Building structure design optimization method based on artificial intelligence

Abstract

The invention relates to the field of building structure design, in particular to a building structure design optimization method based on artificial intelligence, and the method comprises the steps: obtaining a design file and environment data of a building structure; according to design requirements, the design file and the environment data, a finite element model is established, and the finite element model and the building structure are consistent in mechanical property; inputting a load working condition scene into the finite element model to obtain a real working condition under the design requirement; performing multi-objective optimization on the real working condition by using a deep neural network to obtain an optimization scheme meeting the design requirements; the optimization scheme is converted into an executable mathematical model, the building structure is constructed, and by establishing a two-channel deep neural network architecture combining physical constraint and data driving, it is ensured that an optimization result conforms to the mechanics principle, implicit modes in historical data can be fully utilized, and optimization efficiency and precision are improved.

Description

Technical Field

[0001] The present invention relates to the field of building structure design, and in particular to an artificial intelligence-based building structure design optimization method, which combines deep neural network technology with building structure engineering principles to achieve intelligent optimization of building structure design. Background Art

[0002] Traditional structural design optimization often relies on the designer's experience and trial and error. This approach is not only time-consuming and labor-intensive, but also struggles to find the optimal solution under multiple constraints. While finite element analysis has been widely used in structural optimization in recent years, it still suffers from high computational complexity and low efficiency when faced with complex working conditions and multiple constraints. Furthermore, traditional optimization methods often fail to fully consider the structure's full lifecycle performance and its comprehensive performance under various load conditions.

[0003] With the development of artificial intelligence (AI) technology, structural optimization methods based on intelligent algorithms such as neural networks and genetic algorithms have gradually emerged. However, most existing AI optimization methods suffer from the following problems: First, they overly rely on historical data and lack physical constraints, resulting in optimization results that may not conform to mechanical principles; second, they typically optimize for a specific goal, making it difficult to balance multiple design objectives; third, the optimization models are disconnected from actual engineering application scenarios and lack sufficient consideration of real-world working conditions; and fourth, their knowledge transfer capabilities are limited, making it difficult to apply existing experience to new design scenarios.

[0004] Therefore, there is an urgent need for a building structure design optimization method that can combine physical models and data-driven models, adapt to complex working conditions, achieve multi-objective balance, and have knowledge transfer capabilities. Summary of the Invention

[0005] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide an artificial intelligence-based building structure design optimization method, which can organically combine deep learning technology with structural mechanics principles to achieve efficient, accurate, and multi-objective optimization of building structures.

[0006] The present invention proposes an artificial intelligence-based building structure design optimization method, comprising:

[0007] Obtain design documents and environmental data of building structures;

[0008] Establishing a finite element model based on the design requirements, the design documents, and the environmental data, wherein the finite element model is consistent with the building structure in terms of mechanical properties;

[0009] Inputting a load condition scenario into the finite element model to obtain a real condition under the design requirements;

[0010] Using a deep neural network to perform multi-objective optimization on the actual working conditions to obtain an optimization solution that meets the design requirements;

[0011] The optimization plan is converted into an executable mathematical model, and the building structure is constructed.

[0012] Preferably, the inputting of the load condition scenario into the finite element model to obtain the actual condition under the design requirements is specifically as follows:

[0013] Performing modal analysis on the building structure to obtain modal characteristics under the design requirements;

[0014] A load condition scenario database that affects the vibration of the building structure is established, and the real working condition of the building structure is matched from the load condition scenario database based on the modal characteristics.

[0015] Preferably, the deep neural network includes a working condition feature extractor, a true optimization objective function, a first fitness evaluator and a second fitness evaluator; the multi-objective optimization of the true working condition using the deep neural network is specifically as follows:

[0016] The working condition feature extractor performs data preprocessing on the real working condition to obtain working condition features corresponding to the real working condition;

[0017] The first fitness evaluator calculates a first fitness based on the working condition characteristics and the load condition scenario parameters through the true optimization objective function;

[0018] The second fitness evaluator performs performance evaluation based on the operating condition characteristics to obtain a second fitness;

[0019] In which, the dimensions of the operating condition features extracted by the operating condition feature extractor and the number of performance evaluation indicators are consistent, and the first fitness is obtained by calculating the multi-objective optimization objective function; if both the first fitness evaluator and the second fitness evaluator meet the preset threshold, the multi-objective optimization is completed, otherwise, the multi-objective optimization continues.

[0020] Preferably, before obtaining the design documents and environmental data of the building structure, the method further includes the following steps:

[0021] Construct building structure design models based on engineering knowledge;

[0022] Deep learning based on the building structure design model and topology optimization of the building structure.

[0023] Preferably, the multi-objective optimization of the actual working condition is performed using a deep neural network to obtain an optimization solution that meets the design requirements, specifically:

[0024] Inputting the actual working condition into the deep neural network, extracting a non-dominated solution set of the actual working condition using a non-dominated sorting algorithm, and evaluating the non-dominated solution set based on a preset evaluation algorithm to obtain an optimal solution;

[0025] The optimized solution of the building structure is generated according to the best solution.

[0026] Preferably, the method of performing multi-objective optimization on the actual working condition using a deep neural network and obtaining an optimization solution that meets the design requirements further includes the following steps:

[0027] Determine whether the preset optimization termination condition is reached;

[0028] If yes, then terminate the multi-objective optimization;

[0029] Otherwise, continue with the multi-objective optimization.

[0030] Preferably, the deep neural network adopts a dual-channel architecture, including:

[0031] Physical constraint channels, incorporating structural mechanics laws and material nonlinear models;

[0032] Data-driven channel, learning implicit patterns from historical optimization cases;

[0033] Channel interaction mechanism to achieve cross-channel knowledge complementarity;

[0034] The weighted fusion layer dynamically adjusts the contribution weights of the physical constraint channel and the data-driven channel.

[0035] Preferably, the multi-objective optimization of the real working condition using a deep neural network comprises the following steps:

[0036] Implement global-local alternating optimization, optimizing the overall structural layout and main parameters at the global level, and fine-tuning key nodes and components at the local level;

[0037] Perform multi-scale parallel computations, including macro-scale overall structural performance assessment, meso-scale critical component and node analysis, and micro-scale material properties and stress concentration area simulation;

[0038] Implement a hierarchical decision-making process, including topology layout and structural form decisions, main component size and material selection, connection node and detail optimization.

[0039] As an advantage, the method further comprises the following steps:

[0040] Conduct multi-scenario performance evaluation on the optimization solution, including normal use scenario, extreme scenario and degradation scenario;

[0041] Implement a multi-dimensional validation mechanism, including theoretical validation by comparing with classical mechanical models, numerical validation by comparing with high-precision finite element analysis, and experimental validation by comparing with scaled model test results;

[0042] Analyze the impact of parameter uncertainty and model uncertainty on the optimization scheme, and perform decision robustness analysis and sensitivity testing.

[0043] Preferably, the deep neural network has knowledge transfer and incremental learning capabilities, including:

[0044] Achieve knowledge transfer between different architectural structures through feature mapping layer;

[0045] Use meta-learning methods to accelerate the optimization process of new projects;

[0046] Build a model architecture that separates general and specific knowledge;

[0047] Implement progressive model updates to integrate new optimization experiences;

[0048] Adopt elastic weight integration to avoid catastrophic forgetting;

[0049] Design experience replay buffer to preserve key historical knowledge.

[0050] The beneficial effects of the present invention include:

[0051] 1. By establishing a dual-channel deep neural network architecture that combines physical constraints with data-driven methods, we ensure that the optimization results not only conform to mechanical principles but also make full use of the implicit patterns in historical data, thereby improving optimization efficiency and accuracy.

[0052] 2. A working condition feature extractor and a real working condition matching system were designed to accurately capture the dynamic response characteristics of building structures in complex environments, making the optimization results more in line with actual engineering needs.

[0053] 3. The use of multi-objective optimization technology can simultaneously consider multiple design goals such as structural safety, economy, sustainability, etc., find the Pareto optimal solution set, and provide designers with a variety of optimization solutions.

[0054] 4. Multi-scale and multi-granularity optimization is achieved from macro to micro, taking into account the overall structural layout, key component design and local detail processing, thereby improving the feasibility of the optimization results.

[0055] 5. Through knowledge transfer and incremental learning mechanisms, the system can continuously accumulate optimization experience, accelerate the optimization process of new projects, and improve its adaptability to different types of building structures.

[0056] 6. Equipped with a complete verification and evaluation system, the optimization results are comprehensively verified from three dimensions: theoretical, numerical and experimental. The impact of uncertainty on the optimization results is analyzed to improve the reliability of the results. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is an overall flow chart of an artificial intelligence-based building structure design optimization method of the present invention;

[0058] Figure 2 A workflow diagram for the finite element model building module of the present invention;

[0059] Figure 3 Schematic diagram of the structure of the real working condition matching system of the present invention;

[0060] Figure 4 Schematic diagram of the dual-channel deep neural network architecture of the present invention;

[0061] Figure 5 This is a working principle diagram of the working condition feature extractor of the present invention;

[0062] Figure 6 It is a workflow diagram of the multi-objective optimization system of the present invention;

[0063] Figure 7 It is a working principle diagram of the non-dominated sorting algorithm of the present invention;

[0064] Figure 8 Schematic diagram of the global-local alternating optimization strategy of the present invention;

[0065] Figure 9 Schematic diagram of the multi-scale parallel computing architecture of the present invention;

[0066] Figure 10 Schematic diagram of the knowledge transfer and incremental learning framework of the present invention. DETAILED DESCRIPTION

[0067] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0068] Example 1: Basic implementation method

[0069] like Figure 1 As shown, the present invention provides a building structure design optimization method based on artificial intelligence, comprising the following steps:

[0070] Acquire design documents and environmental data for the building structure. Specifically, the acquired design documents are typically architectural design drawings in CAD or BIM format, including floor plans, elevations, sections, and detailed structural drawings. Environmental data includes climate data, geological data, seismic zone data, and surrounding environmental information for the building's location. Preferably, the present invention utilizes a data preprocessing module to standardize the acquired design documents, converting them into a unified format, and screen key parameters within the environmental data, such as annual average wind speed, wind direction frequency distribution, earthquake intensity, and soil bearing capacity.

[0071] According to the design requirements and the design documents and environmental data, a finite element model is established, and the finite element model is consistent with the mechanical properties of the building structure. Figure 2 As shown, the present invention uses an intelligent meshing algorithm to construct a high-precision finite element model. The algorithm can be expressed as:

[0072] G=F mesh (E,P,B,C),

[0073] Where G represents the generated finite element mesh, F mesh Represents the meshing function, E represents the element type set, including beam elements, shell elements, solid elements, etc., P represents the meshing parameters, such as mesh size, shape quality control parameters, etc., B represents the boundary conditions, and C represents the mesh encryption area.

[0074] The present invention preferably automatically refines the mesh in areas of concentrated stress and de-sparses it in areas of gentle stress variation, balancing accuracy and efficiency. For example, the mesh size for beam-column joints can be set to 10-20 mm, while for ordinary wall areas, the mesh size can be set to 50-100 mm.

[0075] A load scenario is input into the finite element model to obtain the actual operating conditions under the design requirements. Specifically, the present invention first performs a modal analysis on the building structure to obtain the modal characteristics under the design requirements. Then, a database of load scenario scenarios that affect the vibration of the building structure is established. Based on the modal characteristics, the actual operating conditions of the building structure are matched from the load scenario database.

[0076] The modal analysis of the present invention adopts the subspace iteration method, which can be expressed as:

[0077] Kφ=λMφ,

[0078] Among them, K represents the structural stiffness matrix, M represents the structural mass matrix, λ represents the eigenvalue (i.e. the square of the structural natural frequency), and φ represents the eigenvector (i.e. the structural vibration mode).

[0079] Preferably, the present invention analyzes the first 10 modes of the building structure and extracts the frequency and mode coefficients of each mode as modal characteristic parameters. For a typical high-rise building, the first-order frequency is usually between 0.1-0.5 Hz, depending on the building height, stiffness, and mass distribution.

[0080] like Figure 3 As shown in Figure 1, the load condition scenario database established by the present invention contains a variety of typical load conditions, such as static loads (dead loads, live loads), wind loads, seismic loads, temperature deformation loads, etc. The present invention uses a matching algorithm based on a graph neural network to match the modal characteristics of the building structure with the load conditions in the database. The algorithm can be expressed as:

[0081]

[0082] Among them, S represents the most matching real working condition, S represents the load condition set, and M GNN represents the graph neural network matching function, and Φ represents the set of modal characteristics of the building structure.

[0083] A deep neural network is used to perform multi-objective optimization on the actual working conditions to obtain an optimization solution that meets the design requirements. Figure 4 As shown, the deep neural network of the present invention adopts a dual-channel architecture, including a physical constraint channel and a data-driven channel.

[0084] The physical constraint channel incorporates the laws of structural mechanics and material nonlinear models to ensure that the optimization results conform to mechanical principles. The core of this channel is a structural response prediction model based on the finite element method, which can be expressed as:

[0085] R p =F FEM (S,M,P),

[0086] Among them, R p represents the structural response predicted by the physical channel, F FEM Represents the finite element calculation function, S represents the structural parameters, M represents the material parameters, and P represents the load parameters.

[0087] The data-driven channel learns implicit patterns from historical optimization cases to improve optimization efficiency. This channel uses a deep convolutional neural network (DCNN) structure and can be expressed as:

[0088] R p =F FEM (S,M,P),

[0089] Among them, R p represents the structural response predicted by the physical channel, F FEM Represents the finite element calculation function, S represents the structural parameters, M represents the material parameters, and P represents the load parameters.

[0090] The outputs of the two channels are fused through a weighted fusion layer, and the weight coefficients are dynamically adjusted according to the data reliability:

[0091] R d =F DCNN (X),

[0092] Among them, R d represents the structural response predicted by the data channel, F DCNN Represents a deep convolutional neural network function, and X represents the input feature vector.

[0093] The outputs of the two channels are fused through a weighted fusion layer, and the weight coefficients are dynamically adjusted according to the data reliability:

[0094] R = αR p +(1-α)R d ,

[0095] Among them, R represents the final predicted structural response, α represents the weight coefficient, and the initial value is generally set to 0.7. It gradually decreases as data accumulates and can be reduced to a minimum of 0.3.

[0096] like Figure 5 As shown, the working condition feature extractor of the present invention adopts a hierarchical feature extraction structure to perform data preprocessing on the real working condition to obtain the working condition features corresponding to the real working condition. The working condition feature extraction process can be expressed as:

[0097] F=T(W,D,L),

[0098] Among them, F represents the extracted working condition characteristics, T represents the feature extraction function, W represents the structural parameters, D represents the material parameters, and L represents the load parameters.

[0099] Preferably, the operating characteristics of the present invention include global characteristics (such as overall deformation, base shear, etc.) and local characteristics (such as key node stress, displacement, etc.), and usually 50-100 characteristic parameters are extracted to ensure comprehensive characterization of structural performance.

[0100] The first fitness evaluator calculates the first fitness based on the working condition characteristics and load condition scenario parameters through the true optimization objective function. The true optimization objective function considers multiple objectives, including structural safety, economy, and sustainability, and can be expressed as:

[0101] f1=w1f safety +w2f cost +w3f sustainability ,

[0102] Among them, f1 represents the first fitness, w1, w2, w3 represent weight coefficients, usually set to w1 = 0.5, w2 = 0.3, w3 = 0.2, f safety 、f cost 、f sustainability represent the safety, economy and sustainability evaluation functions respectively.

[0103] The second fitness evaluator performs performance evaluation based on the working condition characteristics to obtain the second fitness. The second fitness mainly considers the robustness and construction difficulty of the structure and can be expressed as:

[0104] f2=w4f robustness +w5f constructability ,

[0105] Among them, f2 represents the second fitness, w4 and w5 represent weight coefficients, which are usually set to w4=0.6 and w5=0.4, f robustness 、f constructability Represent the robustness and construction difficulty evaluation functions, respectively. Preferably, the fitness evaluation of the present invention employs normalization processing so that each evaluation indicator is mapped to the range of 0-1. The preset threshold is typically set to 0.8. That is, when f1 > 0.8 and f2 > 0.8, the optimization result is considered to meet the requirements and the optimization process is completed; otherwise, the multi-objective optimization process continues.

[0106] The optimization scheme is converted into an executable mathematical model and the building structure is constructed. Specifically, the present invention converts the optimization scheme into a CAD or BIM model through parametric modeling technology, generates construction drawings and component parameter tables, and guides the actual construction process.

[0107] Example 2: Implementation of a real working condition matching system

[0108] Based on Example 1, the present invention further optimizes the real working condition matching system. Figure 3 As shown, before obtaining the design documents and environmental data of the building structure, the present invention further includes the following steps:

[0109] Constructing a building structure design model based on engineering knowledge. This paper uses an object-oriented parametric modeling approach to establish a comprehensive design model that encompasses structural topology, geometry, materials, and connectivity. This model employs a hierarchical structure comprising five levels: project, building, structural system, component, and connectivity.

[0110] The present invention uses a graph convolutional neural network (GCN) for topology optimization based on the building structure design model and the building structure. The network can be expressed as:

[0111]

[0112] Among them, H (l) represents the node feature matrix of the lth layer, Indicates that the adjacency matrix with self-connection is added, express The degree matrix, W (l) Represents the weight matrix of the lth layer, σ represents the activation function, and the ReLU function is usually used.

[0113] Preferably, the topology optimization of the present invention takes into account the two goals of maximizing structural stiffness and minimizing material usage, and achieves optimization by adjusting density distribution. The optimized structural topology can guide subsequent specific structural design.

[0114] Example 3: Implementation of a multi-objective optimization system

[0115] Based on Example 1, the present invention further optimizes the multi-objective optimization system. Figure 6 As shown, the present invention uses a deep neural network to perform multi-objective optimization on the actual working conditions and obtains an optimization solution that meets the design requirements, specifically:

[0116] The actual working condition is input into the deep neural network, and a non-dominated sorting algorithm is used to extract a non-dominated solution set of the actual working condition. The non-dominated solution set is evaluated based on a preset evaluation algorithm to obtain the optimal solution.

[0117] like Figure 7 As shown, the present invention adopts the fast non-dominated sorting algorithm (NSGAII) to extract the non-dominated solution set. The core steps of the algorithm include:

[0118] (1) For each individual p in the population, calculate two entities:

[0119] Dominating set S p : the set of all individuals dominated by individual p;

[0120] Dominance count n p : the number of individuals that dominate individual p;

[0121] (2) Perform non-dominated sorting on all individuals and divide them into multiple frontiers F1, F2, ...;

[0122] The first frontier F1 contains all non-dominated individuals (n p =0);

[0123] For each individual p in F1, visit its dominating set S p For each individual q in , reduce the dominance count of q by 1;

[0124] When the dominance count of an individual q becomes 0, it is placed in the second frontier F2, and this process is repeated to establish all frontiers;

[0125] (3) Calculate the congestion distance to ensure the diversity of solutions:

[0126]

[0127] Among them, d i represents the crowding distance of individual i, M represents the number of objective functions, and represents the function value of two individuals adjacent to individual i on target m, and Indicates the maximum and minimum values ​​of the target m.

[0128] Preferably, the present invention uses a genetic algorithm with an elite retention strategy for population evolution, setting the population size to 100-200, the evolutionary generations to 50-100, the crossover probability to 0.8-0.9, and the mutation probability to 0.1-0.2. The preset evaluation algorithm uses a weighted summation method, taking into account the quality and diversity of the solution, which can be expressed as:

[0129] E(S)=β1Q(S)+β2D(S),

[0130] Among them, E(S) represents the evaluation value of the solution set S, Q(S) represents the solution set quality evaluation function, D(S) represents the solution set diversity evaluation function, β1 and β2 represent weight coefficients, which are usually set to β1 = 0.7 and β2 = 0.3.

[0131] The optimized solution of the building structure is generated according to the optimal solution. The present invention adopts parametric modeling technology to automatically generate a building structure model based on optimized parameters, including geometric shape, size, material properties and connection methods.

[0132] Determine whether the preset optimization termination condition is met; if so, terminate the multi-objective optimization; otherwise, continue the multi-objective optimization. The preset optimization termination condition includes:

[0133] Reach the maximum number of iterations (usually set to 100);

[0134] The change rate of the optimal solution for 10 consecutive generations is less than 0.1%;

[0135] The algorithm running time exceeds the preset threshold (usually 2 hours);

[0136] The optimization goal reaches a preset threshold (e.g., a 20% weight reduction in the structure and all constraints are met).

[0137] Example 4: Deep Neural Network Dual-Channel Architecture

[0138] Based on Example 1, the present invention further optimizes the deep neural network architecture. Figure 4 As shown, the deep neural network of the present invention adopts a dual-channel architecture, including:

[0139] The physical constraint channel incorporates the laws of structural mechanics and material nonlinear models. Based on the principles of finite element analysis, this channel pre-sets various physical constraints, such as equilibrium equations, constitutive equations, and geometric compatibility equations. For linear elastic problems, the basic equilibrium equation can be expressed as:

[0140]

[0141] Where σ represents the stress tensor, b represents the body force, and ρ represents the density. represents the acceleration vector. The constitutive equation (Hooke's law) can be expressed as:

[0142] σ=C:ε,

[0143] Where C represents the elastic tensor and ε represents the strain tensor. The geometric compatibility equation can be expressed as:

[0144]

[0145] Where u represents the displacement vector.

[0146] Preferably, the present invention also considers material nonlinearity and adopts elastic-plastic models, such as the Drucker-Prager model or the multifaceted yield criterion, which are applicable to building materials such as concrete. 2. Data-driven channel, learning implicit patterns from historical optimization cases. This channel uses a deep learning architecture, including multiple convolutional layers and fully connected layers. The convolutional layer can be expressed as:

[0147] h l =σ(W l *h l-1 +b l ),

[0148] Among them, h l represents the output feature map of the lth layer, W l represents the convolution kernel of the lth layer, * represents the convolution operation, b l represents the bias, and σ represents the activation function.

[0149] The fully connected layer can be expressed as:

[0150] h l =σ(W l h l-1 +b l ),

[0151] Among them, Wl represents the weight matrix, h l -1 indicates the output of the previous layer, b l Represents the bias vector. 3. Channel interaction mechanism to achieve cross-channel knowledge complementarity. This paper designs a channel interaction module based on the attention mechanism, which can be expressed as:

[0152]

[0153] Among them, A represents the attention output, Q, K and V represent the query, key and value matrices respectively, which are obtained by linear transformation of the input data. k Indicates the dimension of the key vector.

[0154] The weighted fusion layer dynamically adjusts the contribution weights of the physical constraint channel and the data-driven channel. The weight adjustment mechanism can be expressed as:

[0155] α t =α t-1 +λ·ΔE,

[0156] Among them, α t Represents the weight coefficient at the current moment, α t-1 represents the weight coefficient at the previous moment, λ represents the learning rate (usually set to 0.01-0.05), and ΔE represents the difference in prediction errors between the two channels.

[0157] Preferably, the present invention assigns a higher weight (such as 0.7-0.8) to the physical constraint channel in the initial stage, and gradually increases the weight of the data-driven channel as data accumulates and model accuracy improves.

[0158] Example 5: Multi-scale and multi-granularity optimization strategy

[0159] Based on Example 1, the present invention further optimizes the multi-objective optimization strategy. Figure 8 As shown, the present invention uses a deep neural network to perform multi-objective optimization on the actual working conditions, including the following steps:

[0160] Implement global-local alternating optimization to optimize the overall structural layout and main parameters at the global level, and fine-tune key nodes and components at the local level. The global-local alternating optimization algorithm can be expressed as:

[0161]

[0162] Among them, X t represents the design variable vector for the tth iteration, F g represents the global objective function, F l represents the local objective function, α and β represent the step size parameters, which are usually set to α = 0.1 and β = 0.05.

[0163] Preferably, the present invention adopts an alternating optimization strategy, that is, first performing 10 global optimization iterations, then performing 5 local optimization iterations, and so on, alternating until convergence.

[0164] Perform multi-scale parallel calculations, including macro-scale overall structural performance evaluation, meso-scale key component and node analysis, micro-scale material properties and stress concentration area simulation. Figure 9 As shown in Figure 2, the multi-scale parallel computing architecture consists of three levels:

[0165] (1) Macroscale: Use simplified models (such as equivalent frame models) to evaluate the overall structural performance and calculate global response indicators such as period, base shear, and inter-story displacement angle.

[0166] (2) Mesoscale: Use detailed models to analyze key components and nodes, such as beam-column joints and shear wall connections, and calculate local stress, deformation, and energy dissipation indicators.

[0167] (3) Microscale: Use a sophisticated model to simulate the material constitutive behavior and stress concentration areas, and calculate cracking, plastic development, and local failure indicators.

[0168] The three-scale models communicate and collaborate via a data exchange interface. Multi-scale parallel computing significantly improves computational efficiency, reducing the time required for optimizing the structure of a typical 30-story building from the traditional 2448 hours to just 24 hours.

[0169] Implement a hierarchical decision-making process, including topology layout and structural form decision-making, main component size and material selection, connection node and detail optimization. The hierarchical decision-making process adopts a top-down strategy, namely:

[0170] (1) First-level decision: determine the structural type (such as frame structure, shear wall structure, frame shear wall structure, etc.) and the overall layout.

[0171] (2) Second-level decision-making: Determine the size, shape, and material properties of major components (such as beams, columns, and walls).

[0172] (3) Third-level decision-making: Determine the structural details and reinforcement measures of connection nodes (such as beam-column nodes and wall-beam connections).

[0173] Preferably, the present invention adopts a hierarchical genetic algorithm (HGA) to implement the above decision-making process, which sets chromosomes at different levels and performs evolutionary operations in hierarchical order.

[0174] Example 6: Evaluation and Verification System

[0175] Based on Example 1, the present invention further optimizes the evaluation and verification system. In addition to the basic optimization process, the present invention also includes the following steps:

[0176] 1. Conduct multi-scenario performance evaluation of the optimization scheme, including normal usage scenarios, extreme scenarios, and degradation scenarios.

[0177] (1) Normal use scenario: Evaluate the service performance of the structure under daily loads (dead load, live load, wind load, etc.), and calculate deformation, vibration and long-term durability indicators. Indicators include:

[0178] Maximum deformation: usually limited to 1 / 2501 / 500 of the span;

[0179] Natural frequency: The first-order frequency is usually required to be greater than 0.3Hz;

[0180] Material stress: usually limited to 70% of the allowable stress;

[0181] (2) Extreme scenarios: Evaluate the performance of the structure under extreme loads such as earthquakes, typhoons, and floods, and calculate the degree of damage and safety margin. Indicators include:

[0182] Inter-story displacement angle: usually limited to 1 / 1201 / 50, depending on the seismic fortification intensity;

[0183] Number of plastic hinges: usually required to meet the strong column and weak beam principle;

[0184] Base shear coefficient: usually between 0.0350.09, depending on the earthquake zone;

[0185] (3) Degradation scenario: Evaluate the performance of the structure under long-term effects such as material aging, corrosion, and fatigue, and calculate the structural life and maintenance requirements. Indicators include:

[0186] Fatigue life: usually required to be greater than 100 years;

[0187] Corrosion rate: usually limited to less than 0.1mm / year;

[0188] Cracking width: usually limited to 0.20.3mm;

[0189] 2. Implement a multi-dimensional verification mechanism, including theoretical verification by comparison with classical mechanical models, numerical verification by comparison with high-precision finite element analysis, and experimental verification by comparison with scaled model test results.

[0190] (1) Theoretical Verification: The optimization results are compared with the analytical solution of classical mechanics to verify their rationality. This is applicable to simple components and load conditions, such as simply supported beam deflection and frame lateral displacement.

[0191] (2) Numerical Verification: Compare the optimization results with high-precision nonlinear finite element analysis results to verify their accuracy. Usually, fine meshing and high-order elements are used to consider geometric and material nonlinearities.

[0192] (3) Experimental verification: Compare the optimization results with the scaled model test results to verify their practicality. For important structures, physical models with a scale of 1:10 or 1:5 can be made and static or vibration table tests can be carried out.

[0193] Preferably, the acceptance criteria for verification indicators set in the present invention are: theoretical verification error does not exceed 10%, numerical verification error does not exceed 5%, and experimental verification error does not exceed 15%.

[0194] 3. Analyze the impact of parameter uncertainty and model uncertainty on the optimization scheme, and perform decision robustness analysis and sensitivity testing.

[0195] (1) Parameter uncertainty analysis: Monte Carlo simulation method is used to randomly sample material parameters, load parameters and geometric parameters to evaluate their impact on the optimization results. It can be expressed as:

[0196] Y=f(X,Θ),

[0197] Where Y represents the model output vector, f represents the model function, X represents the design variable vector, and Θ represents the uncertain parameter vector, which follows a known probability distribution. Through a large number of sampling calculations, the probability distribution of the output Y is obtained and the robustness of the optimization solution is evaluated.

[0198] (2) Model uncertainty analysis: Using the Bayesian model averaging method, the prediction results of multiple models are integrated to evaluate the impact of model selection on the optimization results. It can be expressed as:

[0199]

[0200] Among them, p(y|x) represents the predicted distribution, p(y|x,M k ) represents the model M k The predicted distribution of Represents model M k Based on training data (3) Sensitivity test: Use the analysis of variance (ANOVA) or Sobol index method to quantify the influence of each uncertainty factor on the optimization result. The Sobol first-order sensitivity index can be expressed as:

[0201]

[0202] Among them, S i Represents parameter X i The sensitivity index, V iIndicated by parameter X i V(Y) represents the total variance of the output Y, V[E(Y|X i )] represents the conditional expectation E(Y|X i )’s variance.

[0203] Preferably, the present invention focuses on parameters with a sensitivity greater than 0.1 and takes corresponding control measures, such as increasing the material safety factor, adjusting the structural layout, etc.

[0204] Example 7: Knowledge Transfer and Incremental Learning Framework

[0205] Based on Example 1, the present invention further optimizes the knowledge transfer and incremental learning framework. Figure 10 As shown, the deep neural network of the present invention has the ability of knowledge transfer and incremental learning, including:

[0206] The knowledge transfer between different building structures is achieved through the feature mapping layer. The feature mapping layer can be expressed as:

[0207] F target =W trans ·F source +b trans ,

[0208] Among them, F target represents the feature representation of the target task, F source represents the feature representation of the source task, W trans represents the transformation matrix, b tRans Represents the bias vector.

[0209] Preferably, the present invention adopts a transfer learning strategy, first pre-training the model on a large-scale building structure dataset, and then fine-tuning it on a specific building type, which significantly improves the model's adaptability to new building types.

[0210] Use meta-learning to accelerate the optimization process of new projects. This paper uses the model-agnostic meta-learning algorithm (MAML), which can be expressed as:

[0211]

[0212] Among them, θ represents the model parameters, α represents the outer loop learning rate, and L i represents the loss function of task i, f θ represents the model with parameter θ, and N represents the number of tasks.

[0213] For the new task T new , the model can adapt quickly with a small number of gradient steps:

[0214]

[0215] Among them, β represents the inner loop learning rate, Represents the loss function for the new task.

[0216] Construct a model architecture that separates general and specific knowledge. The model architecture of the present invention includes the following parts:

[0217] (1) General knowledge encoder: extracts general features of building structures, such as geometric shapes, topological relationships, etc.

[0218] (2) Specific knowledge encoder: extracts specific features of building structures, such as material properties, loading conditions, etc.

[0219] (3) Hybrid reasoning module: integrates general knowledge and specific knowledge to generate optimized decisions.

[0220] Implement progressive model updates to integrate new optimization experiences. The progressive update strategy can be expressed as:

[0221]

[0222] Among them, θ t Represents the current model parameters, θ t +1 represents the updated model parameters, γ represents the learning rate, L n ew represents the loss function for new data.

[0223] Elastic weight integration is used to avoid catastrophic forgetting. The elastic weight integration method can be expressed as:

[0224]

[0225] Among them, L EWC represents the elastic weight integration loss function, L new represents the new task loss function, λ represents the regularization coefficient, and F1 represents the parameter θ i Fisher information matrix, θ i,old Represents the parameter value in the old task.

[0226] Preferably, the present invention sets λ=0.1-1.0, and adjusts it according to the similarity between the new and old tasks. The Fisher information matrix is ​​obtained by calculating the second-order derivative of the log-likelihood function with respect to the parameters.

[0227] The experience replay buffer is designed to preserve key historical knowledge. The experience replay strategy can be expressed as:

[0228] L ER (θ)=(1-α)·L new (θ)+α·L buffer (θ),

[0229] Among them, LER represents the experience replay loss function, L new Represents the loss function of new data, L buffer represents the loss function of the buffer data, α represents the balance coefficient, which is usually set to 0.3-0.5.

[0230] Preferably, the experience replay buffer of the present invention adopts an importance sampling strategy to prioritize the retention of representative historical samples. The buffer size is usually set to 10%-20% of the original data set.

[0231] In summary, this invention provides an AI-based method for optimizing building structure design. This method utilizes a dual-channel deep neural network architecture, a multi-objective optimization system, a multi-scale and multi-granularity optimization strategy, and a knowledge transfer and incremental learning framework to achieve intelligent optimization of building structure design. This method can significantly improve the efficiency and quality of building structure design, reduce material consumption and construction costs, and enhance the safety, affordability, and sustainability of structures, thus possessing significant application value and promising prospects for widespread adoption.

[0232] It should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A building structure design optimization method based on artificial intelligence, characterized in that: include: Obtain design documents and environmental data of building structures; Establishing a finite element model based on the design requirements, the design documents, and the environmental data, wherein the finite element model is consistent with the building structure in terms of mechanical properties; Inputting a load condition scenario into the finite element model to obtain a real condition under the design requirements; Using a deep neural network to perform multi-objective optimization on the actual working conditions to obtain an optimization solution that meets the design requirements; The optimization plan is converted into an executable mathematical model, and the building structure is constructed.

2. The method for optimizing building structure design based on artificial intelligence according to claim 1, characterized in that: The inputting of the load condition scenario into the finite element model to obtain the actual working condition under the design requirements is specifically as follows: Performing modal analysis on the building structure to obtain modal characteristics under the design requirements; A load condition scenario database that affects the vibration of the building structure is established, and the real working condition of the building structure is matched from the load condition scenario database based on the modal characteristics.

3. The method for optimizing building structure design based on artificial intelligence according to claim 1, characterized in that: The deep neural network includes a working condition feature extractor, a true optimization objective function, a first fitness evaluator and a second fitness evaluator; the multi-objective optimization of the true working condition using the deep neural network is specifically as follows: The working condition feature extractor performs data preprocessing on the real working condition to obtain working condition features corresponding to the real working condition; The first fitness evaluator calculates a first fitness based on the working condition characteristics and the load condition scenario parameters through the true optimization objective function; The second fitness evaluator performs performance evaluation based on the operating condition characteristics to obtain a second fitness; In which, the dimensions of the operating condition features extracted by the operating condition feature extractor and the number of performance evaluation indicators are consistent, and the first fitness is obtained by calculating the multi-objective optimization objective function; if both the first fitness evaluator and the second fitness evaluator meet the preset threshold, the multi-objective optimization is completed, otherwise, the multi-objective optimization continues.

4. The method for optimizing building structure design based on artificial intelligence according to claim 1, characterized in that: Before obtaining the design documents and environmental data of the building structure, the following steps are also included: Construct building structure design models based on engineering knowledge; Deep learning based on the building structure design model and topology optimization of the building structure.

5. The method for optimizing building structure design based on artificial intelligence according to claim 1, characterized in that: The multi-objective optimization of the actual working conditions using a deep neural network and obtaining an optimization solution that meets the design requirements is specifically as follows: Inputting the actual working condition into the deep neural network, extracting a non-dominated solution set of the actual working condition using a non-dominated sorting algorithm, and evaluating the non-dominated solution set based on a preset evaluation algorithm to obtain an optimal solution; The optimized solution of the building structure is generated according to the best solution.

6. The method for optimizing building structure design based on artificial intelligence according to claim 5, characterized in that: The method of using a deep neural network to perform multi-objective optimization on the actual working condition and obtaining an optimization solution that meets the design requirements further includes the following steps: Determine whether the preset optimization termination condition is met; If yes, then terminate the multi-objective optimization; Otherwise, continue with the multi-objective optimization.

7. The method for optimizing building structure design based on artificial intelligence according to claim 3, characterized in that: The deep neural network adopts a dual-channel architecture, including: Physical constraint channels, incorporating structural mechanics laws and material nonlinear models; Data-driven channel, learning implicit patterns from historical optimization cases; Channel interaction mechanism to achieve cross-channel knowledge complementarity; The weighted fusion layer dynamically adjusts the contribution weights of the physical constraint channel and the data-driven channel.

8. The method for optimizing building structure design based on artificial intelligence according to claim 1, characterized in that: The multi-objective optimization of the real working condition using a deep neural network comprises the following steps: Implement global-local alternating optimization, optimizing the overall structural layout and main parameters at the global level, and fine-tuning key nodes and components at the local level; Perform multi-scale parallel computations, including macro-scale overall structural performance assessment, meso-scale critical component and node analysis, and micro-scale material properties and stress concentration area simulation; Implement a hierarchical decision-making process, including topology layout and structural form decisions, main component size and material selection, connection node and detail optimization.

9. The method for optimizing building structure design based on artificial intelligence according to claim 1, characterized in that: The following steps are also included: Conduct multi-scenario performance evaluation on the optimization solution, including normal use scenario, extreme scenario and degradation scenario; Implement a multi-dimensional validation mechanism, including theoretical validation by comparing with classical mechanical models, numerical validation by comparing with high-precision finite element analysis, and experimental validation by comparing with scaled model test results; Analyze the impact of parameter uncertainty and model uncertainty on the optimization scheme, and perform decision robustness analysis and sensitivity testing.

10. The method for optimizing building structure design based on artificial intelligence according to claim 1, characterized in that: The deep neural network has knowledge transfer and incremental learning capabilities, including: Achieve knowledge transfer between different architectural structures through feature mapping layer; Use meta-learning methods to accelerate the optimization process of new projects; Build a model architecture that separates general and specific knowledge; Implement progressive model updates to integrate new optimization experiences; Adopt elastic weight integration to avoid catastrophic forgetting; Design experience replay buffer to preserve key historical knowledge.

Images