Generative model-based synthetic data generation method and system

Abstract

This invention provides a method and system for generating sophisticated synthetic data with high quality. [Solution] The synthetic data generation method includes the steps of acquiring source shape data and a target defect profile, extracting command information from the acquired data, extracting topology information of the source shape from the acquired data, and using the target defect profile as a condition for a generative model to generate synthetic data of a defect shape in which the target defect profile is reflected in the source shape from the command information and topology information. This method generates sophisticated synthetic defect data in a controllable manner, thereby effectively solving the problem of securing anomaly data for the defect detection model.

Description

Technical Field

[0001] The present disclosure relates to a technique for generating synthetic data using a generative model.

Background Art

[0002] Anomaly detection is recognized as an important task in various industrial fields such as manufacturing, healthcare, security, and finance. Machine learning or deep learning-based anomaly detection models are utilized as powerful tools for effectively performing such tasks. For example, anomaly detection models are used to automatically detect product defects in manufacturing and to detect diseases at an early stage in the medical field.

[0003] By the way, in order to effectively train such an anomaly detection model, a fairly large amount of training sets are required, and securing abnormal data acts as a major challenge. Usually, normal data can be easily secured, but abnormal data has a low occurrence frequency, so it is realistically difficult to secure sufficient actual cases.

[0004] Abnormal data is difficult to predict due to its inherent characteristics and exists in various forms. Therefore, if sufficient data cannot be secured, the performance of the anomaly detection model deteriorates. Therefore, in order to build a high-performance anomaly detection model, a method for generating abnormal data close to reality (so-called "data augmentation method") is required. In particular, a method for generating various forms of abnormal data in a controllable manner is required so that the anomaly detection model can learn various types of abnormal patterns.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

[0006] The various technical challenges that we aim to address through some embodiments of this disclosure relate to methods and systems for generating synthetic data.

[0007] Specifically, the objective is to provide a method and system capable of generating sophisticated synthetic data using a generative model. For example, the objective is to provide a method and system capable of generating sophisticated synthetic data of anomaly classes, such as defect shape data (e.g., 3D model data, shape images, etc.).

[0008] Another technical challenge is to provide a method and system that can utilize generative models in a controllable manner to generate diverse synthetic data. For example, to provide a method and system that can accurately generate synthetic data of anomaly classes (e.g., defect shape data corresponding to a defect profile) according to a target anomaly profile. [Means for solving the problem]

[0009] A method for generating synthetic data to solve the above-mentioned technical problems is a method performed by at least one processor and may include the steps of: acquiring source shape data and a target defect profile; extracting command information from the acquired data (the command information is information relating to commands performed to generate the source shape); extracting topology information of the source shape from the acquired data; and using the target defect profile as a condition for a generative model, generating synthetic data of a defect shape in which the target defect profile is reflected in the source shape from the command information and the topology information.

[0010] The acquired data is CAD data, and the command information may include a sequence of CAD commands performed at the user interaction level.

[0011] The target defect profile may include information regarding the type and weight of the morphological defect.

[0012] The step of extracting the topology information may include: mapping the face elements, edge elements, and vertex elements constituting the source shape to nodes at a first level, a second level, and a third level, respectively, to construct a topology tree (the second level being a sub-level of the first level, and the third level being a sub-level of the second level); encoding the geometric characteristic information of the face elements to generate features for the first level nodes; encoding the geometric characteristic information of the edge elements to generate features for the second level nodes; and encoding the geometric characteristic information of the vertex elements to generate features for the third level nodes.

[0013] The generative model is a diffusion-based model, and the step of generating the synthesized data may include the steps of preparing noise data and performing a denoising process on the noise data based on the command information, the topology information, and the target defect profile.

[0014] The step of preparing the noise data may include the step of generating the noise data by adding noise to at least one of the acquired data, the command information, and the topology information.

[0015] The steps of performing the denoising process may include inputting the command information, topology information, and target defect profile into a neural network-based denoiser to predict noise at a specific time step, and denoising the noise data using the predicted noise.

[0016] The steps of performing the denoising process may include: inputting the command information and the topology information into a neural network-based denoiser to predict the noise at a specific time step; adjusting the predicted noise based on the target defect profile; and denoising the noise data using the adjusted noise.

[0017] The steps of performing the denoising process may include: generating a defect shape image in which the target defect profile is reflected in the source shape based on the acquired data and the target defect profile; inputting the command information, topology information and the defect shape image into a neural network-based denoiser to predict noise at a specific time step; and denoising the noise data using the predicted noise.

[0018] The target defect profile includes information about the type and weight of the defect, and the step of generating the composite data may include the step of generating multiple composite data while changing the type and the weight values.

[0019] The aforementioned composite data sets are used to construct a defect detection model.

[0020] The training process for the generative model may include the steps of: generating a synthetic defect data sample from a source shape data sample according to a specified defect profile; determining the defect profile of the synthetic defect data sample via a discriminator; and updating the parameters of the generative model using a loss based on the result of the determination.

[0021] The training process for the generative model may include the steps of: generating a synthetic defect data sample from a source shape data sample according to a specified defect profile; obtaining a target defect data sample corresponding to the specified defect profile; determining the similarity between the synthetic defect data sample and the target defect data sample via a discriminator; and updating the parameters of the generative model using a loss based on the result of the determination.

[0022] A method for generating synthetic data to solve the above-mentioned technical problems is a method performed by at least one processor and may include the steps of: acquiring source shape data and a target shape profile; extracting command information from the acquired data (the command information is information relating to commands performed to generate the source shape); extracting topology information of the source shape from the acquired data; and using the target shape profile as a condition for a generative model, generating synthetic data of a target shape in which the target shape profile is reflected in the source shape from the command information and the topology information.

[0023] A synthetic data generation system for solving the above-described technical problems includes one or more processors and a memory storing a computer program executed by the one or more processors. The computer program can include instructions for obtaining data in a source shape and a target defect profile, extracting command information from the obtained data (the command information is information regarding commands performed to generate the source shape), extracting topology information of the source shape from the obtained data, and using the target defect profile as a condition of a generative model to generate synthetic data of a defective shape in which the target defect profile is reflected in the source shape from the command information and the topology information.

[0024] A non-transitory computer-readable recording medium for solving the above-described technical problems is a non-transitory computer-readable recording medium storing instructions that cause at least one processor to perform a synthetic data generation method when executed by the at least one processor. The synthetic data generation method can include steps of obtaining data in a source shape and a target defect profile, extracting command information from the obtained data (the command information is information regarding commands performed to generate the source shape), extracting topology information of the source shape from the obtained data, and using the target defect profile as a condition of a generative model to generate synthetic data of a defective shape in which the target defect profile is reflected in the source shape from the command information and the topology information. [Effect of the Invention]

[0025] According to the present invention, by using a generative model, various synthetic data can be easily generated from source data. For example, by using a diffusion-based generative model, sophisticated synthetic data can be generated with high quality.

[0026] Also, by generating synthetic defect data from source shape data, the problem of securing anomaly data is solved, and a high-performance defect detection model can be easily constructed. Furthermore, the difficulty of securing a training set is effectively solved in other anomaly detection fields as well.

[0027] Also, by using a defect profile as a condition of the generative model, various synthetic defect data can be generated in a controllable manner. For example, by conditioning a generative model with a defect profile including defect type and weight information, synthetic defect data with different defect types and degrees can be freely generated.

[0028] Also, by providing both command information (i.e., user interaction level information) and topology information (i.e., shape data level information) extracted from source shape data to the generative model, high-quality synthetic defect data can be easily generated.

[0029] Also, by constructing a topology tree in which the hierarchical relationship between shape elements is reflected from source shape data and generating node features of the topology tree based on the geometric characteristic information of each shape element, the topology information of the source shape is accurately extracted.

Brief Description of Drawings

[0030] [Figure 1] It is an exemplary diagram for explaining the operation of the synthetic data generation system at the system level. [Figure 2]This is an illustrative diagram to supplement the explanation of how the synthetic data generation system works. [Figure 3] This is an illustrative diagram illustrating the structure and operation of a generative model. [Figure 4] This is an illustrative diagram illustrating the operation of a topology information extractor. [Figure 5] This is an illustrative diagram illustrating the structure and operation of a diffusion-based generator. [Figure 6] This is an illustrative diagram illustrating methods for conditioning and denoising defect profiles. [Figure 7] This is an illustrative diagram illustrating methods for conditioning and denoising defect profiles. [Figure 8] This is an illustrative diagram illustrating methods for conditioning and denoising defect profiles. [Figure 9A] This is an illustrative diagram illustrating methods for conditioning and denoising defect profiles. [Figure 9B] This is an illustrative diagram illustrating methods for conditioning and denoising defect profiles. [Figure 9C] This is an illustrative diagram illustrating methods for conditioning and denoising defect profiles. [Figure 9D] This is an illustrative diagram illustrating methods for conditioning and denoising defect profiles. [Figure 10] This is an illustrative diagram illustrating the operation of the discriminator. [Figure 11] This is an illustrative diagram illustrating the operation of the discriminator. [Figure 12] This is an illustrative flowchart that provides a general overview of the method for generating synthesized data. [Figure 13] Figure 12 is an illustrative diagram to supplement the training set preparation steps. [Figure 14]This is an illustrative diagram to supplement the generative model building steps in Figure 12. [Figure 15] Figure 12 is an exemplary flowchart illustrating the detailed process of the synthetic data generation step. [Figure 16] This diagram shows an exemplary computing device capable of realizing a synthetic data generation system. [Best Mode for Carrying Out the Invention]

[0031] Various embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The advantages and features of this disclosure, and how they are achieved, will become clear when referring to the embodiments described in detail below with reference to the accompanying drawings. However, the technical idea of ​​this disclosure is not limited to the following embodiments and can be realized in various different forms, and the following embodiments are provided merely to complete the technical idea of ​​this disclosure and to fully inform those who are ordinary skill in the art to which this disclosure pertains, and the technical idea of ​​this disclosure is defined by the scope of the claims.

[0032] In describing various embodiments of this disclosure, if it is determined that a specific description of such known configurations or functions would obscure the gist of this disclosure, such detailed description will be omitted.

[0033] Unless otherwise defined, the terms used in the following embodiments (including technical and scientific terms) are used in a way that is commonly understood by a person of ordinary skill in the art to which this disclosure pertains, although this may change depending on the intent of the technicians in the relevant field, case law, the emergence of new technologies, etc. The terms used in this disclosure are for illustrative purposes only and do not limit the scope of this disclosure.

[0034] In the following examples, singular expressions include plural concepts unless explicitly stated to be singular in context. Similarly, plural expressions include singular concepts unless explicitly stated to be plural in context.

[0035] Furthermore, terms such as 1st, 2nd, A, B, (a), (b), etc., used in the following embodiments are merely used to distinguish one component from another, and do not limit the nature, order, or sequence of the component in question.

[0036] The components described with reference to terms such as “part,” “unit,” “module,” “block,” and “or,” as used in the following embodiments, and the functional blocks shown in the drawings, are embodied in the form of software, hardware, or a combination thereof. The software may be, for example, machine code, firmware, embedded code, or application software. The hardware may include, for example, electrical circuits, electronic circuits, processors, computers, integrated circuits, integrated circuit cores, passive components, or a combination thereof.

[0037] Various embodiments of this disclosure will be described in detail below with reference to the attached drawings.

[0038] Figure 1 is an illustrative diagram illustrating the operation of the synthetic data generation system 10 at the system level.

[0039] As shown in Figure 1, the synthetic data generation system 10 is a computing device / system capable of generating diverse synthetic data 13 from source data 12 using a generative model 11. For example, the synthetic data generation system 10 can generate diverse synthetic shape data from source shape (or geometry) data via the generative model 11. As a more specific example, the synthetic data generation system 10 can generate diverse synthetic defect data from source shape data using the generative model 11. For details on this, please refer to the explanation in Figure 2 and other figures.

[0040] Such a synthetic data generation system 10 can also be described using expressions such as "data augmentation device / system," "virtual data generation device / system," "artificial data generation device / system," or "defective / abnormal data generation device / system," and can also be abbreviated as "data generation device / system."

[0041] For the sake of understanding, the following explanation will assume that the source data 12 is "data indicating the shape" of an object (e.g., a person, thing, product, etc.). However, the scope of this disclosure is not limited to this, and the technical ideas described later are also applicable when the source data 12 is other types of data.

[0042] Furthermore, for the sake of clarity, the synthetic data generation system 10 will be abbreviated as "the system" below.

[0043] Figure 2 is an illustrative diagram that provides supplementary explanation of the operation of system 10.

[0044] As shown in Figure 2, the system 10 can generate various composite defect data 23 from source shape data 22 via the generative model 11. For example, the system 10 can use a defect profile as a condition in the generative model 11 to generate defective shape data that reflects the defect profile from normal shape data 22 (i.e., data showing a shape without defects). In some cases, the system 10 can also use a shape profile (or transformation / deformation profile) as a condition in the generative model 11 to generate data of other shapes (e.g., shapes with deformed length, width, size, bend, ratio, structure, etc.) from the source shape data 22. Alternatively, the system 10 can use a style profile as a condition in the generative model 11 to generate shape data of other styles from the source shape data 22.

[0045] Source shape data 22 is the underlying data used to generate composite defect data, and as described above, it means data that shows (or represents) the shape of a specific object. Examples of such data include, but are not limited to, data relating to 2D models / shapes, 3D models / shapes (e.g., 3D CAD, 3D mesh models, point clouds, etc.), and images (e.g., 2D images, 3D images, etc.). Examples of 3D model data include, but are not limited to, command information for generating 3D models / shapes (e.g., CAD command sequences and parameter values ​​for each command performed at the user interaction level to generate a specific shape), B-rep (Boundary representation), point clouds, and mesh models. Source shape data 22 may or may not be considered as a condition of the generative model 11. Source shape data 22 may be actual data or composite data. For example, composite defect data generated via the generative model 11 may again be used as source shape data 22.

[0046] The source shape data 22 can also be referred to as "basic shape data," "original shape data," "reference shape data," or "input shape data," and abbreviations such as "source data" or "basic data" can also be used.

[0047] A defect profile is a set of information that defines a geometric defect and is used to control the operation (or output) of the generative model 11. Examples of geometric defects include, but are not limited to, wear, dents, holes, scratches, cuts, crushing, and distortion. The scope of geometric defects includes not only surface defects but also structural defects.

[0048] A defect profile consists of one or more attributes that indicate a morphological defect, and examples of such attributes include, but are not limited to, the type of defect, weight (i.e., degree / level / severity of the defect), location, orientation, size, and shape.

[0049] Defect profiles can also be expressed using terms such as "anomaly profile," "defect information," "shape profile," "control profile," and "deformation / transformation profile."

[0050] The composite defect data is composite data generated by reflecting the defect profile in the source shape data 22. In other words, the composite defect data can be understood as data that shows the defect shape generated by reflecting the defect profile in the source shape. The composite defect data may be in the same format as the source shape data 22, or it may be in a different format.

[0051] Synthetic defect data can also be expressed using terms such as "synthetic shape data," "defect shape data," or "target defect data," and abbreviations such as "defect data" or "synthetic data" can also be used.

[0052] The generative model 11 is a model equipped with data generation capabilities, in which a defect profile is input as a condition, and the model generates synthetic defect data from source geometry data 22. Such a generative model 11 may be a diffusion-based model or a conditional model, but the scope of this disclosure is not limited thereto. Depending on the case, the generative model 11 may be embodied in other types of models, such as a Variational Autoencoder (VAE). Examples of diffusion-based models include the Denoising Diffusion Probabilistic Model (DDPM) and score-based generative models.

[0053] Detailed information regarding the structure, operation, and training methods of the generative model 11 will be described later with reference to the diagrams in Figure 3 and below.

[0054] As shown in the figure, the diverse composite defect data 23 generated from the source shape data 22 can be used to train (or build) a defect detection model 21. In this case, the system 10 can generate diverse composite defect data 23 from the source shape data 22 while changing the attribute values ​​of the defect profile (e.g., values ​​such as defect type and weight), and can also generate further composite defect data using other source shape data. Furthermore, the system 10 can also generate further composite defect data using specific composite defect data as source shape data.

[0055] For example, suppose the composite defect data is B-rep data (e.g., 3D model data in B-rep format). In this case, system 10 can render each of the composite defect data 23 to generate defect shape image samples and add these to extend the existing training set (e.g., a training set consisting of many normal shape image samples and a small number of defect shape image samples). Then, system 10 can use the extended training set to construct a defect detection model 21. In this way, a high-performance defect detection model 21 can be easily constructed.

[0056] For reference, the term "sample" can be used to refer to individual data points in a dataset (e.g., a training set), or to refer to individual data points in a training step, distinguishing them from inference steps.

[0057] Furthermore, a portion of the synthetic defect data 23 generated by the generative model 11 may be used as a normal data sample when constructing the defect detection model 21. For example, suppose that a particular synthetic defect data is synthetic data of a shape in which only minute defects exist (hereinafter referred to as "miniature defect shape"). And suppose that such a minute defect shape is judged as normal according to the actual industrial defect judgment criteria. In such a case, the synthetic data of the minute defect shape is used as a normal data sample when constructing the defect detection model 21 (i.e., a normal label is assigned to the image sample of the minute defect shape). The synthetic data of the minute defect shape may be generated, for example, by setting the weight values ​​of the defect profile to be below a standard value, but the scope of this disclosure is not limited thereto.

[0058] The system 10 described above is implemented by at least one computing device. For example, all functions of system 10 may be implemented by a single computing device, or the first function of system 10 may be implemented by the first computing device and the second function by the second computing device. Alternatively, specific functions of system 10 may be implemented by multiple computing devices.

[0059] The term "computing device" can encompass any device equipped with computing capabilities; see Figure 16 for an example of such a device. Since a computing device is a collection of interacting components (e.g., memory, processors, etc.), the term "computing system" can also be used. Of course, the term "computing system" can also encompass the concept of a collection of interacting computing devices.

[0060] The operation of system 10 has been described above with reference to Figures 1 and 2. Below, the structure and operation of the generative model 11 will be described in detail with reference to Figures 3 to 11.

[0061] Figure 3 is an illustrative diagram illustrating the structure and operation of the generative model 11. Blocks 31-34 shown in Figure 3 may be understood as components of system 10.

[0062] As shown in Figure 3, the generative model 11 comprises a command information extractor 31, a topology information extractor 32, a generator 33, and a discriminator 34. However, from some perspectives, the command information extractor 31, the topology information extractor 32, and / or the discriminator 34 may not be considered components of the generative model 11.

[0063] The command information extractor 31 is a module that extracts command information from the source shape data 35. As described above, the command information may include a command sequence performed to generate a specific shape at the user interaction level, and the parameter values ​​for each command.

[0064] For example, if the source shape data 35 contains a command history (e.g., if it is a CAD file containing command records), the command information extractor 31 can extract the command history and, if necessary, process it into an appropriate format.

[0065] As another example, the command information extractor 31 can extract command information by analyzing the source shape data 35. For example, the command information extractor 31 can analyze the source shape data 35 based on a predefined command description or specification to derive the command sequence and parameter values ​​for each command necessary to generate the source shape. In some cases, the command information extractor 31 can also determine the final command information by repeatedly performing the steps of deriving command information, generating a temporary shape from the command information, comparing the temporary shape with the source shape and identifying parts where the difference is greater than or equal to a threshold value (e.g., parts such as surfaces and lines), and correcting the command information of the identified parts (e.g., repeating until the similarity of the shapes exceeds a threshold).

[0066] As another example, the command information extractor 31 can also extract command information based on various combinations of the examples described above.

[0067] For information on the types of commands and parameters used in CAD (Computer-Aided Design) tools, please refer to Table 1 below.

[0068] [Table 1]

[0069] Next, the topology information extractor 32 is a module that extracts topology information from the source shape data 35. The topology information extractor 32 can analyze the source shape data 35 in various ways and extract topology information of the source shape. For example, the topology information extractor 32 can analyze the source shape data 35 to identify (or recognize) shape elements such as faces, edges, and vertices that constitute the source shape, and the connections between them, and generate topology information from the identification results. If the source shape data 35 includes B-rep data, such shape elements can be easily identified from the B-rep data.

[0070] In some embodiments, topology information can be extracted through the process by which the topology information extractor 32 hierarchically organizes the elements of the source shape to construct a topology tree. For example, as shown in Figure 4, the topology information extractor 32 can analyze the source shape data 35 to identify the faces, lines, and point elements that constitute the source shape 41. Next, the topology information extractor 32 can construct a topology tree 42 by mapping the faces (e.g., F1), lines (e.g., E1), and point elements (e.g., V1) to nodes (e.g., 44-48) at the first, second, and third levels, respectively (for reference, the root node at the 0th level corresponds to the source shape 41). At this point, the nodes of the topology tree 42 are connected according to the linkage relationships (or topology / inclusion relationships) between the elements that constitute the source shape 41. Next, the topology information extractor 32 encodes the geometric characteristic information of each face element to generate features for mapping nodes (i.e., face nodes), and can also generate features for mapping nodes of each line element (i.e., line nodes) and each point element (i.e., point nodes) in the same manner. The node features consist of, for example, a position feature (e.g., bounding box information) and a shape feature (however, the feature of a point node consists only of a position feature), and the shape feature may be generated by, for example, encoding the shape characteristic information of the shape element (e.g., using an encoder such as an autoencoder or VAE (Variational AutoEncoder)), but the scope of this disclosure is not limited thereto. For reference, if a particular point element (e.g., V2) or a particular line element (e.g., E2) belongs to two or more higher-level shape elements, duplicate nodes (e.g., 47 and 48) will occur in the topology tree 42, but such duplicate nodes can serve to more clearly represent the topological relationships of the source shape 41.

[0071] Let's refer to Figure 3 again for further explanation.

[0072] The generator 33 is a module that receives command information and topology information of the source shape, as well as a defect profile 36, as input, and generates composite defect data 37 that conforms to the defect profile 36. The command information, topology information, and defect profile 36 are each embedded (or encoded) in an appropriate form and input to the generator 33.

[0073] The generator 33 may be embodied, for example, by a diffusion-based conditional model. However, the scope of this disclosure is not limited thereto. For example, the generator 33 may be embodied based on a neural network such as UNet. However, for the sake of understanding, the following explanation will assume that the generator 33 is embodied by a "diffusion-based conditional model". Examples of such models include conditional DDPM and conditional score-based generative models. Since those working in the art are already familiar with the operating principles of conditional DPPM and conditional score-based generative models, a detailed explanation of these will be omitted.

[0074] The structure and operation of the diffusion-based generator 33 will be further explained below with reference to Figures 5 to 8.

[0075] Figure 5 is an illustrative diagram illustrating the structure and operation of the diffusion-based generator 33. Figure 5 and others assume that the generator 33 is implemented using a conditional DDPM (or diffusion stochastic model).

[0076] As shown in Figure 5, the generator 33 is configured to generate composite defect data 53 corresponding to the defect profile 36 via a denoising process (or reverse process). However, in some cases, the generator 33 can also generate intermediate composite defect data (e.g., topology tree, command information, etc.) via the denoising process and then post-process this to generate composite defect data (e.g., B-rep data, command information, defect shape image, etc.).

[0077] The generator 33 may include a denoiser 51 that predicts noise at specific time steps in order to perform the denoising process. The defect profile 36 can then be used as a condition for the generator 33 or the denoiser 51. Although not explicitly shown in Figure 5, command information and topology information of the source geometry can also be used to generate the composite defect data 53. The composite defect data 53 may correspond to the composite defect data 37 in Figure 3.

[0078] For reference, if the generator 33 is implemented as a conditional score-based generative model, the denoiser 51 can be replaced by a "score predictor," which is a neural network that predicts the score at a specific time step.

[0079] Specifically, the generator 33 can generate high-quality composite defect data 53 from noise data 52 through a denoising process that repeatedly predicts noise at a specific time step via a denoiser 51 and removes (or purifies) the predicted noise. In this case, the noise data 52 may be generated by adding noise to at least one of the source shape data 35, command information, and topology information.

[0080] In Figure 5 and other diagrams, "X" represents data, the subscript represents a time step, and "T" represents the maximum time step. Also, in the denoising process, "XT " represents noise data, and "X0" represents the synthetic data generated as a result of the process. Also, although Figure 5 shows that the generator 33 generates an image, this is only for the sake of understanding, and the generator 33 can also generate model data such as B-rep.

[0081] The denoiser 51 can be understood as a neural network that predicts noise at a specific time step. The denoiser 51 is trained using noise added in a diffusion process (or forward process), which is explained in Figure 14.

[0082] Denoiser 51 can also be described as a "noise predictor" or a "noise predictive neural network."

[0083] On the other hand, the specific methods for conditioning the defect profile 36 and the denoising methods will vary depending on the embodiment.

[0084] In some embodiments, as shown in Figure 6, the defect profile 36 is directly conditioned by the denoiser 51. That is, the defect profile 36 is input to the denoiser 51 as a condition for predicting noise at a specific time step. For example, the denoiser 51 is configured to receive command information 61, topology information 62, and the defect profile 36 as inputs and output predicted noise at a specific time step. In such a case, the noise for generating synthetic defect data is accurately predicted. The denoiser 51 can also be configured to receive further inputs of time step values.

[0085] In some other embodiments, as shown in Figure 7, a defect profile 36 is used to adjust the predicted noise 73. Specifically, the denoiser 51 is configured to receive command information 71 and topology information 72 as inputs and to output predicted noise 73 at a specific time step, and the defect profile 36 is used to adjust the predicted noise 73. The adjustment of the predicted noise 73 may be performed, for example, via a neural network layer trained with the denoiser 51 (e.g., a neural network layer that receives predicted noise 73 and defect profile 36 as inputs and outputs adjusted noise 74), but the scope of this disclosure is not limited thereto. In this embodiment, adjusted noise 74 is used for the denoising process instead of predicted noise 73.

[0086] In some other embodiments, as shown in Figure 8, the source shape image 83 is converted to a temporary defect shape image 84 that reflects the defect profile 36. In some cases, the temporary defect shape image 84 (or defect shape data) may be generated by reflecting the defect profile 36 in the source shape data 35. The defect shape image 84 is conditioned by a denoiser 51, which is configured to take command information 81, topology information 82, and the defect shape image 84 as inputs and output predictive noise. If the source shape data 35 is not an image, the source shape image 83 is generated, for example, by rendering the source shape data 35. However, the scope of this disclosure is not limited thereto. The source shape image 83 may also be converted to a defect shape image 84, for example, via a deep learning-based image conversion model, but the scope of this disclosure is not limited thereto.

[0087] In some other embodiments, as shown in Figures 9A to 9D, the topology information is organized in the form of a tree (see Figure 4), and the denoiser 51 is configured to include multiple sub-denoisers 91 to 94. Then, via the sub-denoisers 91 to 94, the denoising work at each time step can proceed (or be performed) sequentially along the lower levels of the topology tree. This will be explained in more detail later.

[0088] In several other embodiments, the defect profile 36 may be conditioned or denoised based on various combinations of the embodiments described above.

[0089] Below, we will provide supplementary explanations of the denoising method by referring to Figures 9A to 9D.

[0090] Figures 9A to 9D are illustrative diagrams showing the process of sequentially performing denoising operations according to several embodiments of this disclosure. Figures 9A to 9D assume that the node features of the topology tree consist of "location features" and "shape features" (however, node features of point elements consist only of location features), that the topology tree is generated through the denoising process (i.e., the topology tree is intermediate composite defect data), and that composite defect data (e.g., B-rep data) is finally generated through post-processing of the topology tree. However, the scope of this disclosure is not limited thereto, and the number and types of detailed shape features constituting the node features, the resulting data of the denoising process, etc., can be designed in a variety of ways.

[0091] As shown in Figures 9A to 9D, the denoiser 51 is composed of multiple subdenoisers 91 to 94, which can be used to perform denoising along the lower levels of the topology tree at each time step. Each of the subdenoisers 91 to 94 is implemented based on a neural network.

[0092] Specifically, as shown in Figure 9A, denoising of the node location feature 96-1 is performed at the face level 95-1 of the topology tree via the first subdenoiser 91. The first subdenoiser 91 receives the noisy location feature 96-1, command information 97-1, and defect profile 97-2 as input, predicts the noise for a specific time step, and uses this to denoise the noisy location feature 96-1.

[0093] Next, as shown in Figure 9B, denoising of the node's shape feature 96-3 is performed at the face level 95-1 of the topology tree via the second subdenoiser 92. The second subdenoiser 92 is input with the denoised position feature 96-2, the noisy / noise shape feature 96-3, command information 97-1, and defect profile 97-2 to predict the noise for a specific time step, and can use this to denoise the noisy / noise shape feature 96-3.

[0094] Next, as shown in Figure 9C, denoising of the node's location feature 98-1 is performed at the line level 95-2 of the topology tree via the third subdenoiser 93. The third subdenoiser 93 is input with the features 96 of the denoised face node (i.e., location feature and shape feature), the noisy location feature 98-1, command information 97-1, and defect profile 97-2 to predict the noise at a specific time step, and can use this to denoise the noisy location feature 98-1.

[0095] Next, as shown in Figure 9D, denoising is performed on the remaining node features 99-1 (i.e., shape features of line nodes and position features of point nodes) of the line and point levels 95-3 of the topology tree via the fourth subdenoiser 94. The fourth subdenoiser 94 is input with the features 96 (i.e., position features and shape features) of the denoised face nodes, the position features 98-2 of the denoised line nodes, the noisy node features 99-1, command information 97-1, and defect profiles 97-2 to predict the noise at a specific time step, and can use this to denoise the noisy node features 99-1 (see 99-2 for the denoising results).

[0096] Through the sequential denoising process described above, a topology tree and its node features reflecting the defect profile 97-2 are accurately generated, resulting in the creation of sophisticated synthetic defect data.

[0097] The denoising methods have been explained sequentially with reference to Figures 9A to 9D.

[0098] Referring again to Figure 3, we will continue the explanation of the generator 33.

[0099] In some embodiments, weights for command information and topology information are derived via a neural network layer. The command information and topology information, with the weights reflected, are then input to the generator 33. The neural network layer is configured to receive defect profile 36 and source shape data 35 as input and output weights for command information and topology information, and is trained together with the generative model 11. In such cases, the generator 33 can dynamically adjust the degree of concentration on command information and topology information according to the source shape and defect profile 36, thereby further improving the quality of the generated synthetic defect data 37.

[0100] Next, the discriminator 34 is a module that performs discrimination operations related to the composite defect data 37, and the specific discrimination method varies depending on the embodiment.

[0101] In some embodiments, the discriminator 34 is configured to determine whether the input synthetic defect data is synthetic data or actual data. If the synthetic defect data is not an image, the discriminator 34 can either accept a defect shape image generated by rendering the synthetic defect data as input, or it can accept the synthetic defect data as input. Such technical details are applicable to other examples as well.

[0102] In some other embodiments, as shown in Figure 10, the discriminator 34 is configured to determine the defect profile 101 (e.g., defect type, weight, etc.) of the input synthetic defect data 37.

[0103] In some other embodiments, as shown in Figure 3 or Figure 11, the discriminator 34 is configured to determine the similarity 111 between the composite defect data 37 and the target defect data 38 (hereinafter referred to as "target defect data"). Here, the target defect data 38 means defect shape data that corresponds to the defect profile 36, and can be understood as a kind of ground truth data. The discriminator 34 may or may not receive both the composite defect data 37 and the target defect data 38 as input. For example, the discriminator 34 may determine the defect profile of the composite defect data 37 and compare it with the defect profile 36 of the target defect data 38 to derive the similarity.

[0104] In some cases, the discriminator 34 may be configured to output a patch-level similarity map. For example, the discriminator 34 can receive both a composite defect shape image and a target defect shape image as input, decompose and compare them on a patch-by-patch basis, and then output a similarity map. Such a similarity map can further improve the performance of the generator 33 by providing direct information about the missing parts in the composite defect shape image.

[0105] In several other embodiments, the discriminator 34 can also operate based on various combinations of the embodiments described above.

[0106] The structure and operation of the generative model 11 have been described above with reference to Figures 3 to 11. Below, various methods performed in the system 10 described above will be explained with reference to Figures 12 and subsequent drawings.

[0107] In the following explanation, for the sake of clarity, we will assume that all steps / operations of the method described below are performed on the system 10 described above. Therefore, if the subject of a particular step / operation is omitted, it can be understood that that step / operation is performed on the system 10. However, in a real environment, some steps / operations of the method described below may be performed on other computing devices. For example, training of the generative model 11 may be performed on other computing devices.

[0108] Figure 12 is an exemplary flowchart illustrating a method for generating composite data. However, this is merely an exemplary embodiment for achieving the objectives of this disclosure, and of course, some steps may be added or omitted as needed. Figure 12 assumes that the generated composite data is "defect shape data".

[0109] As shown in Figure 12, the synthetic data generation method according to the embodiment can be started from step S121, which involves preparing a training set for the generative model 11.

[0110] For example, as shown in Figure 13, the training set 131 consists of pairs of source shape data samples (e.g., S1, hereinafter referred to as "source shape sample") and target defect data samples (e.g., T2, hereinafter referred to as "target defect sample"), and the source shape sample (e.g., S1) and target defect samples (e.g., T1-1, T1-2) may have a one-to-many relationship. The source shape sample (e.g., S1) may be a data sample that shows (or represents) a normal shape, but the scope of this disclosure is not limited thereto. The training set 131 may further include defect profiles (e.g., P2) of the target defect samples (e.g., T2).

[0111] Let's refer to Figure 12 again for further explanation.

[0112] In step S122, a generative model 11 is constructed (or trained) using the training set to generate synthetic defect data based on the input defect profile. However, the specific training method and process will vary depending on the form of the generative model 11.

[0113] For ease of understanding, the training process of the generative model 11 will be described below with reference to Figure 14.

[0114] Figure 14 is an illustrative diagram showing the training process of the generative model 11. Figure 14 assumes that the generative model 11 (more precisely, the generator 33) is realized based on "conditional DDPM".

[0115] As shown in Figure 14, the system 10 can train the generator 33 (more precisely, the denoiser 51 of the generator 33) through a diffusion process and a denoising process.

[0116] Specifically, in the diffusion process, system 10 can progressively noise the target defect sample 144 by adding noise from a prior distribution (e.g., a Gaussian distribution) at each time step. The added noise is used as ground truth noise to update the parameters of the denoiser 51.

[0117] In the denoising process, system 10 can predict the noise at each time step via the denoiser 51 and use this to progressively denoise the noise data sample 145. At this time, system 10 can predict the noise using the defect profile 142 as a condition for the denoiser 51, as explained in Figures 5 to 9D. Next, system 10 can calculate a first loss L1 which represents the difference between the predicted noise and the noise added in the diffusion process.

[0118] For reference, the noise data sample 145 may be generated by sampling from a prior distribution, or by adding noise from the prior distribution to at least one of the source shape sample 141, command information, and topology information. For example, if information / data in the form of a topology tree is generated via a denoising process, the noise data sample 145 is generated by adding noise to the topology information. Similarly, if information / data in the form of a command sequence is generated, the noise data sample 145 is generated by adding noise to the command information.

[0119] Furthermore, the system 10 can calculate a second loss L2 based on the similarity between the synthetic defect data sample 143 (hereinafter referred to as "synthetic defect sample") and the target defect sample 144 via the discriminator 34. Alternatively, the system 10 can also determine, via the discriminator 34, whether or not the synthetic defect sample 143 corresponds to actual data and calculate a loss based on the determination result.

[0120] In some cases, system 10 may also compare the synthetic defect sample 143 with the target defect sample 144 at the image level or feature level to calculate a third loss (e.g., perceptual loss, pixel-level loss, etc.).

[0121] Next, system 10 can update the parameters of the generative model 11 based on the calculated losses L1 and L2. For example, system 10 can calculate the total loss by combining the losses L1 and L2, and update the parameters of the generator 33 (more precisely, the denoiser 51) based on this.

[0122] System 10 can repeat the above process for other source shape samples and target defect samples. In this way, a generative model 11 is constructed that can precisely control the generation of synthetic defect data based on the defect profile.

[0123] The exemplary training process of the generative model 11 has been described above with reference to Figure 14.

[0124] Let's refer to Figure 12 again for further explanation.

[0125] In step S123, source geometry data and target defect profile are acquired. For example, the system 10 may receive source geometry data and target defect profile input from the user. As mentioned above, the source geometry data may include command information, model data such as B-rep, images, etc. The target defect profile may include information (or values) about the type and weight of the geometric defect, for example.

[0126] In step S124, the target defect profile is used as a condition for the generative model 11 to generate composite data of the defect shape in which the target defect profile is reflected in the source shape. The detailed process of step S124 is shown in Figure 15.

[0127] Figure 15 is an exemplary flowchart illustrating the detailed process of step S124. However, this is merely an exemplary embodiment for achieving the objectives of this disclosure, and of course, some steps may be added or omitted as needed. Figure 15 assumes that the generative model 11 (more precisely, the generator 33) is embodied based on a "conditional DDPM".

[0128] In step S151, command information and topology information are extracted from the source shape data. As described above, the command information may include the command sequence performed at the user interaction level to generate the source shape and the parameter values ​​of each command. Step S151 corresponds to the operation of the command information extractor 31 and the topology information extractor 32, so please refer to the explanations in Figures 3 and 4 for further details.

[0129] In step S152, noise data is prepared. For example, system 10 can sample noise data from a prior distribution. As another example, system 10 can generate noise data by adding noise from a prior distribution to at least one of the source shape data, command information, and topology information. In such a case, the inclusion of source shape information in the noise data generates sophisticated composite defect data that accurately reflects the characteristics of the source shape.

[0130] In step S153, a denoising process is performed on the noise data based on the command information, topology information, and defect profile, resulting in the generation of synthetic defect data. However, the specific generation method varies depending on the embodiment.

[0131] In some embodiments, system 10 can input command information, topology information, and defect profiles to the denoiser 51 to predict noise at a specific time step, and use this to denoise the noise data. Such denoising operations are repeated for other time steps. Such embodiments are further described in Figure 6.

[0132] In some other embodiments, system 10 can input command information and topology information to the denoiser 51 to predict noise for a specific time step, and then adjust the predicted noise based on the target defect profile. System 10 can then denoise the noise data using the adjusted noise. Such embodiments are further described in Figure 7.

[0133] In several other embodiments, the system 10 can generate a temporary defect shape image based on source shape data and a target defect profile, and input command information, topology information, and the defect shape image to the denoiser 51 to predict noise for a specific time step. The system 10 can then denoise the noise data using the predicted noise. Such embodiments are further described in Figure 8.

[0134] In several other embodiments, the noise data is in the form of a topology tree, and the system 10 can perform sequential denoising while moving in the lower levels of the topology tree. For such embodiments, please refer to the descriptions of Figures 9A to 9D.

[0135] In several other embodiments, the denoising process may be carried out based on various combinations of the embodiments described above.

[0136] The method for generating synthetic data has been explained above with reference to Figures 12 to 15. As described above, sophisticated synthetic data can be generated with high quality by using the diffusion-based generative model 11.

[0137] Furthermore, by generating synthetic defect data from source geometry data, the problem of securing anomaly data is solved, and high-performance defect detection models can be easily constructed. In addition, the difficulty of securing training sets is effectively resolved in other fields of anomaly detection.

[0138] Furthermore, by using the defect profile as a condition for the generative model 11, a variety of synthetic defect data can be generated in a controllable manner. For example, by conditioning the generative model 11 with a defect profile that includes defect type and weight information, synthetic defect data with different defect types and degrees can be freely generated.

[0139] Furthermore, by providing both command information (i.e., user interaction level information) and topology information (i.e., shape data level information) extracted from the source shape data to the generative model, high-quality synthetic defect data can be easily generated.

[0140] Furthermore, by constructing a topology tree that reflects the hierarchical relationships between shape elements from the source shape data, and generating node features of the topology tree based on the geometric characteristic information of each shape element, the topology information of the source shape is accurately extracted.

[0141] On the other hand, the aforementioned method for generating synthetic data can also be used to generate other types of synthetic data besides defective synthetic data.

[0142] For example, the synthetic data generation method described above can be used to generate synthetic data of a target shape (e.g., a shape with deformed length, width, size, bending, ratio, structure, etc.) from source shape data by using a target shape profile as a condition for the generative model 11 (i.e., shape transformation / deformation can be performed using a shape profile instead of a defect profile). In this case, the generative model 11 is constructed using a training set consisting of source shape data samples, target shape data samples, and a shape profile, as described with reference to Figure 14. The shape profile may consist of attributes such as the type of shape deformation (or target shape) and weights (i.e., degree / level / intensity of deformation), but the scope of this disclosure is not limited thereto.

[0143] In the following section, an exemplary computing device 160 that can embody the system 10 described above will be explained with reference to Figure 16.

[0144] Figure 16 is an exemplary hardware configuration diagram showing the computing device 160.

[0145] As shown in Figure 16, the computing device 160 may include one or more processors 161, a bus 163, a communication interface 164, a memory 162 for loading computer programs 166 performed by the processors 161, and a storage 165 for storing the computer programs 166. However, only components relevant to the embodiments of this disclosure are shown in Figure 16. Therefore, a person ordinary in the art to which this disclosure belongs will see that in addition to the components shown in Figure 16, other general-purpose components (e.g., input devices such as keyboards and mice, output devices such as speakers and displays) are also included. In other words, the computing device 160 includes a variety of components in addition to those shown in Figure 16. Furthermore, the computing device 160 may be configured in a form in which some of the components shown in Figure 16 are omitted. The components of the computing device 160 will be described below.

[0146] The processor 161 can control the overall operation of each configuration of the computing device 160. The processor 161 may consist of at least one of the following: a CPU (Central Processing Unit), an MPU (Micro Processor Unit), an MCU (Micro Controller Unit), a GPU (Graphic Processing Unit), an NPU (Neural Processing Unit), a TPU (Tensor Processing Unit), a VPU (Vision Processing Unit), an APU (Accelerated Processing Unit), or any form of processor well known in the art of this disclosure. The processor 161 can also perform calculations for at least one application or program to perform a specific operation / step / method. The computing device 160 may comprise one or more processors.

[0147] Next, memory 162 can store various data, instructions, and / or information. Memory 162 can load computer programs 166 from storage 165 to perform specific operations / steps / methods. Memory 162 may be embodied in volatile memory such as RAM, but the technical scope of this disclosure is not limited thereto.

[0148] Next, bus 163 can provide communication functions between the components of the computing device 160. Bus 163 can be implemented in various forms, such as an address bus, a data bus, and a control bus.

[0149] Next, the communication interface 164 can support wireless internet communication of the computing device 160. Furthermore, the communication interface 164 can also support a variety of communication methods other than internet communication. To this end, the communication interface 164 is configured to include communication modules well known in the art of this disclosure.

[0150] Next, the storage 165 can non-temporarily store one or more computer programs 166. The storage 165 comprises non-volatile memory such as ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), flash memory, a hard disk, a removable disk, or any form of computer-readable recording medium well known in the art to which this disclosure belongs.

[0151] Next, when the computer program 166 is loaded into memory 162, it may include instructions that cause the processor 161 to perform a specific action / step / method. That is, the processor 161 can perform a specific action / step / method by executing the loaded instructions.

[0152] For example, the computer program 166 may include instructions for the following operations: acquiring source geometry data and a target defect profile; extracting command information from the acquired data; extracting topology information of the source geometry from the acquired data; and using the target defect profile as a condition for the generative model 11, generating composite data of a defect geometry in which the target defect profile is reflected in the source geometry from the command information and topology information.

[0153] As another example, the computer program 166 may include instructions for the following operations: acquiring source shape data and target shape profile; extracting command information from the acquired data; extracting topology information of the source shape from the acquired data; and using the target shape profile as a condition for the generative model 11, generating composite data of the target shape in which the target shape profile is reflected in the source shape from the command information and topology information.

[0154] As yet another example, computer program 166 may include instructions to perform at least some of the actions / steps / methods described with reference to Figures 1 to 15.

[0155] In the example shown, the system 10 is realized through the computing device 160.

[0156] On the other hand, the computing device 160 shown in Figure 16 may represent a virtual machine implemented based on cloud technology. For example, the computing device 160 may be a virtual machine running on one or more physical servers included in a server farm. In this case, at least a portion of the processor 161, memory 162, and storage 165 shown in Figure 16 may be virtual hardware, and the communication interface 164 may also be implemented as a virtualized networking element such as a virtual switch.

[0157] In summary, an exemplary computing device 160 capable of realizing a treatment support system has been described with reference to Figure 16.

[0158] The above describes an exemplary computing device 160 that can embody system 10 with reference to Figure 16.

[0159] The various embodiments of this disclosure and the effects thereof have been described above with reference to Figures 1 to 16. The effects of the technical concept of this disclosure are not limited to those described above, and other effects not mentioned above will be clearly understood by a person of the ordinary skill from the following description.

[0160] Furthermore, even though the above embodiments have described multiple components as being combined into one or operating in combination, the technical idea of ​​this disclosure is not necessarily limited to such embodiments. That is, within the scope of the purpose of the technical idea of ​​this disclosure, all of its components can also be selectively combined into one or more components and operate in combination.

[0161] The technical concept of this disclosure, as described above, is embodied as computer-readable code on a computer-readable recording medium. A computer program recorded on a computer-readable recording medium is transmitted to another computing device via a network such as the Internet, installed on that computing device, and thereby used on that computing device.

[0162] While the drawings show operations in a specific order, it cannot be understood that the operations must necessarily be performed in the specific order or sequentially shown, or that the desired result can only be obtained by performing all the illustrated operations. In certain situations, multitasking and parallel processing may be advantageous. Although various embodiments of this disclosure have been described above with reference to the attached drawings, a person with ordinary skill in the art to which this disclosure belongs will understand that the technical idea of ​​this disclosure can be implemented in other specific forms without altering the technical idea or essential features. Therefore, the embodiments described above should be understood in all respects as illustrative and not limiting. The scope of protection of this disclosure should be interpreted in accordance with the attached claims, and all technical ideas within an equivalent scope should be interpreted as being included in the scope of rights of the technical idea as defined by this disclosure.

Claims

1. A method performed by at least one processor, The steps include obtaining source geometry data and target defect profile, The steps include: extracting command information from the acquired data (the command information is information about the commands performed to generate the source shape); The steps include extracting topology information of the source shape from the acquired data, A method for generating composite data, comprising the step of using the target defect profile as a condition for a generative model, and generating composite data of a defect shape in which the target defect profile is reflected in the source shape from the command information and the topology information.

2. The acquired data is CAD data. The method for generating composite data according to claim 1, wherein the command information includes a CAD command sequence performed at the user interaction level.

3. The method for generating composite data according to claim 1, wherein the target defect profile includes information regarding the type and weight of the morphological defect.

4. The step of extracting the topology information is: The steps include: constructing a topology tree by mapping the face elements, edge elements, and vertex elements constituting the source shape to nodes at the first, second, and third levels, respectively (the second level is a sublevel of the first level, and the third level is a sublevel of the second level); The steps include encoding the geometric characteristic information of the surface element to generate the features of the first level node, The steps include encoding the geometric characteristic information of the line element to generate the features of the second level node, A method for generating composite data according to claim 1, comprising the step of encoding geometric characteristic information of the point elements to generate features of the third level node.

5. The aforementioned generative model is a diffusion-based model, The step of generating the aforementioned composite data is: Steps to prepare noise data, A method for generating composite data according to claim 1, comprising the step of performing a denoising process on the noise data based on the command information, the topology information, and the target defect profile.

6. The step of preparing the noise data is: The composite data generation method according to claim 5, comprising the step of generating noise data by adding noise to at least one of the acquired data, the command information, and the topology information.

7. The step of performing the aforementioned denoising process is: A step of inputting the command information, topology information, and target defect profile into a neural network-based denoiser to predict noise at a specific time step, The method for generating composite data according to claim 5, comprising the step of denoising the noise data using the predicted noise.

8. The step of performing the aforementioned denoising process is: A step of inputting the command information and topology information into a neural network-based denoiser to predict the noise at a specific time step, The steps include adjusting the predicted noise based on the target defect profile, The method for generating composite data according to claim 5, comprising the step of denoising the noise data using the adjusted noise.

9. The step of performing the aforementioned denoising process is: The steps include generating a defect shape image in which the target defect profile is reflected in the source shape based on the acquired data and the target defect profile, A step of inputting the command information, topology information, and defect shape image into a neural network-based denoiser to predict noise at a specific time step, The method for generating composite data according to claim 5, comprising the step of denoising the noise data using the predicted noise.

10. The aforementioned target defect profile includes information about the type and weight of the defects, The step of generating the aforementioned composite data is: A method for generating composite data according to claim 1, comprising the step of generating a plurality of composite data while changing the type and the weight values.

11. The composite data generation method according to claim 10, wherein the plurality of composite data are used to construct a defect detection model.

12. The training process of the generative model is as follows: A step of generating a composite defect data sample from a source shape data sample according to a specified defect profile, The steps include: determining the defect profile of the synthetic defect data sample via a discriminator; A method for generating synthetic data according to claim 1, comprising the step of updating the parameters of the generative model using the loss based on the result of the discrimination.

13. The training process of the generative model is as follows: A step of generating a composite defect data sample from a source shape data sample according to a specified defect profile, The steps include: obtaining a target defect data sample corresponding to the specified defect profile; The steps include determining the similarity between the synthesized defect data sample and the target defect data sample via a discriminator, A method for generating synthetic data according to claim 1, comprising the step of updating the parameters of the generative model using the loss based on the result of the discrimination.

14. A method performed by at least one processor, Steps include obtaining source shape data and target shape profile, The steps include: extracting command information from the acquired data (the command information is information about the commands performed to generate the source shape); The steps include extracting topology information of the source shape from the acquired data, A method for generating composite data, comprising the step of using the target shape profile as a condition for a generative model, and generating composite data of a target shape in which the target shape profile is reflected in the source shape from the command information and the topology information.

15. One or more processors, Includes a memory for storing computer programs executed by one or more processors, The aforementioned computer program, The operation involves acquiring source geometry data and target defect profiles. The operation of extracting command information from the acquired data (the command information is information about the commands performed to generate the source shape), The operation of extracting topology information of the source shape from the acquired data, A synthetic data generation system including instructions for an operation to generate synthetic data of a defect shape in which the target defect profile is reflected in the source shape, using the target defect profile as a condition for a generative model, from the command information and the topology information.

16. The aforementioned generative model is a diffusion-based model, The operation to generate the aforementioned composite data is: The process of preparing noise data, A composite data generation system according to claim 15, comprising the operation of performing a denoising process on the noise data based on the command information, the topology information, and the target defect profile.

17. A non-transitory computer-readable recording medium, which stores instructions for causing the at least one processor to perform a synthetic data generation method when executed by the at least one processor, The aforementioned method for generating composite data is: The steps include obtaining source geometry data and target defect profile, The steps include: extracting command information from the acquired data (the command information is information about the commands performed to generate the source shape); The steps include extracting topology information of the source shape from the acquired data, A non-temporary computer-readable recording medium comprising the steps of: using the target defect profile as a condition for a generative model, generating composite data of a defect shape in which the target defect profile is reflected in the source shape from the command information and the topology information.

Images