Truss structure transmission tower generation method and related device based on generative adversarial network

Abstract

The present application relates to the field of graph generation technology and discloses a method and related apparatus for generating a truss structure transmission tower based on a generative adversarial network. The method comprises: performing a transmission tower graph structure generation operation based on the vertical height, bottle mouth width, and root opening width of the transmission tower to obtain a transmission tower graph structure dataset; using the transmission tower graph structure dataset to perform adversarial training on an initial TGN model for truss structure transmission tower generation to obtain a target TGN model, wherein the initial TGN model is constrained by a conditional control vector and includes a GVAE generator and a GCN discriminator; and using the target TGN model to generate a truss structure transmission tower in a target scenario to obtain a generated transmission tower. The present application can solve the technical problem that existing truss structure transmission tower generation technologies have defects such as complex model operation, high training difficulty, and low efficiency, which lead to poor transmission tower design effects.

Description

Technical Field

[0001] The present application relates to the field of graph generation technology, and in particular to a method and related device for generating a truss structure transmission tower based on a generative adversarial network. Background Art

[0002] The structural design of transmission towers requires comprehensive consideration from multiple aspects, including safety, environmental protection, convenience, and economy. Traditional transmission tower structural design methods often use digital design forms assisted by SolidWorks and CAD as hardware. Digital models are manually constructed, and the models are verified by mechanical calculations to perform structural optimization. This method requires a lot of manual operation, and designers must also have many years of experience in tower design. With the development of AI generative algorithms, tower design has begun to extend into the field of automation. Computers can automatically help implement structural modeling, mechanical calculations, system optimization, and other operations. However, this method requires designers to set constraints and specify optimization directions in advance. Computers themselves do not have the ability to learn independently. At the same time, the priority of each constraint must be considered to alleviate potential conflicts.

[0003] Existing intelligent tower structure design generally involves topology generation and subsequent tower material configuration using data-driven neural networks as the computational core. Unlike the initial exploration of shear wall placement in house-style buildings and the beam-and-rod structure placement problem using generative adversarial networks as the computational core, truss structure design requires training with a non-given node set, a graph generation problem. Current neural network-based truss structure graph generation technology suffers from limitations such as complex model operation, difficulty in training, and low efficiency. These limitations result in relatively poor efficiency and effectiveness in actual transmission tower design, making it difficult to adapt to practical application requirements. Summary of the Invention

[0004] The present application provides a truss structure transmission tower generation method and related devices based on a generative adversarial network, which are used to solve the technical problems that the existing truss structure transmission tower generation technology has defects such as complex model operation, high training difficulty and low efficiency, resulting in poor transmission tower design effects.

[0005] In view of this, the first aspect of the present application provides a method for generating a truss structure transmission tower based on a generative adversarial network, comprising:

[0006] Based on the vertical height, bottle mouth width and root opening width of the transmission tower, a transmission tower graph structure generation operation is performed to obtain a transmission tower graph structure dataset;

[0007] Using the transmission tower graph structure dataset to perform generative adversarial training on an initial TGN model for truss structure transmission towers, to obtain a target TGN model, wherein the initial TGN model is constrained by a conditional control vector and includes a GVAE generator and a GCN discriminator;

[0008] The target TGN model is used to generate a truss structure transmission tower in a target scenario to obtain a generated transmission tower.

[0009] Preferably, the transmission tower graph structure generation operation is performed based on the vertical height of the tower section, the bottle mouth width, and the root opening width of the transmission tower to obtain the transmission tower graph structure dataset, including:

[0010] The transmission towers are sorted and screened based on the vertical height, bottle mouth width and root opening width of the transmission tower sections, and multiple transmission tower foundation structural elements are generated;

[0011] Performing a graph structure splicing operation according to the transmission tower basic structure element to obtain graph structure data;

[0012] The graph structure data is converted into a node feature file and an adjacency relationship file to obtain a transmission tower graph structure data set.

[0013] Preferably, the transmission tower graph structure dataset is used to perform generative adversarial training on the initial TGN model for truss structure transmission towers to obtain a target TGN model, wherein the initial TGN model is constrained by a conditional control vector and includes a GVAE generator and a GCN discriminator, and also includes:

[0014] By introducing latent variables, combining GCN encoder and FNN decoder to build an initial generator, and configuring the generation constraints of the conditional control vector for the initial generator, we can get the GVAE generator.

[0015] Construct a GCN discriminator through the GCN network and FNN network framework;

[0016] The initial TGN model is obtained by combining the GVAE generator and the GCN discriminator according to the comprehensive loss function.

[0017] Preferably, the comprehensive loss function of the initial TGN model is expressed as:

[0018]

[0019] in, is the prior loss, is the reconstruction loss, To generate the loss, 、 is a variable weighting parameter.

[0020] The second aspect of the present application provides a truss structure transmission tower generation device based on a generative adversarial network, comprising:

[0021] A data generation unit is used to generate a transmission tower graph structure based on the vertical height, bottle mouth width and root opening width of the transmission tower to obtain a transmission tower graph structure data set;

[0022] A model training unit is used to perform generative adversarial training of a truss structure transmission tower on an initial TGN model using the transmission tower graph structure dataset to obtain a target TGN model, wherein the initial TGN model is constrained by a conditional control vector and includes a GVAE generator and a GCN discriminator;

[0023] The target generation unit is used to generate a truss structure transmission tower in a target scenario by using the target TGN model to obtain a generated transmission tower.

[0024] Preferably, the data generating unit is specifically used to:

[0025] The transmission towers are sorted and screened based on the vertical height, bottle mouth width and root opening width of the transmission tower sections, and multiple transmission tower foundation structural elements are generated;

[0026] Performing a graph structure splicing operation according to the transmission tower basic structure element to obtain graph structure data;

[0027] The graph structure data is converted into a node feature file and an adjacency relationship file to obtain a transmission tower graph structure data set.

[0028] Preferably, it also includes:

[0029] A generator construction unit is used to construct an initial generator by combining the introduced latent variables with the GCN encoder and the FNN decoder, and configure the generation constraints of the conditional control vector for the initial generator to obtain a GVAE generator;

[0030] The discriminator construction unit is used to build a GCN discriminator through the GCN network and FNN network framework;

[0031] A model generation unit is used to obtain an initial TGN model by combining the GVAE generator and the GCN discriminator according to a comprehensive loss function.

[0032] Preferably, the comprehensive loss function of the initial TGN model is expressed as:

[0033]

[0034] in, is the prior loss, is the reconstruction loss, To generate the loss, 、 is a variable weighting parameter.

[0035] A third aspect of the present application provides a truss structure transmission tower generation device based on a generative adversarial network, the device comprising a processor and a memory;

[0036] The memory is used to store program code and transmit the program code to the processor;

[0037] The processor is used to execute the truss structure transmission tower generation method based on the generative adversarial network described in the first aspect according to the instructions in the program code.

[0038] In a fourth aspect, the present application provides a computer-readable storage medium for storing program code, wherein the program code is used to execute the method for generating a truss structure transmission tower based on a generative adversarial network as described in the first aspect.

[0039] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0040] In the present application, a method for generating a truss structure transmission tower based on a generative adversarial network is provided, comprising: performing a transmission tower graph structure generation operation based on the vertical height, bottle mouth width, and root opening width of the transmission tower to obtain a transmission tower graph structure dataset; using the transmission tower graph structure dataset to perform generative adversarial training of a truss structure transmission tower on an initial TGN model to obtain a target TGN model, wherein the initial TGN model is constrained by a conditional control vector and includes a GVAE generator and a GCN discriminator; and using the target TGN model to generate a truss structure transmission tower in a target scenario to obtain a generated transmission tower.

[0041] The present application provides a method for generating a truss structure transmission tower based on a generative adversarial network. The method generates a transmission tower graph structure dataset based on the transmission tower structure feature data, and performs generative adversarial training on the constructed initial TGN model based on this dataset. This method can optimize the model's truss structure transmission tower graph generation performance, thereby obtaining a target TGN model. The generative adversarial model can quickly generate a truss structure transmission tower graph with good results. This process constructs a targeted training dataset for model training, making the model training more feasible. In addition, the constructed generative adversarial model structure is relatively simple, and the data processing is not complicated, which can improve the model's graph structure generation efficiency to a certain extent. Therefore, the present application can solve the technical problem that the existing truss structure transmission tower generation technology has defects such as complex model operation, high training difficulty, and low efficiency, which leads to poor transmission tower design results. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic diagram of a flow chart of a method for generating a truss structure transmission tower based on a generative adversarial network according to an embodiment of the present application;

[0043] Figure 2 A schematic diagram of the structure of a truss structure transmission tower generation device based on a generative adversarial network provided in an embodiment of the present application;

[0044] Figure 3 Schematic diagrams of 8 typical transmission tower truss structural elements provided in the embodiments of this application;

[0045] Figure 4 An example diagram of a transmission tower structure generated based on a transmission tower truss structure element provided in an embodiment of the present application;

[0046] Figure 5 A schematic diagram of the initial TGN model network structure provided in an embodiment of the present application;

[0047] Figure 6 A schematic diagram comparing the discriminant values of different samples from different batches provided for the experimental case of this application;

[0048] Figure 7 Schematic diagram of the comprehensive loss function change of the TGN model provided for the experimental case of this application;

[0049] Figure 8 Schematic diagram of the changes in the prior loss and reconstruction loss functions provided for the experimental case of this application;

[0050] Figure 9 An example diagram of the transmission tower segment structure generated under different training batches provided for the experimental case of this application. DETAILED DESCRIPTION

[0051] In order to help those skilled in the art better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0052] For easier understanding, see Figure 1 The present application provides an embodiment of a method for generating a truss structure transmission tower based on a generative adversarial network, comprising:

[0053] Step 101: Perform a transmission tower graph structure generation operation based on the vertical height, bottle mouth width, and root opening width of the transmission tower to obtain a transmission tower graph structure dataset.

[0054] Furthermore, step 101 includes:

[0055] The transmission towers are sorted and screened based on the vertical height, bottle mouth width and root opening width of the transmission tower sections, and multiple transmission tower foundation structural elements are generated;

[0056] Perform graph structure splicing operations based on the basic structural elements of the transmission tower to obtain graph structure data;

[0057] The graph structure data is converted into node feature files and adjacency relationship files to obtain the transmission tower graph structure dataset.

[0058] It should be noted that in order to achieve network training, the application of creating graph structure datasets based on established generator solutions in the field of truss structure design is not mature, so there is currently a lack of transmission tower truss graph structure datasets available for training. Therefore, this embodiment generates multiple transmission tower basic structure elements based on the vertical height, bottleneck width, and root opening width of the transmission tower segments, sorted and filtered by the number of occurrences; please refer to the detailed information. Figure 3 The figure shows eight typical transmission tower truss structural elements. These eight structures are then used as the basic structural elements of the transmission tower. The dataset is constructed by randomly splicing the eight structures according to the upper and lower root opening ranges. Please refer to Table 1. Taking structure 1 as the height unit, it can be inferred that structure 2 has two segments, structures 3-4 have three segments, and structures 5-8 have four segments. The total number of spliced segments is mainly 5-10. Each spliced structure case contains 2-3 structural elements. A total of 200 graph structure data were produced and converted into node feature and adjacency relationship files for graph generation network training.

[0059] Table 1 Boundary condition range of transmission tower foundation structure element

[0060]

[0061] According to the above transmission tower basic structure elements, we can get the following Figure 4 The tower section shown in the figure is randomly spliced when the width of the bottle mouth and the width of the root opening meet the splicing conditions. Among them, the node feature is the node coordinate [X n ,Y n ], which is embedded and fused with the adjacency matrix as node embedding features, and the tower section contour features are used as conditions to control the generated graph structure at the graph level. In the graph dataset, the node adjacency matrix is in the form of a pair of vector groups, the maximum number of graph edges is 60, the number of nodes is 32, and all graphs are symmetrical structures. Based on this, a transmission tower graph structure dataset for network training can be constructed, which contains a node feature file and an adjacency relationship file. The node feature file contains the positional relationship of the nodes, and the adjacency relationship file contains the topological features of the graph. The training set and test set of the entire dataset are divided into a ratio of 8:2.

[0062] Step 102: Use the transmission tower graph structure dataset to perform generative adversarial training on the initial TGN model for truss structure transmission towers to obtain a target TGN model. The initial TGN model is constrained by a conditional control vector and includes a GVAE generator and a GCN discriminator.

[0063] Furthermore, before step 102, the following steps are also included:

[0064] By introducing latent variables, combining GCN encoder and FNN decoder to build the initial generator, and configuring the generation constraints of the conditional control vector for the initial generator, we get the GVAE generator.

[0065] Construct a GCN discriminator through the GCN network and FNN network framework;

[0066] The initial TGN model is obtained by combining the GVAE generator and the GCN discriminator based on the comprehensive loss function.

[0067] Furthermore, the comprehensive loss function of the initial TGN model is expressed as:

[0068]

[0069] in, is the prior loss, is the reconstruction loss, To generate the loss, 、 is a variable weighting parameter.

[0070] See also Figure 5The TGN (Truss Generation Network) model used in this embodiment for truss transmission tower generation consists of a GVAE generator (top layer) and a GCN discriminator (bottom layer). The GVAE (Generation Variational Auto Encoder) generator uses latent space conjugation to convert node embeddings into envelope-invariant graph-level encodings, allowing learning and generation of resizable graphs on any number of graphs. The GCN discriminator (bottom layer) performs discriminant analysis of images. Specifically, the network consists of two parts: the generator and the discriminator. The generator is actually a variational autoencoder. It uses the latent vector distribution encoded in the dataset and the sample labels (i.e., vertical height, bottle mouth width, and root opening width) as network inputs. The network outputs the reconstructed graph structure and the generated graph structure. The reconstructed graph structure is input to Discriminator 1 and the previous latent distribution to calculate the reconstruction loss, which is used to measure the training status within the variational autoencoder. The generated graph structure means relying on the new structure generated by the label and decoder. The generation loss is calculated through Discriminator 1 and Discriminator 2 and is used to measure the training status of the overall network architecture. When the node order changes, the model can maintain a one-to-one mapping between the graph and the selected structure without retraining. The model can explore graphs of different sizes and arbitrary node orders. It also strengthens the boundary conditions to control the mapping between the vector space and the node structure space, ensuring model efficiency and robustness.

[0071] It can be found that the GVAE generator of this embodiment is constructed based on a GCN (Graph Convolutional Network) encoder and an FNN (Feedforward Neural Network) decoder, and is also configured with a conditional control vector for image generation constraints; while the GCN discriminator is constructed based on a GCN network and a two-layer FNN network. In addition, the conditional control vector mainly considers three factors, namely the vertical height of the tower segment, the bottle mouth width, and the root opening width. To facilitate network parameter propagation, this embodiment performs a small-scale staged merging. The vertical height of the tower segment after merging is divided into 6 categories, namely 5 to 10 basic tower segment structures, the bottle mouth width is divided into 8 categories, each 1000mm is a category, namely 0 to 8000mm; the root opening width is divided into 10 categories, each 1000mm is a category, namely 0 to 10000mm. All three are composed of the conditional control vector of the graph structure in the form of a one-hot vector. ; Based on the classification, we can see that the dimension of the conditional control vector is 6+8+10=24.

[0072] In the GVAE generator, the graph generation structure adjacency matrix is feature-embedded through the Encoder graph convolutional network encoder. That is, given a graph structure G={V, E}, the node features X=X(A) are generated according to the graph structure adjacency matrix A and embedded with the standard k-dim spectrum. Gaussian random vectors can also be used instead of spectral embedding. There is no specific limitation.

[0073] It should be noted that this embodiment introduces latent variables in the encoding stage of the GVAE generator , which is calculated by the node feature X and the structural adjacency matrix A, can be expressed as:

[0074]

[0075] Where n represents the number of nodes, Indicates that under the given observation data X and graph structure A, the i-th hidden variable The posterior distribution of Can be regarded as a graph structure node Node embedding features, where Representing a set of graph-structured nodes by using a single latent space vector distribution For all Modeling, specifically expressed as:

[0076]

[0077] in, 、 、 Represents the mean of the latent space vector, represented by the function The calculated mean, obtained by the function The calculated standard deviation is represents a diagonal matrix of squared standard deviations, Indicates is the Gaussian distribution with the covariance matrix of the mean; It is a two-layer GCN network, where is the symmetric normalized adjacency matrix of the graph structure G, expressed as , is its degree matrix, expressed as , 、 These are all network parameters; 、 Represent the mean and standard deviation matrices respectively, and share the first layer network parameters ; yes The mean or standard deviation of the i-th row.

[0078] This embodiment introduces latent variables in the encoding stage of the generator Not only can the model be trained on different sets of node graph structures, but it can also enter graph-level variational coding, and at the same time consider the influence of conditional control vectors on network parameters, and ensure the permutation invariance of the model. Specifically, the graph structure generator used in this embodiment can be obtained by Based on this, a graph of any size m can be generated; not only that, the conditional control vector can be synchronously embedded during the training of the generator, thereby realizing the training control of the network parameters.

[0079] For the ideal number After sampling, in order to improve the decoding performance of the GCN decoder in the GVAE generator, several layers of fully connected feedforward neural networks f are attached to the GCN decoder before the sigmoid function operation. The specific expression is:

[0080]

[0081] in, Indicates that given latent variables Under the condition of , the probability of graph structure A is, Indicates that given latent variables and Under the condition of , the probability that there is an edge between node i and node j is, Represents the sigmoid function, which is used to map the input to the (0,1) interval. Represents latent variables The result after the fully connected feedforward neural network transformation can be expressed as follows by minimizing the negative variation lower bound:

[0082]

[0083]

[0084]

[0085] in, is the overall loss of the variational autoencoder, i.e., the loss of the GVAE generator; The loss function is essentially a reconstruction term of the variational autoencoder loss function, which means that its decoder reconstructs the expected log-likelihood of the input data Z based on the latent variable Z; The loss function is essentially the KL divergence term of the variational autoencoder loss function, which is used to describe the potential posterior distribution of the generated graph structure. With Gaussian prior distribution At the same time, the KL divergence term plays a role in restraining and balancing the reconstruction term to avoid overfitting of the model; Indicates that under the given observation data X and graph structure A, the hidden variable The posterior distribution of represents the Kullback-Leibler divergence, represents the Gaussian prior distribution, which can be expressed as:

[0086]

[0087] in, represents the identity matrix; is the mean of the latent space vector; Represents n mutually independent vectors with a mean of 0 and a covariance matrix that is the identity matrix The joint probability distribution model of the multivariate Gaussian distribution. is the Gaussian prior distribution based on Kullback-Leibler divergence The prior loss of The above is the encoding and decoding calculation process in the GVAE generator of the TGN model in this embodiment.

[0088] Since the graph structure data generation process of the GVAE generator is still not permutation-invariant, this embodiment constructs a permutation-invariant graph structure discriminator based on the GCN network and the two-layer FNN network, namely the GCN discriminator; the discriminator can learn based on the link reconstruction loss The generated adjacency matrix With the original adjacency matrix Specifically, a two-layer GCN network is constructed and then connected to a two-layer FNN network. From the overall model point of view, it is trained together with the encoder and decoder of the GVAE generator mentioned above. The network training of the graph structure is completed by optimizing the GAN generation loss of the minimax game. The GAN generation loss is expressed as:

[0089]

[0090] in, 、 、 Represents spectral embedding, through After the operation, the encoding 、 is permutation invariant, that is ,in, Represents a matrix rearrangement operation.

[0091] Comprehensive loss function of the initial TGN model Variable weighting parameters in 、 The weights of different losses can be adjusted. In the process of using the transmission tower graph structure dataset to perform adversarial training on the initial TGN model for truss structure transmission tower generation, the combination function is not used to update all model parameters. Instead, the following update rules are used in each training batch to update the model parameters:

[0092]

[0093] in, 、 、 Represent the weight matrices of the encoder, decoder / generator, and discriminator respectively, 、 、 Represent the back propagation gradients of the three respectively; Represents the update operation of the model parameters, that is, updating the corresponding model weight matrix based on the current gradient calculation.

[0094] Combined loss function , the transmission tower graph structure dataset and the above update rule formula can be used to perform generative adversarial training on the initial TGN model to obtain the target TGN model.

[0095] Step 103: Use the target TGN model to generate a truss structure transmission tower in the target scenario to obtain a generated transmission tower.

[0096] The target TGN model can be directly applied to actual target scenarios. By inputting data requirements such as the actual vertical height, bottleneck width, and root opening width, the target TGN model can generate truss transmission towers. The specific generation process is similar to the data processing during the training process and is not detailed here. The model is simple to operate and highly efficient, meeting the application requirements of real-world scenarios. Furthermore, since its training process is supported by a reliable loss function and a large pre-prepared transmission tower structure dataset, the overall training process is not difficult. Overall, this model has great engineering significance.

[0097] For ease of understanding, this application provides an experimental case of a truss structure transmission tower generation method based on a generative adversarial network. This process uses an NVIDIA GeForce RTX3090 GPU for training. The dataset is a typical truss transmission tower graph structure dataset that has been constructed. The data includes a conditional control vector file, a node feature file, and an adjacency relationship file. The network loss is calculated using a cross-entropy loss function, a learning rate of 0.003, and a maximum training round of 200. The discriminant values of different batches and samples can be found in [1]. Figure 6The three curves represent the discriminant values after pre-training sampling, which are used to measure the discriminant performance of the variational autoencoder during pre-training. Figure 6 It can be seen that during the pre-training process of the discriminator D, through the cross entropy loss calculation of the down-sampling value and the target value, and the gradient back propagation, it can be seen that the discriminator's discrimination value for real samples gradually increases, and the discrimination value for the variational autoencoder reconstruction amount and the noise sample gradually decreases, indicating that the discriminator's ability to distinguish between real and false samples continues to improve with the progress of training. At 200 epochs, it already has a strong judgment ability, and the discrimination value K after downsampling of true and false samples is Dis It has reached 0.9 and 0.1. As the discriminator's capabilities improve, during the formal training period, the discriminator and the generator are trained alternately. The generator gradually has the ability to generate graph structures that the discriminator cannot distinguish between true and false, reaching a Nash equilibrium state.

[0098] After the discriminator pre-training is completed, the decoder / G generator is added to the joint training process. The total loss of the generator G and the discriminator D can be found in Figure 7 The two loss curves represent the overall loss of the generator and discriminator, respectively. As the number of epochs increases, the generator and discriminator losses gradually approach the intermediate value of 0 (after regularization), indicating that the generator and discriminator have reached Nash equilibrium. This indicates that the truss solutions generated by the generative adversarial network are indistinguishable from real samples in the discriminator's view.

[0099] See also Figure 8 , the prior loss L during the overall training process prior and reconstruction loss L rec It approaches zero and maintains a low level for a long time. It can be seen that the Encoder-Decoder structure has a good effect on the reconstruction of the truss structure. The model converges very quickly. During the reconstruction process, the generated truss model embedding matrix has a high similarity with the original real sample.

[0100] After 200 epochs of training, the truss transmission tower structure generated based on the model can be obtained according to different boundary conditions and control vectors. For details, please refer to Figure 9 After 200 epochs of training, the generator has the ability to conditionally generate specific truss structures, and can judge them from the number of basic tower segment units. The network's generation results are controlled by the conditional control vector. As the label value increases, the vertical height, bottleneck width, and root opening width of the generated tower segment continue to increase, and the truss complexity also increases. It can be seen that the constructed graph generation adversarial network framework can be used for the overall intelligent design of the truss structure inside the transmission tower, and can also be connected with the material selection and tower segment merging work in the subsequent design process.

[0101] The embodiment of the present application provides a truss structure transmission tower generation method based on a generative adversarial network. The method generates a transmission tower graph structure dataset based on transmission tower structural feature data, and performs generative adversarial training on the constructed initial TGN model based on this dataset. This can optimize the model's truss structure transmission tower graph generation performance, thereby obtaining a target TGN model. The generative adversarial model can quickly generate a truss structure transmission tower graph with good results. This process constructs a targeted training dataset for model training, making model training more feasible. In addition, the constructed generative adversarial model structure is relatively simple, and data processing is not complicated, which can improve the model's graph structure generation efficiency to a certain extent. Therefore, the embodiment of the present application can solve the technical problem that the existing truss structure transmission tower generation technology has defects such as complex model operation, high training difficulty, and low efficiency, which leads to poor transmission tower design results.

[0102] For easier understanding, see Figure 2 The present application provides an embodiment of a truss structure transmission tower generation device based on a generative adversarial network, including:

[0103] The data generating unit 201 is configured to generate a transmission tower graph structure based on the vertical height, the bottle mouth width, and the root opening width of the transmission tower to obtain a transmission tower graph structure dataset;

[0104] A model training unit 202 is configured to perform generative adversarial training of a truss structure transmission tower on an initial TGN model using a transmission tower graph structure dataset to obtain a target TGN model, wherein the initial TGN model is constrained by a conditional control vector and includes a GVAE generator and a GCN discriminator;

[0105] The target generation unit 203 is configured to generate a truss structure transmission tower in a target scenario using a target TGN model to obtain a generated transmission tower.

[0106] Furthermore, the data generating unit 201 is specifically configured to:

[0107] The transmission towers are sorted and screened based on the vertical height, bottle mouth width and root opening width of the transmission tower sections, and multiple transmission tower foundation structural elements are generated;

[0108] Perform graph structure splicing operations based on the basic structural elements of the transmission tower to obtain graph structure data;

[0109] The graph structure data is converted into node feature files and adjacency relationship files to obtain the transmission tower graph structure dataset.

[0110] Furthermore, it also includes:

[0111] The generator construction unit 204 is used to construct an initial generator by combining the introduced latent variables with the GCN encoder and the FNN decoder, and configure the generation constraints of the conditional control vector for the initial generator to obtain a GVAE generator;

[0112] The discriminator construction unit 205 is used to construct a GCN discriminator through the GCN network and the FNN network framework;

[0113] The model generation unit 206 is used to obtain an initial TGN model by combining the GVAE generator and the GCN discriminator according to the comprehensive loss function.

[0114] Furthermore, the comprehensive loss function of the initial TGN model is expressed as:

[0115]

[0116] in, is the prior loss, is the reconstruction loss, To generate the loss, 、 is a variable weighting parameter.

[0117] The present application also provides a truss structure transmission tower generation device based on a generative adversarial network, the device including a processor and a memory;

[0118] The memory is used to store program codes and transmit the program codes to the processor;

[0119] The processor is configured to execute the method for generating a truss structure transmission tower based on a generative adversarial network in the above method embodiment according to the instructions in the program code.

[0120] The present application also provides a computer-readable storage medium, which is used to store program code, and the program code is used to execute the truss structure transmission tower generation method based on the generative adversarial network in the above method embodiment.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0122] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0123] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0124] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the method described in each embodiment of the present application through a computer device (which can be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (full name: Read-Only Memory, English abbreviation: ROM), random access memory (full name: Random Access Memory, English abbreviation: RAM), disk or optical disk, and other media that can store program code.

[0125] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A truss structure transmission tower generation method based on generative adversarial networks, characterized in that: include: The transmission tower graph structure generation operation is performed based on the vertical height, bottle mouth width, and root opening width of the transmission tower to obtain the transmission tower graph structure dataset. The specific process is as follows: The transmission towers are sorted and screened based on the vertical height, bottle mouth width and root opening width of the transmission tower sections, and multiple transmission tower foundation structural elements are generated; Performing a graph structure splicing operation according to the transmission tower basic structure element to obtain graph structure data; Converting the graph structure data into a node feature file and an adjacency relationship file to obtain a transmission tower graph structure dataset; Among them, the node feature is the node coordinate [Xn, Yn], which is embedded and fused with the adjacency matrix as the node embedding feature. The tower segment contour feature is used as a condition to control the generated graph structure at the graph level. In the graph structure dataset, the node adjacency matrix is in the form of paired vector groups, the maximum number of graph edges is 60, the number of nodes is 32, all graphs are symmetrical structures, and the node feature file contains the positional relationship of the nodes, and the adjacency relationship file contains the topological characteristics of the graph; By introducing latent variables, combining the GCN encoder and FNN decoder to construct an initial generator, and configuring the generation constraints of the conditional control vector for the initial generator, we obtain the GVAE generator. The adopted graph structure generator can generate a graph of any size m by sampling m times from the shared distribution of the latent space vector. The conditional control vector considers three factors, namely the vertical height of the tower segment, the bottle mouth width, and the root opening width. Construct a GCN discriminator through the GCN network and FNN network framework; An initial TGN model is obtained by combining the GVAE generator and the GCN discriminator according to a comprehensive loss function; Using the transmission tower graph structure dataset to perform generative adversarial training on the initial TGN model for truss structure transmission towers, to obtain a target TGN model, wherein the initial TGN model is constrained by a conditional control vector and includes a GVAE generator and a GCN discriminator; The target TGN model is used to generate a truss structure transmission tower in a target scenario to obtain a generated transmission tower.

2. The method for generating a truss structure transmission tower based on a generative adversarial network according to claim 1, characterized in that: The comprehensive loss function of the initial TGN model is expressed as: ； in, is the prior loss, is the reconstruction loss, To generate the loss, 、 is a variable weighting parameter.

3. A truss structure transmission tower generation device based on a generative adversarial network, characterized in that: include: A data generation unit is configured to generate a transmission tower graph structure based on the vertical height, bottle mouth width, and root opening width of the transmission tower to obtain a transmission tower graph structure dataset. The data generation unit is specifically configured to: The transmission towers are sorted and screened based on the vertical height, bottle mouth width and root opening width of the transmission tower sections, and multiple transmission tower foundation structural elements are generated; Performing a graph structure splicing operation according to the transmission tower basic structure element to obtain graph structure data; Converting the graph structure data into a node feature file and an adjacency relationship file to obtain a transmission tower graph structure dataset; Among them, the node feature is the node coordinate [Xn, Yn], which is embedded and fused with the adjacency matrix as the node embedding feature. The tower segment contour feature is used as a condition to control the generated graph structure at the graph level. In the graph structure dataset, the node adjacency matrix is in the form of paired vector groups, the maximum number of graph edges is 60, the number of nodes is 32, all graphs are symmetrical structures, and the node feature file contains the positional relationship of the nodes, and the adjacency relationship file contains the topological characteristics of the graph; The generator construction unit is used to construct an initial generator by combining the introduced latent variables with the GCN encoder and the FNN decoder, and configure the generation constraints of the conditional control vector for the initial generator to obtain the GVAE generator. The adopted graph structure generator can generate a graph of any size m by sampling m times from the shared distribution of the latent space vector. The conditional control vector considers three factors, namely the vertical height of the tower segment, the bottle mouth width, and the root opening width; The discriminator construction unit is used to build a GCN discriminator through the GCN network and FNN network framework; A model generation unit, configured to obtain an initial TGN model by combining the GVAE generator and the GCN discriminator according to a comprehensive loss function; A model training unit is configured to perform generative adversarial training of a truss structure transmission tower on the initial TGN model using the transmission tower graph structure dataset to obtain a target TGN model, wherein the initial TGN model is constrained by a conditional control vector and includes a GVAE generator and a GCN discriminator; The target generation unit is used to generate a truss structure transmission tower in a target scenario by using the target TGN model to obtain a generated transmission tower.

4. The truss structure transmission tower generation device based on generative adversarial network according to claim 3, characterized in that: The comprehensive loss function of the initial TGN model is expressed as: ； in, is the prior loss, is the reconstruction loss, To generate the loss, 、 is a variable weighting parameter.

5. A truss structure transmission tower generation device based on a generative adversarial network, characterized in that: The device includes a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the truss structure transmission tower generation method based on a generative adversarial network according to any one of claims 1-2 according to the instructions in the program code.

6. A computer-readable storage medium, characterized in that The computer-readable storage medium is used to store program code, and the program code is used to execute the method for generating a truss structure transmission tower based on a generative adversarial network according to any one of claims 1 to 2.

Images