle. i1 and L i2 Indicates the length of the wall segment in two directions in the plane. The value represents the state S at time t. t Take action below The expected return, where π represents the strategy of the action. This indicates the time discount rate; the superscript * represents the optimal strategy. Indicates the next time step (t+1), superscript To represent different neural networks, the parameters that need to be adjusted throughout the reinforcement learning process are... express.

[0137] Example 8:

[0138] The main steps in this embodiment are the same as in embodiment 6. Furthermore, in steps 3-5), the similarity evaluation method uses the following formula for calculation:

[0139]

[0140] Where x and y represent two images being compared, and μ x μ y σ x σ y σ xy Let x and y represent the mean gray level of image x, the mean gray level of image y, the gray level variance of image x, the gray level variance of image y, and the gray level covariance of images x and y, respectively. L is the dynamic range of pixel values. The final structural planar layout diagram has a similarity range of [0, 1], and the larger the value, the more similar the images are.

[0141] The loss function for transfer learning is constructed based on the maximum mean difference:

[0142]

[0143] Where H indicates that this distance is caused by The data is mapped to a regenerated Hilbert space for measurement.

Claims

1. A method for intelligent structural design of high-rise buildings, characterized in that, Includes the following steps: 1) Intelligent generation of floor plans for high-rise buildings, including the following steps: 1-1) Collect existing high-rise building floor plans, extract apartment types from the high-rise building floor plans, and form an apartment type library; 1-2) Using residential unit types, office unit types, public passages, elevators and stairs as basic elements, semantics and outer contour pixel extraction are performed on the floor plan of high-rise buildings. Batch processing of the floor plan of high-rise buildings is carried out to form a high-dimensional feature library of floor plan of high-rise buildings. 1-3) Based on the text and pixel fusion method, a neural network architecture with text information and outer contour pixels as input is built. Samples are randomly extracted from the high-dimensional feature library to train the neural network architecture and perform intelligent deployment of multiple sets. 1-4) A heuristic algorithm is used to refine the layout of multiple apartment types; among them, the corner coordinates of each building unit are optimization variables, the input outer contour pixels and the non-overlapping of building units are constraints, and the difference between the total area of ​​the building units and the input total area is the objective function. 2) Intelligent modeling of high-rise building structures, including the following steps: 2-1) Collect the floor plans of existing high-rise buildings and the floor plans of high-rise buildings intelligently generated in steps 1-4), and perform manual preprocessing on the floor plans to form a drawing database; 2-2) Based on the drawing database, the naming rules of the layers are summarized, and then a high-rise building component extraction algorithm is developed based on the layer analysis method to realize the automatic extraction and positioning of walls, windows, ordinary doors and elevator doors; 2-3) The automatically extracted high-rise building components are converted into point clouds and binarized into images. The binarized images are then processed using a connected component analysis algorithm to perform intelligent room segmentation. 2-4) Establish the room graph structure through the connectivity of doors, clean up, merge, cluster and segment the graph structure to perform intelligent segmentation of apartment types; use Protodyakonov analysis to match the corner points of apartment types to complete the intelligent identification of the same apartment type. 2-5) Based on the drawing database, current structural design codes and expert experience, summarize the arrangement rules of anti-side members and load-bearing members to form prior knowledge of the arrangement of load-bearing members; 2-6) By writing the corresponding software data format, the initial design scheme of high-rise building structure can be automatically generated; 3) Intelligent optimization of high-rise building structures, including the following steps: 3-1) Based on current structural design codes and expert experience, we propose specific forms of the objective function and penalty function for intelligent optimization of high-rise building structures. Combining the parameter value range and the inherent constraint relationship between parameters, we form a mathematical model for intelligent optimization of high-rise building structures. 3-2) For the mathematical model of intelligent optimization of high-rise building structure, a deep reinforcement learning architecture for intelligent optimization of high-rise building structure is built. 3-3) Use heuristic algorithms to obtain the structure of high-rise buildings and perform intelligent optimization to form a database of <initial state, feasible solution>; collect samples from the database to complete the training of the deep reinforcement learning architecture and form a deep reinforcement learning training strategy. 3-4) Intelligent design is performed on high-rise building structures that do not meet the set requirements to form a database of <structural plan layout and deep reinforcement learning parameters>. 3-5) Construct a network architecture for transfer learning, propose a method for evaluating the similarity of structural planar layout diagrams, and then establish the loss function form for transfer learning; 3-6) Several <structural plan layout diagrams and deep reinforcement learning parameters> samples are used to train the transfer learning and complete the intelligent design of the new high-rise building.

2. The intelligent design method for high-rise building structures according to claim 1, characterized in that, Steps 2-4) include the following sub-steps: 2-4-1) Construct a graph structure of rooms using the connectivity of doors, where a wall, represented by l, is partitioned if it has at least one door. If there are no doors, it is divided into ;Room The surrounding walls gathered into It can be expressed using formula (1): In the formula, This indicates that room i# is connected to room j#, by removing M. i For rooms with a value of 0, clean up the diagram structure; 2-4-2) Perform important node analysis on the room plan structure to obtain the public area, living room, and apartment layout structure with the living room as the core of the entire floor plan; 2-4-3) Identify similar apartment types; among them, a set of walls The intersection point represents the housing unit and another housing unit. The indicated housing units are matched, and Protodyakonov analysis is used to match the corner points of the housing units. The calculation method is formula (2) to (6): Where R is the rotation matrix and t is the translation matrix; the diagonal matrix Σ, the left singular matrix U, and the right singular matrix V are all obtained by singular value decomposition of matrix W; μ t and μ s Let S be the mean coordinates of point sets T and S, respectively; S i and T i These are control point pairs determined by the nearest neighbor algorithm.

3. The intelligent design method for high-rise building structures according to claim 1, characterized in that, In steps 3-5), the similarity evaluation method uses the following formula for calculation: Where x and y represent two images being compared, and μ x μ y σ x σ y σ xy Let x and y represent the mean gray level of image x, the mean gray level of image y, the gray level variance of image x, the gray level variance of image y, and the gray level covariance of images x and y, respectively. L is the dynamic range of pixel values. The final structural planar layout diagram has a similarity range of [0, 1], and the larger the value, the more similar the images are. The loss function for transfer learning is constructed based on the maximum mean difference: Where H indicates that this distance is caused by The data is mapped to a regenerated Hilbert space for measurement.

Images