Pipe truss structure design and model selection method based on structured generative adversarial network

Abstract

The invention relates to the technical field of structural engineering optimization design and intelligent model selection, in particular to a structured generative adversarial network-based pipe truss structural design and model selection method, which comprises the following steps of: obtaining pipe truss load data and corresponding structural design schemes under different load conditions; respectively coding the load data and the design scheme to obtain a multi-dimensional dynamic load feature vector and graph structure information; inputting the two schemes into a generator, and generating a design scheme and a model selection scheme by taking the minimum steel use amount as a target under the constraint of physical information; discriminating the generation scheme and the real scheme through a discriminator, calculating an adversarial loss, and alternately training the generator and the discriminator until convergence; and finally, inputting the preset load and the structure design scheme into a convergence generator, and outputting an optimal scheme. According to the method, the end-to-end mapping from the dynamic load to the optimal design and model selection scheme is realized, the design efficiency is remarkably improved, and the structural lightweight is realized while the engineering reliability is ensured.

Description

Technical Field

[0001] This invention relates to the field of structural engineering optimization design and intelligent selection technology, and more specifically, to a method for designing and selecting tubular truss structures based on structured generative adversarial networks. Background Technology

[0002] With the global trend towards larger scale and ultra-high hub heights in wind power technology, the design of wind turbine support structures faces unprecedented engineering challenges. In pursuit of higher power generation efficiency and broader utilization of wind energy resources, support structures must withstand increasingly complex multi-dimensional dynamic loads, placing extremely high demands on the coordinated design of structural strength, stability, fatigue life, and dynamic characteristics. Traditional tubular tower structures, when the hub height exceeds 140m, are no longer sufficient to meet the economic and reliability requirements of next-generation ultra-large wind power projects due to inherent limitations such as low material utilization and high manufacturing and transportation costs.

[0003] Truss-type support structures, with their excellent lightweight, high stiffness, and high material utilization characteristics, are considered an ideal solution to overcome the bottlenecks of traditional tubular towers. (See also: Truss-type support structures) Figure 6 As shown. However, existing design and optimization methods for truss structures still have significant technical shortcomings. Truss structures have highly complex topological configurations and numerous design parameters. Currently, their optimization design mainly relies on the following two technical approaches: Path 1: Iterative optimization based on numerical simulation; This path usually combines finite element analysis software with traditional optimization algorithms. Designers need to establish a large number of numerical simulation models with different design parameters and iteratively optimize under multiple constraints such as strength, stability, fatigue and frequency. However, this method has the disadvantages of low computational efficiency, long cycle and high computational cost, which makes it difficult to meet the engineering design requirements for fast response and multiple rounds of iteration. Path Two: Regression Prediction Based on Traditional Artificial Intelligence; To address the efficiency bottleneck of numerical simulation, some studies have adopted technologies such as neural networks to construct surrogate models to predict structural performance or reverse-predict design parameters; However, these methods have new limitations: Traditional AI models are essentially "black box" regression models, highly dependent on the distribution of training data, and their prediction results cannot guarantee strict adherence to the hard constraints in physical laws and engineering specifications; while existing models are difficult to handle direct input of arbitrary dynamic load time series and lack end-to-end intelligent mapping capabilities from arbitrary load conditions to optimal structural parameters.

[0004] In summary, there is an urgent need for a tubular truss structure design and selection scheme that can break through the bottleneck of traditional numerical simulation calculations, embed engineering mechanics principles and design specifications into the model, and achieve collaborative intelligent optimization of lightweight truss structures and engineering reliability. SUMMARY

[0005] The present application aims to provide a pipe truss structure design and selection method based on a structured generative adversarial network to solve the technical problems pointed out in the background art.

[0006] The present application is implemented by the following technical scheme: a pipe truss structure design and selection method based on a structured generative adversarial network, comprising the following steps: Obtain pipe truss load data under different load conditions and corresponding pipe truss structure design schemes, which include pipe truss structure topology, geometry and material properties; Respectively encode the pipe truss load data and its corresponding pipe truss structure design scheme, including encoding the pipe truss load data using a multi-modal encoder to obtain a multi-dimensional dynamic load feature vector, and graph structure encoding the pipe truss structure design scheme to obtain graph structure information; Take the multi-dimensional dynamic load feature vector and graph structure information as the input of the generator, consider the physical information constraint, and minimize the steel consumption as the optimization objective to obtain the generated pipe truss structure design scheme and the corresponding selection scheme, which includes steel consumption and cross-sectional parameters of each member; Use the discriminator to process the generated pipe truss structure design scheme to obtain the generated pipe truss structure design scheme discrimination probability, and process the real pipe truss structure design scheme to obtain the real pipe truss structure design scheme discrimination probability; According to the generated pipe truss structure design scheme discrimination probability and the real pipe truss structure design scheme discrimination probability, determine the adversarial loss, and alternately train the generator and the discriminator until the output of the generator meets the convergence condition; For the preset pipe truss load data and the preset pipe truss structure design scheme, after encoding processing, input the converged generator to obtain the optimal selection scheme.

[0007] According to a preferred embodiment, the pipe truss structure topology includes the shape of the web member arrangement, and the geometry includes the bottom outer diameter of the member, the top-to-bottom diameter ratio, the number of edges, and the number of segments.

[0008] According to a preferred embodiment, the multi-modal encoder is based on a causal time series convolutional network, and the multi-dimensional dynamic load feature vector output by the multi-modal encoder is used to approximate the structural engineering demand parameters under the action of any dynamic load.

[0009] According to a preferred embodiment, the physical information constraint includes strength constraint loss term, stability constraint loss term, fatigue damage loss term and frequency constraint loss term, which is expressed as:

[0010] In the above formula, represents the total loss constrained by physical information, 、 、 and are the learnable weight coefficients of each loss term constrained by physical information, represents the intensity constraint loss term, represents the stability constraint loss term, represents the fatigue damage loss term, represents the frequency constraint loss term.

[0011] According to a preferred embodiment, the intensity constraint loss term needs to satisfy: the ratio of the calculated stress of all rods to the allowable stress is not greater than 1, which is expressed as:

[0012] In the above formula, represents the rod set, is the calculated stress of the rod , represents the allowable stress of the rod.

[0013] According to a preferred embodiment, the stability constraint loss term needs to satisfy: the ratio of the actual slenderness ratio of all rods to the allowable slenderness ratio, and the ratio of the actual diameter-thickness ratio to the allowable diameter-thickness ratio, are not greater than 1, which is expressed as:

[0014] In the above formula, is the slenderness ratio of the rod , is the allowable slenderness ratio of the rod, is the diameter-thickness ratio of the rod , is the allowable diameter-thickness ratio of the rod.

[0015] According to a preferred embodiment, the fatigue damage loss term needs to satisfy: the ratio of the cumulative fatigue damage of all rods to the allowable cumulative fatigue damage value is not greater than 1, which is expressed as:

[0016] In the above formula, represents the cumulative fatigue damage of the rod , represents the allowable cumulative fatigue damage value.

[0017] According to a preferred embodiment, the frequency constraint loss term needs to satisfy: the absolute value of the difference between the first-order characteristic frequency of the tube truss structure and the excitation frequency is not less than 15% of the excitation frequency, which is expressed as:

[0018] In the above formula, represents the excitation frequency of the wind turbine 1P or 3P, represents the first-order characteristic frequency of the tube truss structure.

[0019] According to a preferred embodiment, the total optimization target expression of the generator is as follows:

[0020] In the above formula, represents the total optimization target of the generator, , and is the weight coefficient of each sub-target of the total optimization target, represents the amount of steel used, represents the adversarial loss.

[0021] According to a preferred embodiment, the alternating training of the generator and the discriminator adopts a staged training strategy, specifically: The first stage is set to zero to minimize the physical information constraint as the goal, so that the generator quickly learns and outputs a feasible tube truss structure design scheme and the corresponding selection scheme; The second stage introduces a non-zero , which drives the generator to converge to the optimal solution of the adversarial loss constraint under the supervision of the discriminator, with the goal of minimizing .

[0022] The technical scheme of the tube truss structure design and selection method based on the structured generative adversarial network provided by the present application has at least the following advantages and beneficial effects: (1) The present method can realize end-to-end mapping from multi-dimensional dynamic load input to optimal selection scheme output. Compared with traditional numerical simulation methods, the design efficiency and response speed are both significantly improved while effectively reducing the time-consuming of iterative optimization calculation; (2) By considering the physical information constraint and minimizing the steel usage as the optimization target, the generator can accurately search and converge to the globally optimal and lightest scheme under the premise of ensuring engineering reliability; (3) Through the combination of multi-modal encoder and generator, the model has the ability to handle arbitrary dynamic load conditions and complex structure topologies, greatly expanding its application range in the next generation of wind power structure design. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is the overall flowchart of the tube truss structure design and selection method based on the structured generative adversarial network provided by the present application for Example 1; Figure 2A schematic diagram of the principle of edge condition message passing of the graph neural network provided for Embodiment 1 of the present application is shown in Figure 3 A schematic diagram of the network connection of the generator and the discriminator provided for Embodiment 1 of the present application is shown in Figure 4 A schematic diagram of the principle of embedding of the structural reliability constraint of the physical information constraint loss provided for Embodiment 1 of the present application is shown in Figure 5 An engineering constraint schematic diagram of the physical information constraint loss function provided for Embodiment 1 of the present application is shown in Figure 6 A structural schematic diagram of the truss support structure referred to in the background is shown in DETAILED DESCRIPTION

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0025] Embodiment 1 The embodiments of the present application provide a pipe truss structure design and selection method based on a structured generative adversarial network, Figure 1 A schematic diagram of the overall process of the pipe truss structure design and selection method is shown in Figure 1 The pipe truss structure design and selection method includes the following steps: Step S1, data preparation and coding; First, a data set for model training is obtained, which includes multiple groups of pipe truss load time series under different load conditions, and a verified pipe truss structure design scheme corresponding to each group of load time series. The load time series can be derived from a wind turbine overall simulation software, such as GH-Bladed software. The structure design scheme includes topology (such as web member arrangement shape), geometry (bar member bottom outer diameter, top-to-bottom diameter ratio, number of edges, number of segments), and material properties.

[0026] Subsequently, the two types of data are coded and processed to convert them into feature representations suitable for neural network processing, as follows: Load coding: a multi-modal encoder based on a causal time series convolution network is used to process the load time series, extract and compress its dynamic features, and form a fixed-dimensional multi-dimensional dynamic load feature vector, which aims to represent the comprehensive effect of the load on the structure; Structural encoding: The structural design scheme is abstracted into a graph structure, with nodes representing connection points and edges representing rods. A graph neural network is used to encode the graph structure to obtain graph structure information containing topological and geometric relationships.

[0027] Step S2, Generator Design; In this embodiment, this step aims to construct a generator model that employs a graph neural network (GNN) with an edge-conditional message passing mechanism; see [link to previous section]. Figure 2 As shown, the generator model takes the multidimensional dynamic load feature vector and graph structure information obtained in step S1 as its condition input, and under the premise of taking physical information constraints, takes minimizing the amount of steel used as the optimization direction, and outputs a generated truss structure design scheme and corresponding selection scheme. The selection scheme includes the amount of steel used and the cross-sectional parameters of each member.

[0028] Combination Figure 2 As shown, the generator employs a graph neural network (GNN) edge-conditional message passing mechanism, specifically implemented as follows: The multi-dimensional dynamic load feature vector and graph structure information are used as initial inputs; for nodes in the graph structure... and nodes ,side For members of a tubular truss, the edge condition information they carry includes the member's geometric parameters and material properties; edge It will receive from the node and nodes The message, and generate a message accordingly. Subsequently, regarding the message With nodes The node's own state is aggregated, and the node state is iteratively optimized using an update function. After the GNN iterates through L layers, the output layer is decoded by an MLP to finally output the design scheme and selection scheme for the truss structure.

[0029] The physical information constraints are embedded in the form of a loss function, including: The strength constraint loss term must satisfy the following condition: the ratio of the calculated stress to the allowable stress of all members is not greater than 1, expressed as:

[0030] In the above formula, Represents a set of rods, For rods Calculated stress, This indicates the allowable stress of the member.

[0031] The stability constraint loss term must satisfy the following condition: the ratio of the actual slenderness ratio to the allowable slenderness ratio of all members, and the ratio of the actual diameter-to-thickness ratio to the allowable diameter-to-thickness ratio, must not exceed 1, expressed as:

[0032] In the above formula, is the slenderness ratio of the rod member is the allowable slenderness ratio of the rod member is the slenderness ratio of the rod member is the diameter-thickness ratio of the rod member is the allowable diameter-thickness ratio of the rod member is the diameter-thickness ratio of the rod member

[0033] The fatigue damage loss term, as shown in Figure 5 , according to the DNV-RP-C203 and IIW criteria, the nominal stress method and SN curve are used for fatigue evaluation, and it needs to meet: the ratio of the cumulative fatigue damage of all rod members to the allowable cumulative fatigue damage value is not greater than 1, which is expressed as:

[0034] In the above formula, represents the cumulative fatigue damage of the rod member represents the cumulative fatigue damage of the rod member represents the allowable cumulative fatigue damage value. As shown in Figure 5 , the fatigue grade FAT of the tower column member is 71, and the FAT grade of the diagonal rod DB and the cross rod CB is 45.

[0035] The frequency constraint loss term, as shown in Figure 5 , in order to avoid resonance, it needs to meet: the absolute value of the difference between the first-order characteristic frequency of the tube truss structure and the excitation frequency is not less than 15% of the excitation frequency, which is expressed as:

[0036] In the above formula, represents the excitation frequency of the wind turbine 1P or 3P represents the first-order characteristic frequency of the tube truss structure. As shown in Figure 5 , for example, a certain 5.5 MW wind turbine, the rated speed of the fan is 9.57 rpm, and the corresponding excitation frequencies are 0.1595 Hz (1P) and 0.4785 Hz (3P), respectively.

[0037] As shown in Figure 4 , in this embodiment, the above loss terms are combined as physical information constraint total loss to guide the generator model to follow the engineering specifications, and the expression is as follows:

[0038] In the above formula, , , and The physical information constrains the learnable weight coefficients of each loss term.

[0039] It should be noted that the combination of the multi-modal encoder and the generator in the embodiment enables the model to process any dynamic load working condition and complex structure topology, greatly expanding its application range in the next generation of wind power structure design.

[0040] Step S3, discriminator design; In this embodiment, this step aims to build a discriminator model for evaluating the rationality of the generated tubular truss structure design scheme output by the generator model; specifically, referring to Figure 3 As shown in the figure, it includes: Step S301, input the generated tubular truss structure design scheme generated in step S2 into the discriminator model to obtain the generated tubular truss structure design scheme discrimination probability; Step S302, input the real tubular truss structure design scheme into the same discriminator model to obtain the real tubular truss structure design scheme discrimination probability; Step S303, calculate the adversarial loss according to the generated tubular truss structure design scheme discrimination probability and the real tubular truss structure design scheme discrimination probability, to measure the difference between the generated scheme and the real scheme distribution.

[0041] In combination with Figure 3 As shown in the figure, after the tubular truss load data and the corresponding structure design scheme provided by the data set are encoded in step S1, the multi-dimensional dynamic load feature vector and the graph structure information are output and sent to the generator; the generator obtains the generated tubular truss structure design scheme and the corresponding selection scheme according to the total optimization target, and inputs them into the discriminator together with the real tubular truss structure design scheme in the data set; the discriminator outputs the respective discrimination probabilities by discriminating the two types of schemes; the adversarial loss calculated from the two types of discrimination probabilities and the physical information constraint loss are fed back to the generator to optimize the generator.

[0042] Step S4, alternating training; In this embodiment, based on steps S2 and S3, the generator model and the discriminator model are alternately iteratively trained; The total optimization target expression of the generator model is as follows:

[0043] In the above formula, denotes the total optimization target of the generator model, , and are the weight coefficients of each sub-target of the total optimization target, denotes the steel consumption, denotes the adversarial loss.

[0044] It should be noted that in the embodiment, by taking into account the physical information constraint and taking minimizing the steel consumption as the optimization goal, the generator can accurately search and converge to the globally optimal and lightest solution under the premise of ensuring engineering reliability.

[0045] To improve the training efficiency and stability, the embodiment adopts a phased training strategy, which is as follows: The first phase is set to zero to minimize the physical information constraint as the goal, so that the generator model learns quickly and outputs a feasible pipe truss structure design scheme and the corresponding selection scheme; The second phase introduces a non-zero , drives the generator model to converge to the optimal solution of the adversarial loss constraint under the supervision of the discriminator model.

[0046] Step S5, optimal selection scheme generation; In the embodiment, for the preset pipe truss load data and the preset pipe truss structure design scheme, after the same encoding processing as step S1 is performed, the converged generator model is input, and the optimal selection scheme can be obtained.

[0047] In summary, the method can realize end-to-end mapping from multi-dimensional dynamic load input to optimal selection scheme output. Compared with the traditional numerical simulation method, the design efficiency and response speed are also significantly improved while effectively reducing the time-consuming of iterative optimization calculation.

[0048] In addition, the preset pipe truss structure design scheme and the corresponding optimal selection scheme can be further subjected to regional seismic damage assessment to determine the seismic economic loss under the action of the earthquake, which will not be described in detail here.

[0049] The above is only a preferred embodiment of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.​

Claims

1. A method for design and selection of tubular truss structure based on structured generative adversarial network, characterized in that, The method comprises the following steps: Obtaining pipe truss load data under different load conditions and corresponding pipe truss structure design schemes, wherein the pipe truss structure design schemes include pipe truss structure topology, geometry and material properties; Respectively, the pipe truss load data and the corresponding pipe truss structure design scheme are coded, including using a multi-modal encoder to encode the pipe truss load data to obtain a multi-dimensional dynamic load feature vector, and graph structure coding is performed on the pipe truss structure design scheme to obtain graph structure information; Taking the multi-dimensional dynamic load feature vector and the graph structure information as the input of the generator, considering the physical information constraint, and taking the minimization of the steel consumption as the optimization objective, the generated pipe truss structure design scheme and the corresponding selection scheme are obtained, wherein the selection scheme includes the steel consumption and the cross-section parameters of each member; Using the discriminator to process the generated pipe truss structure design scheme to obtain the generated pipe truss structure design scheme discrimination probability, and processing the real pipe truss structure design scheme to obtain the real pipe truss structure design scheme discrimination probability; According to the generated pipe truss structure design scheme discrimination probability and the real pipe truss structure design scheme discrimination probability, the adversarial loss is determined, and the generator and the discriminator are alternately trained until the output of the generator meets the convergence condition; For the preset pipe truss load data and the preset pipe truss structure design scheme, after coding processing, the optimal selection scheme is obtained by inputting the converged generator. 2.The method of claim 1, wherein, The pipe truss structure topology includes the shape of the web member arrangement, and the geometry includes the bottom outer diameter of the member, the top-to-bottom diameter ratio, the number of edges and the number of segments. 3.The method of claim 1, wherein, The multi-modal encoder is based on a causal time series convolution network, and the multi-dimensional dynamic load feature vector output by the multi-modal encoder is used to approximate the structural engineering demand parameters under the action of any dynamic load. 4.The method of claim 1, wherein, The physical information constraint includes a strength constraint loss term, a stability constraint loss term, a fatigue damage loss term and a frequency constraint loss term, and is expressed as: In the above formula, represents the total loss of physical information constraints, , , and are the learnable weight coefficients of each loss term of the physical information constraint, represents the intensity constraint loss term, represents the stability constraint loss term, represents the fatigue damage loss term, represents the frequency constraint loss term.

5. The method of claim 4, wherein the method is based on a structured generative adversarial network. The expression of the strength constraint loss term is as follows: In the above formulae, denotes a set of bars, is a bar the calculated stress, denotes the allowable stress of the bar. 6.The method of designing and selecting a tubular truss structure based on a structured generative adversarial network according to claim 4, wherein, The expression of the stability constraint loss term is as follows: In the above formula, Represents a set of rods, For rods The slenderness ratio, For the allowable slenderness ratio of the rod, For rods Aspect ratio, This refers to the allowable diameter-to-thickness ratio of the rod. 7.The method of claim 4, wherein, The expression of the fatigue damage loss term is as follows: In the above formula, denotes a set of bars, denotes a bar cumulative fatigue damage of the bar, denotes an allowable cumulative fatigue damage value. 8.The method of claim 4, wherein, The expression of the frequency constraint loss term is as follows: In the above formula, denotes the excitation frequency of the wind turbine 1P or 3P, denotes the first eigenfrequency of the tube truss structure. 9.The method of claim 4, wherein, The total optimization objective expression of the generator is as follows: In the above formula, denotes the total optimization target of the generator, , and is a weight coefficient of each sub-target of the total optimization target, denotes the amount of steel used, denotes the adversarial loss. 10.The method of designing and selecting a tubular truss structure based on a structured generative adversarial network according to claim 9, wherein, The alternating training of the generator and the discriminator adopts a staged training strategy, specifically as follows: First stage setup to zero to minimize physical information constraints to target, so that the generator learns quickly and outputs a feasible tube truss structure design scheme and the corresponding selection scheme; The second stage introduces non-zero , drives the generator under the supervision of the discriminator, with the goal of minimizing the optimal solution to the adversarial loss constraint.

Images