Generating Method and System Based on the Topology Structure of Variational Autoencoders

Abstract

A method and system for generating a topological structure based on a variational autoencoder. The present invention relates to the technical field of topological structure generation. The method includes inputting an initial topological structure of a nuclear reactor into an encoder structure of a variational autoencoder to obtain a first latent variable; inputting a target performance of the nuclear reactor into a mapping model to obtain a second latent variable; fusing the first latent variable and the second latent variable to obtain a fused latent variable; performing a random operation on the fused latent variable to obtain a plurality of random latent variables; and inputting the fused latent variable and the plurality of random latent variables into a decoder structure of the variational autoencoder to obtain a target topological structure of the nuclear reactor.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear reactor topological structures, and particularly to a method and system for generating a topological structure based on a variational autoencoder. Background Art

[0002] As an important energy facility, a nuclear reactor has high design complexity and strict safety requirements. The topological structure of a nuclear reactor directly affects the performance and efficiency of the reactor. Therefore, optimizing the topological structure is the key to improving the performance of a nuclear reactor. Traditional nuclear reactor design methods rely on expert experience and trial-and-error methods, which are time-consuming and costly. In addition, traditional methods may have limitations in exploring the design space and it is difficult to achieve an optimal solution.

[0003] Therefore, how to automatically generate a topological structure that meets specific performance requirements and improve the design efficiency is an urgent problem to be solved currently. Summary of the Invention

[0004] The main technical problem to be solved by the present invention is how to automatically generate a topological structure that meets specific performance requirements and improve the design efficiency.

[0005] According to a first aspect, the present invention provides a method for generating a topological structure based on a variational autoencoder, including: inputting an initial topological structure of a nuclear reactor into an encoder structure of the variational autoencoder to obtain a first latent variable; inputting the target performance of the nuclear reactor into a mapping model to obtain a second latent variable; fusing the first latent variable and the second latent variable to obtain a fused latent variable; performing random operations on the fused latent variable to obtain a plurality of random latent variables; and inputting the fused latent variable and the plurality of random latent variables into a decoder structure of the variational autoencoder to obtain a target topological structure of the nuclear reactor.

[0006] Furthermore, the random operations include convolution, crossover, and mutation.

[0007] Furthermore, the step of inputting an initial topological structure of a nuclear reactor into an encoder structure of the variational autoencoder to obtain a first latent variable includes: representing the initial topological structure of the nuclear reactor as pixels of 80*80*80 and inputting them into the encoder structure of the variational autoencoder to obtain a first latent variable.

[0008] Further, the encoder includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, and a fully connected layer. The convolutional kernel of the first convolutional layer is 2, 2, 80, 48, the stride is 2, and a feature map group of 40*40*48 is obtained as the output. Then, it is activated by the RELU function. The first pooling layer uses a maximum pooling operation with a size of 2*2 and a stride of 2 to obtain a feature map group of 20*20*48. The second convolutional layer uses a convolutional kernel of 2, 2, 48, 128, the stride is 1, and zero-padding is used to obtain a feature map group of 19*19*128. Then, it is activated by the RELU function. The second pooling layer uses a maximum pooling operation with a size of 3*3 and a stride of 2 to obtain a feature map group of 9*9*128. The third convolutional layer uses a convolutional kernel of 2, 2, 128, 128, the stride is 1, to obtain a feature map group of 8*8*128. Then, it is activated by the RELU function. The third pooling layer uses a maximum pooling operation with a size of 2*2 and a stride of 3 to obtain a feature map group of 2*2*128. The fully connected layer outputs a 1*512-dimensional tensor.

[0009] Further, the step of inputting the fused latent variable and multiple random latent variables into the decoder structure of the variational autoencoder to obtain the target topology structure of the nuclear reactor includes: inputting the fused latent variable and multiple random latent variables into the decoder structure of the variational autoencoder to obtain the fused topology structure and multiple random topology structures of the nuclear reactor; judging the target performance of the nuclear reactor, the fused topology structure of the nuclear reactor, and multiple random topology structures based on the convolutional neural network model to obtain the degree of difference between the fused topology structure of the nuclear reactor and the target performance of the nuclear reactor, and the degree of difference between each random topology structure and the target performance of the nuclear reactor; and taking the topology structure with the smallest degree of difference in the fused topology structure and each random topology structure as the target topology structure.

[0010] Further, the mapping model is a bidirectional recurrent neural network.

[0011] According to a second aspect, the present invention provides a generation system for a topology structure based on a variational autoencoder, including: an acquisition module, configured to input the initial topology structure of the nuclear reactor into the encoder structure of the variational autoencoder to obtain a first latent variable;

[0012] a mapping module, configured to input the target performance of the nuclear reactor into the mapping model to obtain a second latent variable;

[0013] a fusion module, configured to fuse the first latent variable and the second latent variable to obtain a fused latent variable;

[0014] a random latent variable determination module, configured to perform a random operation on the fused latent variable to obtain multiple random latent variables;

[0015] The target topology determination module is used to input the fused latent variable and multiple random latent variables into the decoder structure of the variational autoencoder to obtain the target topology of the nuclear reactor.

[0016] Furthermore, the random operations include convolution, crossover, and mutation.

[0017] Furthermore, the encoder includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, and a fully connected layer. The convolutional kernel of the first convolutional layer is 2, 2, 80, 48, the stride is 2, and a feature map group of 40*40*48 is obtained as the output. Then it is activated by the RELU function. The first pooling layer uses a maximum pooling operation with a size of 2*2 and a stride of 2 to obtain a feature map group of 20*20*48. The second convolutional layer uses a convolutional kernel of 2, 2, 48, 128, the stride is 1, and zero-padding is used to obtain a feature map group of 19*19*128. Then it is activated by the RELU function. The second pooling layer uses a maximum pooling operation with a size of 3*3 and a stride of 2 to obtain a feature map group of 9*9*128. The third convolutional layer uses a convolutional kernel of 2, 2, 128, 128, the stride is 1, to obtain a feature map group of 8*8*128. Then it is activated by the RELU function. The third pooling layer uses a maximum pooling operation with a size of 2*2 and a stride of 3 to obtain a feature map group of 2*2*128. The fully connected layer outputs a 1*512-dimensional tensor.

[0018] Furthermore, the target topology determination is also used for:

[0019] Input the fused latent variable and multiple random latent variables into the decoder structure of the variational autoencoder to obtain the fused topology of the nuclear reactor and multiple random topologies;

[0020] Based on the convolutional neural network model, judge the target performance of the nuclear reactor, the fused topology of the nuclear reactor, and multiple random topologies to obtain the degree of difference between the fused topology of the nuclear reactor and the target performance of the nuclear reactor, and the degree of difference between each random topology and the target performance of the nuclear reactor;

[0021] Take the topology with the smallest degree of difference among the fused topology and each random topology as the target topology.

[0022] A method and system for generating a topological structure based on a variational autoencoder provided by the present invention. The method includes inputting an initial topological structure of a nuclear reactor into an encoder structure of the variational autoencoder to obtain a first latent variable; inputting a target performance of the nuclear reactor into a mapping model to obtain a second latent variable; fusing the first latent variable and the second latent variable to obtain a fused latent variable; performing a random operation on the fused latent variable to obtain a plurality of random latent variables; and inputting the fused latent variable and the plurality of random latent variables into a decoder structure of the variational autoencoder to obtain a target topological structure of the nuclear reactor. This method can automatically generate a topological structure that meets specific performance requirements and improve the design efficiency. Description of the Drawings

[0023] Figure 1 It is a schematic flowchart of a method for generating a topological structure based on a variational autoencoder provided by an embodiment of the present invention;

[0024] Figure 2 It is a schematic diagram of an encoder provided by an embodiment of the present invention;

[0025] Figure 3 It is a schematic flowchart of a method for generating a target topological structure provided by an embodiment of the present invention;

[0026] Figure 4 It is a schematic diagram of a system for generating a topological structure based on a variational autoencoder provided by an embodiment of the present invention. Detailed Embodiments

[0027] In an embodiment of the present invention, there is provided a method for generating a topological structure based on a variational autoencoder as shown in Figure 1 The method for generating a topological structure based on a variational autoencoder includes steps S1 to S5:

[0028] Step S1: Input an initial topological structure of a nuclear reactor into an encoder structure of the variational autoencoder to obtain a first latent variable.

[0029] The initial topological structure of the nuclear reactor refers to the initial layout or structure of the nuclear reactor system. The encoder structure is a part of the VAE and is used to encode the input data into a representation in the latent space. Represent the layout of the nuclear reactor as a pixelated three-dimensional image, and then convert it into a first latent variable through the encoder.

[0030] In some embodiments, the initial topological structure of the nuclear reactor can be represented as an 80*80*80 pixel input into the encoder structure of the variational autoencoder to obtain a first latent variable.

[0031] Figure 2A schematic diagram of an encoder provided by an embodiment of the present invention. In some embodiments, the encoder includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, and a fully connected layer. The convolutional kernel of the first convolutional layer is 2, 2, 80, 48, the stride is 2, and a feature map group of 40*40*48 is obtained as the output, and then activated by the RELU function. The first pooling layer uses a maximum pooling operation with a size of 2*2 and a stride of 2 to obtain a feature map group of 20*20*48. The second convolutional layer uses a convolutional kernel of 2, 2, 48, 128, the stride is 1, and zero-padding is used to obtain a feature map group of 19*19*128, and then activated by the RELU function. The second pooling layer uses a maximum pooling operation with a size of 3*3 and a stride of 2 to obtain a feature map group of 9*9*128. The third convolutional layer uses a convolutional kernel of 2, 2, 128, 128, the stride is 1, to obtain a feature map group of 8*8*128, and then activated by the RELU function. The third pooling layer uses a maximum pooling operation with a size of 2*2 and a stride of 3 to obtain a feature map group of 2*2*128. The fully connected layer outputs a 1*512-dimensional tensor.

[0032] Step S2: Input the target performance of the nuclear reactor into the mapping model to obtain a second latent variable.

[0033] The mapping model is a bidirectional recurrent neural network. The mapping model includes a forward RNN hidden layer and a backward RNN hidden layer. The two hidden layers are combined to extract features of the target performance feature and the second latent variable feature. The hidden state of the forward RNN hidden layer and the hidden state of the backward RNN hidden layer are updated as follows:

[0034] represents the state at time t of the forward hidden layer, represents the state at time t-1 of the forward hidden layer, 、 、 、 、 、 represents the bias, represents the input.

[0035] The forward hidden state and the backward hidden state are connected to obtain the hidden state which is passed to the next bidirectional layer.

[0036] The target performance refers to the expected behavior or performance metrics of a nuclear reactor system under specific conditions. A Bi-directional Recurrent Neural Network (Bi-RNN) is used as a mapping model, and its forward and backward hidden layers are utilized to extract the features of the target performance. Through the mapping model, the complex features of the target performance can be extracted and transformed into a second latent variable, thus better correlating with the topological structure.

[0037] Step S3: Fuse the first latent variable and the second latent variable to obtain a fused latent variable.

[0038] The fused latent variable can be a mixture of Gaussian models. The fused latent variable is a new representation formed by combining the first latent variable and the second latent variable, and is usually used to comprehensively consider information from different sources. Using a mixture of Gaussian models to fuse the first latent variable and the second latent variable can obtain a richer and more complex latent variable representation.

[0039] Step S4: Perform random operations on the fused latent variable to obtain multiple random latent variables.

[0040] The random operations are a series of random transformations or operations on the fused latent variable to increase its diversity and richness. The random operations include convolution, crossover, and mutation. Increasing randomness can introduce more variations and diversities, which helps to generate more diverse and realistic topological structures.

[0041] Step S5: Input the fused latent variable and the multiple random latent variables into the decoder structure of the variational autoencoder to obtain the target topological structure of the nuclear reactor.

[0042] In some embodiments, the variational autoencoder is obtained through training. The variational autoencoder can be trained by maximizing the loss function to minimize the KL divergence.

[0043] where represents the loss function, where represents the KL divergence, is the reconstruction term, representing the expected value of the log-likelihood of the reconstructed data x under the approximate posterior distribution of the latent distribution z. Where is the approximate distribution of the latent variable z after observing the data x, is the regularization term, which prevents overfitting and improves the generalization ability of the model, representing the KL divergence between the approximate posterior distribution and the prior distribution of the latent variable z. The KL divergence measures the difference between two probability distributions.

[0044] The decoder structure is part of the VAE and is used to decode the representation in the latent space into the original data. The fused latent variable and multiple random latent variables are input into the decoder to generate a topology corresponding to the target performance.

[0045] In some embodiments, step S5 can be implemented by Figure 3 the method shown as Figure 3 is a schematic flowchart of a method for generating a target topology provided by an embodiment of the present invention. The method for generating the target topology includes steps S31 to S33:

[0046] Step S31: Input the fused latent variable and multiple random latent variables into the decoder structure of the variational autoencoder to obtain the fused topology of the nuclear reactor and multiple random topologies.

[0047] Taking the fused latent variable and multiple random latent variables as inputs, multiple possible topologies of the nuclear reactor are generated through the decoder structure. Each latent variable corresponds to a different topology. By generating multiple random topologies, different possibilities can be explored, diversity can be increased, and more alternative solutions can be obtained to meet the target performance.

[0048] Step S32: Based on the convolutional neural network model, judge the target performance of the nuclear reactor, the fused topology of the nuclear reactor, and multiple random topologies to obtain the degree of difference between the fused topology of the nuclear reactor and the target performance of the nuclear reactor, and the degree of difference between each random topology and the target performance of the nuclear reactor.

[0049] Use the convolutional neural network (CNN) model to judge the target performance, the fused topology, and multiple random topologies to obtain the degree of difference between them and the target performance. Using the trained CNN model, each topology is evaluated to obtain the degree of difference between it and the target performance, so as to measure the degree to which it meets the target performance.

[0050] Step S33: Select the topology with the smallest degree of difference among the fused topology and each random topology as the target topology.

[0051] From the results evaluated in step S32, select the topology with the smallest degree of difference from the target performance among the fused topology and multiple random topologies as the final target topology.

[0052] Based on the same inventive concept, Figure 4 is a schematic diagram of a system for generating a topology based on a variational autoencoder provided by an embodiment of the present invention. The system for generating a topology based on a variational autoencoder includes:

[0053] An acquisition module 41, configured to input an initial topology of a nuclear reactor into an encoder structure of a variational autoencoder to obtain a first latent variable;

[0054] A mapping module 42, configured to input a target performance of a nuclear reactor into a mapping model to obtain a second latent variable;

[0055] A fusion module 43, configured to fuse the first latent variable and the second latent variable to obtain a fused latent variable;

[0056] A random latent variable determination module 44, configured to perform a random operation on the fused latent variable to obtain a plurality of random latent variables;

[0057] A target topology determination module 45, configured to input the fused latent variable and the plurality of random latent variables into a decoder structure of a variational autoencoder to obtain a target topology of the nuclear reactor.

Claims

1. A method for generating a topological structure based on a variational autoencoder, characterized in that Including: Input the initial topology of the nuclear reactor into the encoder structure of the variational autoencoder to obtain the first latent variable; Input the target performance of the nuclear reactor into the mapping model to obtain the second latent variable; Fuse the first latent variable and the second latent variable to obtain the fused latent variable; Perform random operations on the fused latent variable to obtain multiple random latent variables; Input the fused latent variable and multiple random latent variables into the decoder structure of the variational autoencoder to obtain the target topology of the nuclear reactor. The step of inputting the fused latent variable and multiple random latent variables into the decoder structure of the variational autoencoder to obtain the target topology of the nuclear reactor includes: Input the fused latent variable and multiple random latent variables into the decoder structure of the variational autoencoder to obtain the fused topology of the nuclear reactor and multiple random topologies; Based on the convolutional neural network model, judge the target performance of the nuclear reactor, the fused topology of the nuclear reactor, and multiple random topologies to obtain the degree of difference between the fused topology of the nuclear reactor and the target performance of the nuclear reactor, and the degree of difference between each random topology and the target performance of the nuclear reactor; Take the topology with the smallest degree of difference among the fused topology and each random topology as the target topology.

2. The generation method of the topological structure based on the variational autoencoder according to claim 1, characterized in that, The random operations include convolution, crossover, and mutation.

3. The generation method of the topological structure based on the variational autoencoder according to claim 1, characterized in that The step of inputting the initial topology of the nuclear reactor into the encoder structure of the variational autoencoder to obtain the first latent variable includes: Represent the initial topology of the nuclear reactor as pixels of 80*80*80 and input them into the encoder structure of the variational autoencoder to obtain the first latent variable.

4. The generation method of the topological structure based on the variational autoencoder according to claim 3, wherein, The encoder includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, and a fully connected layer. The convolutional kernel of the first convolutional layer is 2, 2, 80, 48, the stride is 2, and a feature map group of 40*40*48 is output, and then activated by the RELU function. The first pooling layer uses a maximum pooling operation with a size of 2*2 and a stride of 2 to obtain a feature map group of 20*20*48. The second convolutional layer uses a convolutional kernel of 2, 2, 48, 128, the stride is 1, and zero-padding is used to obtain a feature map group of 19*19*128, and then activated by the RELU function. The second pooling layer uses a maximum pooling operation with a size of 3*3 and a stride of 2 to obtain a feature map group of 9*9*128. The third convolutional layer uses a convolutional kernel of 2, 2, 128, 128, the stride is 1, to obtain a feature map group of 8*8*128, and then activated by the RELU function. The third pooling layer uses a maximum pooling operation with a size of 2*2 and a stride of 3 to obtain a feature map group of 2*2*128. The fully connected layer outputs a 1*512-dimensional tensor.

5. The generation method of the topological structure based on the variational autoencoder according to claim 1, characterized in that, The mapping model is a bidirectional recurrent neural network.

6. A generation system based on the topological structure of a variational autoencoder, characterized in that, Including: An acquisition module for inputting the initial topology of the nuclear reactor into the encoder structure of the variational autoencoder to obtain the first latent variable; A mapping module for inputting the target performance of the nuclear reactor into the mapping model to obtain the second latent variable; A fusion module for fusing the first latent variable and the second latent variable to obtain the fused latent variable; A random latent variable determination module for performing random operations on the fused latent variable to obtain multiple random latent variables; A target topology determination module for inputting the fused latent variable and multiple random latent variables into the decoder structure of a variational autoencoder to obtain the target topology of the nuclear reactor. The target topology determination is further used for: Inputting the fused latent variable and multiple random latent variables into the decoder structure of a variational autoencoder to obtain the fused topology of the nuclear reactor and multiple random topologies; Based on a convolutional neural network model, judging the target performance of the nuclear reactor, the fused topology of the nuclear reactor, and multiple random topologies to obtain the degree of difference between the fused topology of the nuclear reactor and the target performance of the nuclear reactor, and the degree of difference between each random topology and the target performance of the nuclear reactor; Taking the topology with the smallest degree of difference among the fused topology and each random topology as the target topology.

7. The generation system based on the topological structure of the variational autoencoder according to claim 6, characterized in that, The random operations include convolution, crossover, and mutation.

8. The generation system based on the topological structure of the variational autoencoder according to claim 6, wherein, The encoder includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, and a fully connected layer. The convolutional kernel of the first convolutional layer is 2, 2, 80, 48, the stride is 2, and a feature map group of 40*40*48 is output, which is then activated by the RELU function. The first pooling layer uses a maximum pooling operation with a size of 2*2 and a stride of 2 to obtain a feature map group of 20*20*48. The second convolutional layer uses a convolutional kernel of 2, 2, 48, 128, the stride is 1, and zero-padding is used to obtain a feature map group of 19*19*128, which is then activated by the RELU function. The second pooling layer uses a maximum pooling operation with a size of 3*3 and a stride of 2 to obtain a feature map group of 9*9*128. The third convolutional layer uses a convolutional kernel of 2, 2, 128, 128, the stride is 1, to obtain a feature map group of 8*8*128, which is then activated by the RELU function. The third pooling layer uses a maximum pooling operation with a size of 2*2 and a stride of 3 to obtain a feature map group of 2*2*128. The fully connected layer outputs a 1*512-dimensional tensor.

Images