KR102576560B1 - Method and system for optimal design based on deep learning - Google Patents

Abstract

Disclosed is a deep learning-based design optimization method and system that calculates the optimal design using only design target data. According to the present invention, an optimal design can be calculated using only data without expert knowledge of the product or system that is the subject of design optimization. Additionally, the deep learning-based design optimization method of the present invention is widely applicable to all fields.

Description

Deep learning based design optimization method and system {Method and system for optimal design based on deep learning}

The present invention relates to a design optimization method and system using various deep learning techniques and genetic algorithms to automate optimal design.

Optimal design can be defined as a design that satisfies engineering needs and consumer requirements (objective function) as much as possible within constraints, and consists of (a) a formalization process and (b) an optimization process that mathematically converts the design requirements. do. The formalization process generally goes through the process of defining design variables, setting objective functions, and deriving constraint conditions, of which configuring design variables and objective functions requires a significant level of expertise. For example, if you want to design a ship shape that minimizes resistance due to wind, waves, etc., domain knowledge about fluid dynamics, hull structure, etc. is required. This creates a barrier to entry into engineering design, and this barrier is expected to increase over time. In addition, due to the nature of requiring specialized knowledge in the field, there is no framework that can be universally applied across multiple fields, leading to inefficiency in the overall industrial structure.

In the present invention, we propose a method to solve the problems of expertise and individuality of the above-mentioned formalization process using various deep learning techniques, which have recently become hot topics, and to automatically perform the overall design process by connecting this with a genetic algorithm-based optimization process. I want to do it.

The purpose of the present invention is to provide a method and system that can optimize a design using only data without expert knowledge of the product or system that is the subject of design optimization.

A deep learning-based design optimization system according to an example of the present invention includes an actual data acquisition unit that acquires actual data about an object to generate a design model; a design variable extraction unit that receives the actual data acquired by the actual data acquisition unit and extracts design variables using the actual data; an objective function constructor that receives the design variables extracted by the design variable extraction unit and constructs an objective function based on the design variables; a learning model generator that receives the design variables extracted by the design variable extraction unit and the objective function constructed by the objective function constructor, and generates a learning model using the objective function and the design variables; an optimal design variable extraction unit that receives the design variables extracted by the design variable extraction unit and extracts optimal design variables using the design variables; And generating an optimal design model that receives the learning model generated by the learning model generation unit and the optimal design variables extracted by the optimal design variable extraction unit, and applies the optimal design variables to the learning model to generate an optimal design model of the object. may include;

The design variable extraction unit extracts the design variables using an autoencoder (AE, Autoencodr), and the objective function constructor configures the objective function using a feedforward neural network (FNN), The optimal design variable extraction unit extracts the optimal design variables using a genetic algorithm (GA), and the learning model generation unit generates the learning model using a conditional adversarial neural network (CGAN, Conditional GAN). , the optimal design model generator may generate an optimal design model by applying the optimal design variables to the learning model generated by the learning model generator.

It further includes a virtual data generator that receives the data acquired by the real data acquisition unit and generates virtual data by augmenting the real data, wherein the design variable extractor is configured to combine the real data acquired by the real data acquisition unit and the Virtual data generated by the virtual data acquisition unit may be received, and design variables may be extracted using the real data and the virtual data.

The virtual data generator may generate the virtual data using a generative adversarial network (GAN).

It further includes a pre-learned design model storage unit that stores a pre-learned design model for another object, wherein the design model generator transfer-learns the pre-learned design model received from the pre-learned design model storage unit to The optimal design model of the target can be created.

The design model generator receives the learned weights when generating the pre-learned optimal design model from the pre-learned design model storage unit, and applies some or all of the weights to learning the optimal design model of the object to be newly created. , an optimal design model of the object can be created.

A deep learning-based design optimization method according to an example of the present invention includes the steps of acquiring actual data about an object to generate a design model; extracting design variables using the actual data; Constructing an objective function based on the design variables; Generating a learning model using the objective function and the design variables; extracting optimal design variables using the design variables; Applying the optimal design variables to the learning model to generate an optimal design model of the object.

In the step of extracting the design variables, the design variables are extracted using an autoencoder (AE, Autoencodr), and in the step of configuring the objective function, the objective function is extracted using a feedforward neural network (FNN). The step of configuring the function and extracting the optimal design variables includes extracting the optimal design variables using a genetic algorithm (GA), and the step of generating the learning model includes using a conditional adversarial generative neural network (CGAN). The learning model is generated using Conditional GAN, and in the step of generating the optimal design model, the optimal design model can be generated by applying the optimal design variable to the learning model generated by the learning model generator.

Before the step of extracting the design variable, it further includes a virtual data generator that generates virtual data by augmenting the real data, wherein the step of extracting the design variable includes designing a design using the real data and the virtual data. Variables can be extracted.

In the step of generating the virtual data, the virtual data may be generated using a generative adversarial network (GAN).

Before the step of generating the optimal design model, it further includes the step of storing a previously learned design model for another object, wherein the step of generating the optimal design model includes transfer learning the pre-learned design model. An optimal design model of the object can be created.

In the step of generating the optimal design model, the learned weights are received when generating the pre-learned optimal design model, and some or all of the weights are applied to learn the optimal design model of the object to be newly created, An optimal design model can be created.

According to the present invention, a design model can be optimized using only data on the design target without expert knowledge of the product or system that is the target of design optimization.

Additionally, the deep learning-based design optimization method of the present invention is widely applicable to all fields.

1 is a flowchart of a deep learning-based design optimization method according to an example of the present invention.
Figure 2 is a configuration diagram of a deep learning-based design optimization system according to an example of the present invention.
Figure 3 is a conceptual diagram showing the structure of AE.
Figure 4 is a conceptual diagram of design variable extraction using AE.
Figure 5 is a conceptual diagram showing the structure of GAN.
Figure 6 is a conceptual diagram of data augmentation using GAN.
Figure 7 is a conceptual diagram of extracting design variables using real data and virtual data.
Figures 8 and 9 are conceptual diagrams showing the construction of the objective function using FNN.
Figure 10 is a conceptual diagram showing the structure of CGAN.
Figure 11 is a conceptual diagram for deriving optimal design variables using GA.
Figure 12 is a conceptual diagram showing the creation of an optimal design model using optimal design variables.
Figure 13 is a conceptual diagram of transfer learning.
Figure 14 is a flow chart showing each specific algorithm (AE, GAN, FNN, GA, CGAN) in the optimal design model process.
Figure 15 shows the design variable extraction step using AE in the method of generating an optimal design model for a vehicle sound system.
Figure 16 shows configuring an objective function in a method for optimizing the design of a vehicle sound system.
Figure 17 shows the creation of a color map corresponding to a design variable.
Figure 18 shows the steps for finding the optimal value of the design variable using GA.
Figure 19 shows steps for calculating the optimal design of vehicle acoustics using optimal values of design variables.

Hereinafter, the present invention will be described in detail with reference to the attached drawings.

FIG. 1 is a flowchart of a deep learning-based design optimization method according to an example of the present invention. Referring to FIG. 1, the method of the present invention includes the steps of acquiring actual data for an object to generate a design model (S100); Extracting design variables using actual data (S200); Constructing an objective function based on design variables (S300); Creating a learning model using the objective function and design variables (S400); Extracting optimal design variables using design variables (S500); Applying optimal design variables to the learning model to generate an optimal design model of the target (S600); generating virtual data by augmenting real data before extracting design variables (S200). It may further include (S150); and, before the step of generating an optimal design model (S600), learning about another object and storing the previously learned design model (S700).

Figure 2 is a configuration diagram of a deep learning-based design optimization system according to an example of the present invention. Referring to Figure 2, the system of the present invention includes an actual data acquisition unit ( 100); a design variable extraction unit 200 that receives the actual data acquired by the actual data acquisition unit and extracts design variables using the actual data; An objective function configuration unit 300 that receives the design variables extracted by the design variable extraction unit and constructs an objective function based on the design variables; A learning model generator 400 that receives the design variables extracted by the design variable extraction unit and the objective function constructed by the objective function configuration unit, and generates a learning model using the objective function and the design variables; An optimal design variable extraction unit 500 that receives the design variables extracted by the design variable extraction unit and extracts optimal design variables using the design variables; and an optimal design model generation unit 600 that receives the learning model generated by the learning model generation unit and the optimal design variables extracted by the optimal design variable extraction unit, and applies the optimal design variables to the learning model to generate an optimal design model of the target. ; and a virtual data generator 150 that receives the data acquired by the real data acquisition unit 100 and generates virtual data by augmenting the real data; and a design model previously learned for another object. It may further include a storage unit 700 for storing the existing learning design model.

Hereinafter, with specific reference to FIGS. 3 to 14, the deep learning-based design model optimization method and system according to the present invention will be described in each step.

(1) Obtaining actual data about the object to create a design model (S100)

As shown in FIG. 2, the actual data acquisition unit 100 may acquire actual data about an object for which a design model is to be created.

(2) Step of extracting design variables using actual data (S200)

The design variable extraction unit 200 may receive the actual data acquired by the actual data acquisition unit 100 and extract design variables using the actual data. Design variables refer to all variables that the designer must determine, such as the thickness, length, shape, and material properties of the product being designed. It can be said that the principle is to set design variables (a) independently of each other and (b) as small as possible. Defining design variables that satisfy these conditions for a given system is a complex process in itself and requires a lot of experience and knowledge from the designer. In the present invention, we intend to extract design variables from data of the design target using Autoencoder (AE), one of the deep learning algorithms.

Figure 3 is a conceptual diagram showing the structure of AE, and Figure 4 is a conceptual diagram of design variable extraction using AE.

, encoder 함수를 f, decoder 함수를 g라고 했을 때,

값을 최소화하는 latent 값

을 구하는 것이 AE의 목적이다. 즉, Z는 X의 차원 축소 결과라고 할 수 있으며, 이는 자연히 위에서 언급된 조건 (a), (b)를 만족하게 된다. 본 발명에서는 Z를 설계 변수로 선정 한다.The AE model largely consists of an encoder that compresses data and a decoder that restores the compressed data to its original state. data

, when the encoder function is called f and the decoder function is called g ,

latent value that minimizes the value

The purpose of AE is to find . In other words, Z can be said to be the result of dimensionality reduction of X , which naturally satisfies the conditions (a) and (b) mentioned above. In the present invention, Z is selected as a design variable.

Another advantage of using AE is that design variables can be easily extracted from multidimensional data, that is, images (2D) or shapes (3D). At this time is expressed as a matrix or tensor rather than a vector, and efficient operations can be performed by replacing the fully-connected layer in FIG. 3 with a convolutional layer.

(2-1) Before extracting design variables, generating virtual data by augmenting real data (S150)

The virtual data generator 150 may receive data acquired by the real data acquisition unit 100 and augment the real data to generate virtual data. The remaining processes of the design optimization method according to the present invention are all design variables derived through AE. Since it is based on , the process of defining design variables can be said to be very important. In AE, design variables are extracted using data, so not only the quality but also the quantity of data used to build the AE model is very important. This means that there may be problems in extracting sophisticated design variables if the amount of data held is insufficient, and the process of increasing the amount of data may be necessary to solve this problem. This is called data augmentation, and a generative adversarial network (GAN) is applied as a method for this. GAN is a generative model and is known to produce high-quality data compared to other algorithms.

Figure 5 is a conceptual diagram showing the structure of GAN, and Figure 6 is a conceptual diagram of data augmentation using GAN.

GAN consists of two neural networks: a discriminator (D) and a generator (G). The generator (G) creates virtual data from random noise, and the discriminator (D) receives real data and virtual data and distinguishes them. The two neural networks can be said to play a zero-sum game, as the generator tries to fool the discriminator by creating sophisticated virtual data that is indistinguishable from real data, and the discriminator is not fooled and tries to accurately distinguish between real and virtual data. . The Nash equilibrium (NE) in this situation is the point when the generator creates virtual data that is indistinguishable from real data, and the discriminator calculates a 50% probability that all data is real.

Figure 7 is a conceptual diagram of extracting design variables using real data and virtual data. The design variable extraction unit 200 may extract design variables using real data and the virtual data. In this way, by using the GAN algorithm to create virtual data with the same distribution as real data and using the virtual data to build an AE model, sophisticated design variables can be extracted even when data is insufficient.

(3) Constructing the objective function based on the design variables (S300)

를 입력받아 목적값 y를 출력하는 구조의 FNN 모델로 구성된다.The objective function constructor 300 may receive the design variables extracted by the design variable extractor 200 and construct an objective function based on the design variables. The objective function serves as a standard for determining the optimal design among various designs. For example, product price, energy consumption, maximum stress, etc. are widely used as objective functions. In the present invention, we propose to construct an objective function using a general feedforward neural network (FNN). The objective function is the design variables extracted from AE

It consists of an FNN model with a structure that receives input and outputs the target value y .

Figures 8 and 9 are conceptual diagrams showing the construction of the objective function using FNN. Referring to this, it represents the case where there is one node of the output layer and there is one objective function. In the case of a multipurpose function with two or more objective functions, it can be expressed by adjusting the number of nodes in the output layer accordingly. In addition to not requiring domain knowledge, this can also be said to be an advantage of the present invention.

(4) Step of creating a learning model based on the objective function and design variables (S400)

The learning model generator 400 receives the design variables extracted by the design variable extraction unit 200 and the objective function constructed by the objective function constructor 300, and can generate a learning model using the objective function and design variables. there is. Since the method proposed in the present invention aims at an end-to-end process, when the value of a design variable is given, a process of deriving the corresponding design is necessary. For this purpose, a deep learning technique called conditional generative adversarial network (CGAN) is used.

Figure 10 is a conceptual diagram showing the structure of CGAN. CGAN is a modified version of the GAN used in the data augmentation step (S150). In the case of a general GAN, the mode of generated data cannot be adjusted. Since the goal of this step (S500) is not to create a random design, but to create a design corresponding to the value of the given design variable, it is not appropriate to apply a general GAN. In CGAN, the mode of the generated data can be adjusted by adding data labels to the input values of the discriminator and generator. Here, the design variable becomes the label.

를 만들어내고, 판별기는 생성기가 만들어낸

를 실제 설계와 구별하는 역할을 수행한다. CGAN에서도 생성기의 입력값으로 랜덤 노이즈가 들어가기 때문에 주어진 설계 변수에 해당하는 다수개의 다양한 설계를 도출할 수 있다. 설계자는 설계 대상을 소비하는 소비자를 고려하여 다양한 설계 중 최적 설계를 선택할 수 있다.The generator in Figure 10 has design variables Receive input and create a corresponding virtual design

and the discriminator is created by the generator.

It plays a role in distinguishing the design from the actual design. In CGAN, random noise is input to the generator, so a number of different designs corresponding to given design variables can be derived. Designers can select the optimal design among various designs by considering the consumers who consume the design object.

(5) Extracting optimal design variables using design variables (S500)

The optimal design variable extraction unit 500 may receive the design variables extracted by the design variable extraction unit 200 and extract optimal design variables using the design variables. Figure 11 is a conceptual diagram for deriving optimal design variables using GA. In the present invention, a genetic algorithm (GA) is used as a method of obtaining the optimal solution (Zopt) of Z , that is, the optimal value of the design variable. GA is a representative method of evolutionary computation, and as the name suggests, it is a method of finding the optimal solution by imitating the process of organisms evolving while adapting to the environment. In GA, the solution to be found is expressed as a chromosome. Here, the chromosome becomes the design variable, and the chromosome is composed of several genes. In GA, a solution population consisting of several chromosomes is created and then the fitness of each chromosome is calculated based on the objective function and constraints. For chromosomes with insufficient fitness, chromosomes with improved fitness are obtained through gene crossover or mutation. By repeating this process, the solution group that maximizes the objective function for given constraints, that is, the optimal value of the design variable, can be obtained.

(6) Creating a design model of the target using optimal design variables (S600)

The optimal design model generation unit 600 receives the learning model generated by the learning model generation unit 400 and the optimal design variables extracted by the optimal design variable extraction unit 500, and applies the optimal design variables to the learning model, The optimal design model of the target can be created.

Figure 12 is a conceptual diagram showing the creation of an optimal design model using optimal design variables. In the present invention, after obtaining the optimal solution (Z _opt ), that is, the optimal design variable, using GA, it is learned in step (4). It can be input into the CGAN generator, which is a model, and the optimal design model G(R|Zopt) corresponding to the optimal value of the design variable can be derived.

(7) Before the step of generating the optimal design model (S600), the step of learning about other objects and saving the already learned design model (S700)

The pre-learned design model storage unit 700 may store a pre-learned design model for another object, and at this time, the design model generator 600 may store the pre-learned design model received from the pre-learned design model storage unit 700. The optimal design model of the target can be created by transfer learning.

In the present invention, transfer learning is applied when there are only a small number of correct answers to the optimal design problem to be solved (New B model), while there may be a large number of correct answers to a problem similar to the problem to be solved (Pretrained A model). This makes it possible to solve the problem.

Figure 13 is a conceptual diagram of transfer learning. Once the final optimal design model is created through the CGAN model in the optimal design model process described above, transfer learning can be applied to a similar engineering system model to derive a new optimal design model result. do. This does not solve the optimal design problem by learning the model again from scratch, but rather takes some or all of the weights from the existing learned CGAN model and trains them, thereby learning in solving the optimal design problem for other models. This has the effect of improving costs. In other words, by introducing the transfer learning concept in the existing optimal design process, it is possible to simultaneously improve the diversity of applied models and the learning cost in configuring the optimal design process.

Figure 14 is a flow chart showing each specific algorithm (AE, GAN, FNN, GA, CGAN) in the optimal design model process.

Hereinafter, with reference to FIGS. 15 to 19, the deep learning-based design optimization method according to the present invention will be described through an example in which it is applied to a vehicle sound system.

In vehicle sound systems, numerous parameters affect the generation of engine sounds. Among these, selecting highly contributing factors and designing acoustics that satisfy the desired conditions requires not only knowledge of optimal design but also specialized knowledge of internal combustion engines and acoustics. By using AE, design variables can be easily extracted using only data without such specialized knowledge. Vehicle sound is converted into image data called a colormap through Fourier transform into the frequency domain and visualization process.

Figure 15 shows the design variable extraction step (S200) using AE in the design optimization method of a vehicle sound system. Referring to Figure 15, it shows the process in which a color map image is compressed through an encoder and then restored to the original image through a decoder. At this time, the node values resulting from the encoder become design variables corresponding to the corresponding acoustic color map.

The next step after completing design variable extraction is to construct the objective function (S300). The objective function can be set by setting each design variable as the input value of the FNN and the objective value corresponding to the design variable as the output value of the FNN.

Figure 16 shows configuring an objective function in a method for optimizing the design of a vehicle sound system. For example, if the goal is to create vehicle sounds targeting men in their 30s, the preferences of men in their 30s for each sound can be obtained through a survey, then scored and set as the output value of the FNN model. . The input value of this objective function is the design variable extracted by AE for each sound (vehicle sound), and the output value is the score corresponding to the input value. This has the advantage of being able to estimate the preferences of target consumers for new vehicle acoustics without having to conduct another survey.

Afterwards, a learning model is created by training CGAN, a learning model, based on the objective function (S400). Figure 17 shows the creation of a color map corresponding to a design variable, and CGAN is trained using the extracted design variable and the corresponding vehicle sound color map (learning model). Since random noise is included in the CGAN model, it allows the creation of multiple color maps for the design variables.

Next, the optimal design variables are obtained using GA. Figure 18 is a conceptual diagram showing the steps to obtain optimal design variables using GA. In GA, several chromosomes corresponding to the design variables are initially created, and then, within given constraints, using evolutionary operations such as crossover and mutation. Find the optimal solution that maximizes the objective function.

Afterwards, the optimal solution extracted from GA can be input into the CGAN generator, which is the learning model learned above, to derive the vehicle sound most preferred by men in their 30s (S400). Figure 19 is a conceptual diagram showing the derivation of vehicle acoustics corresponding to optimal design variables.

Referring to Figure 17, CGAN can be learned using design variables and the corresponding vehicle sound color map. Since random noise is included in the CGAN model, multiple color maps corresponding to design variables can be generated.

Figure 18 shows the step of finding the optimal value of the design variable using GA (S500). In GA, several chromosomes corresponding to design variables are initially created, and then evolutionary operations such as crossover and mutation are used to find the optimal solution that maximizes the objective function within given constraints.

Figure 19 shows the step of calculating the optimal design of the vehicle sound using the optimal solution of the design variable, that is, the optimal value of the design variable (S600). Here, the vehicle sound most preferred by men in their 30s can be derived by inputting the optimal values of the design variables from which GA was extracted into the previously learned CGAN generator.

As can be seen from the above process, when designing according to the deep learning-based design optimization method of the present invention, it can be seen that expert knowledge on acoustics, vehicle engines, etc. was not used in the optimal design process of the vehicle sound system. The only thing used for optimal design of vehicle acoustics is the sound data of vehicle acoustics. Accordingly, if the above-described design optimization method is packaged using a programming language such as Python, it can be easily applied to all fields, even by novice engineers just entering the industry.

As described above, from a technical perspective, the key to the present invention is to greatly reduce the need for domain knowledge by using deep learning technology. At present, algorithms (AE, FNN, GAN, GA, etc.) whose effectiveness has been proven several times were analyzed from the perspective of engineering design and applied to each stage of design to maximize each characteristic. It is significant in that it replaces design variable extraction, which requires great expertise, with AE and presents GAN as a method that can increase the amount of data, which can be the most problematic issue when applying AE.

From an industrial aspect, the present invention can greatly eliminate inefficiencies in industrial structures in product design, etc. In order for novice engineers to participate in the design process in earnest, they need a considerable period of time to build basic knowledge. Considering that science and technology are advancing, this period is expected to become longer. As a result, the process until a product is produced becomes increasingly longer, which leads to a weakening of the competitiveness of the industrial structure. The present invention can help solve these problems.

In addition, the need for domain knowledge is a factor that makes exchange between fields difficult, so design methods tailored to each field are needed, and even if a new design method is developed in a specific industry, there is a problem that it is difficult to immediately apply it to other fields. . By presenting a universally applicable optimization design method, the present invention can help lay the foundation for active exchange and simultaneous development in each field.

The embodiments exemplified above for description of the present invention are only one embodiment in which the present invention is embodied, and can be combined in various forms to realize the gist of the present invention. Therefore, the present invention is not limited to the above-described embodiments, and as claimed in the claims below, various modifications can be made by anyone skilled in the art without departing from the gist of the invention. It will be said that the technical features of the present invention exist to the extent possible.

Claims (12)

As a deep learning-based design optimization system,
an actual data acquisition unit that acquires actual data about the object for which a design model is to be created;
Real data is received from the real data acquisition unit, and using a generative adversarial network (GAN) including a generator and a discriminator, the generator generates virtual data and determines the discriminator. a virtual data generator that generates virtual data by augmenting the real data by calculating data that has a 50% probability of being real among the real data and the virtual data;
An autoencoder (AE) that receives the real data and the virtual data from the real data acquisition unit and the virtual data generation unit, respectively, and includes an encoder that compresses the data and a decoder that restores the compressed data. , Autoencodr) to combine the real data and the virtual data. , when the encoder function is called f and the decoder function is called g , latent value that minimizes the value A design variable extraction unit that obtains and extracts design variables;
Receives the design variable from the design variable extractor, uses a feedforward neural network (FNN), and obtains the design variable An objective function configuration unit that configures an objective function by receiving input and outputting an objective function with an objective value y ;
Data that receives the design variables and the objective function from the design variable extracting unit and the objective function constructing unit, respectively, and is a design variable in the input values of the discriminator and generator of the adversarial generative neural network (GAN) used in the virtual data generating unit Using a conditional adversarial generative neural network (CGAN, Conditional GAN), which modifies the corresponding adversarial generative neural network (GAN) to adjust the mode of data generated by adding a label, the generator generates the design variables z and random. a learning model generator in which noise is input and a virtual design model corresponding thereto is generated, and the discriminator distinguishes the virtual design model created by the generator from an actual design model to generate a learning model;
The design variable is received from the design variable extraction unit, and using a genetic algorithm (GA), the solution to be obtained is expressed as a chromosome corresponding to the design variable and composed of several genes, and the solution composed of several chromosomes is displayed. After creating a group, the fitness of each chromosome is calculated based on the objective function and constraint conditions, and an optimal design variable corresponding to the optimal value of the design variable, which is the solution group that maximizes the objective function for the constraint conditions. An optimal design variable extraction unit that extracts; and
Receives the learning model and the optimal design variables from the learning model generation unit and the optimal design variable extraction unit, respectively, and inputs the optimal design variables into the conditional adversarial generative neural network (CGAN), which is the learning model, to generate an optimal design model. It includes an optimal design model generation unit,
It further includes a previously learned design model storage unit that stores a previously learned design model for another object,
The optimal design model generation unit,
An optimal design model of the target is generated by transfer learning the previously learned design model received from the pre-learned design model storage unit, and when generating the pre-learned optimal design model, learned weights are received, and a portion of the weights or A deep learning-based design optimization system, characterized in that it generates the optimal design model by applying it to learning the optimal design model of the object to be newly created.
(Here, m and n are each natural numbers)
delete delete delete delete delete A deep learning-based design optimization method using a deep learning-based design optimization system, the system comprising:
Obtaining actual data about the object for which a design model is to be created;
The real data is received, and using a generative adversarial network (GAN) including a generator and a discriminator, the generator generates virtual data and the discriminator matches the real data and the discriminator. generating virtual data by augmenting the real data by calculating data with a 50% probability of being real among the virtual data;
Using an autoencoder (AE, Autoencoder) that receives the real data and the virtual data and includes an encoder that compresses the data and a decoder that restores the compressed data, the real data and the virtual data are received. data , when the encoder function is called f and the decoder function is called g , latent value that minimizes the value Obtaining and extracting design variables;
After receiving the design variable, using a feedforward neural network (FNN), the design variable Constructing an objective function by receiving input and outputting an objective function with an objective value y ;
The mode of data generated by receiving the design variables and the objective function and adding a data label, which is a design variable, to the input values of the discriminator and generator of the adversarial neural network (GAN) used in the virtual data generation step Using a conditional adversarial generative neural network (CGAN, Conditional GAN) that modifies the adversarial generative neural network (GAN) to be adjustable, the generator receives the design variables z and random noise and generates a virtual design model corresponding thereto. and the discriminator distinguishing the virtual design model created by the generator from the actual design model to generate a learning model;
After receiving the design variables and using a genetic algorithm (GA), the solution to be obtained is expressed as a chromosome composed of several genes corresponding to the design variable, and a solution group composed of several chromosomes is generated, calculating the fitness of each chromosome based on the objective function and constraint conditions, and extracting an optimal design variable corresponding to the optimal value of the design variable, which is a solution group that maximizes the objective function for the constraint conditions; and
Receiving the learning model and the optimal design variables, and inputting the optimal design variables into a conditional adversarial generative neural network (CGAN), which is the learning model, to generate an optimal design model,
Further comprising: storing a previously learned design model for another target,
In the step of generating the optimal design model,
An optimal design model of the target is generated by transfer learning the previously learned design model received from the pre-learned design model storage unit, and when generating the pre-learned optimal design model, learned weights are received, and a portion of the weights or A deep learning-based design optimization method, characterized in that the optimal design model is generated by applying it to learning the optimal design model of the object to be newly created.
(Here, m and n are each natural numbers) delete delete delete delete delete