JP7627235B2 - Manufacturing support system and manufacturing support method - Google Patents

Description

The present invention relates to a manufacturing support system and a manufacturing support method.

At product production sites, there is a huge amount of information related to each process of product design, process design, procurement, manufacturing, inspection, operation and maintenance, and this information is communicated between each process at the production site. It is expected that such information can be utilized to improve business efficiency and prevent product defects.

The issues in the information management process at production sites are broadly categorized into information accumulation, search, and utilization, and are described below.

The information accumulation process requires the accumulation of a huge amount of information for the entire production process. The challenge with this process is that due to the large amount of information between processes, the labor required by the person in charge to enter information is large, making it easy for errors to occur in the information entered. Furthermore, since errors in the entered information can have an impact on related processes, it is a very important challenge to improve the efficiency of information entry by the person in charge and reduce information deficiencies. Here, examples of input information include processing instruction information and inspection information. Currently, at sites where processing instructions and inspection work are carried out, the entered information is carefully checked by increasing the number of people in charge and the labor required to prevent information errors. However, this careful checking work creates a new challenge in that it increases labor costs, and there is a need to improve the accuracy and efficiency of information entry work.

In the information search process, it is necessary to instantly retrieve desired information from accumulated information. However, because the information accumulated at product production sites is diverse, there is no common standard for comparing data when searching for information, and there is an issue in that it is not easy to retrieve information for various purposes.

In the information utilization process, it is necessary to prevent accidents and streamline the entire information management process by utilizing past information. However, there is a problem that information utilization is not progressing very well because data collection in a data format suitable for information utilization has not progressed. For example, in the process design process, there are many similar cases in the past for processing instructions on drawings, and while it is possible to improve efficiency by utilizing similar information, currently the person in charge inputs the processing instructions while checking them as they are made. Therefore, if the system could appropriately identify similar cases and complement the work, it is expected that the work of the person in charge to input processing instructions could be simplified.

There are systems that can prevent human error and improve work efficiency by streamlining the information management process at product production sites; these are called Product Life Management (PLM) tools. Current PLM tools manage information by linking the information to be managed to the 3D-CAD information of the 3D design drawings that form the basis of the product. However, there is no mechanism for realizing data search and utilization of generic information linked to product design drawings, so it is not easy to utilize similar data from other products or past cases, and the scope of application of the information management process is severely limited, which is a problem. Due to these issues, the current situation is that there has been little progress in streamlining information management processes that utilize similar data.

Patent Document 1 discloses an invention of a device that manages information on welding conditions. In detail, the welding control device has a drawing storage unit that receives and stores drawing data of a processing target including one or more welds, a welding condition database that stores welding conditions previously determined for each combination of multiple types of welded members and multiple types of base materials, a weld extraction unit, and a welding condition generation unit, and the drawing data in the drawing storage unit includes data indicating the type of welded members constituting the welds of the processing target and data indicating the type of base material, the weld extraction unit extracts data indicating the type of welded members and data indicating the type of base material for each of the one or more welds from the drawing data, and the welding condition generation unit determines the welding conditions corresponding to the combination of the type of welded members and the type of base material read by the weld extraction unit for each of the one or more welds by referring to the information stored in the welding condition database.

JP 2020-182973 A

In the device described in Patent Document 1, the target structure is limited to the welded part. In addition, as the feature for identifying the welding information, a specific set of feature values, namely, the type of the welded member and the base material and the plate thickness, which are acquired in advance, is used to identify the information. The technical problem of Patent Document 1 is that since it is a system for identifying the desired welding conditions using a specific set of feature values, the use applications are severely limited. For example, the identification rules cannot be changed in response to the following changes (1) to (3): (1) change in the object of information management, (2) expansion of the identification range of welding conditions, (3) introduction of new feature values. In addition, in the method described in Patent Document 1, when it becomes necessary to change the identification rules due to a change in the number of identification objects or the nature of the acquired data, the data format of the feature values must be redefined each time. In particular, with regard to (1) change in the object of information management, the information of the product production process includes not only welding information, but also the material and shape of the screw, jig information, sintering processing information, adhesion information, friction stir welding information, inspection information, past defects, etc. In the information management method (information identification method) based on the data of a specific feature element as in Patent Document 1, a wide variety of information in the product production process cannot be managed. This means that the application of the system as an information management system is severely limited.

A manufacturing support system according to a first aspect of the present invention comprises an accumulation database storing three-dimensional design information including a plurality of partial structures, at least one of the plurality of partial structures being assigned attention information, an input unit that accepts data operations by a user on the three-dimensional design information, a data conversion unit that converts information of the partial structures into graph data, an extraction unit that extracts attention information corresponding to the input graph data, and a suggestion unit that inputs a selected structure, which is the partial structure selected based on an input by the user to the input unit, to the extraction unit and suggests to the user that a candidate attention information, which is the attention information extracted by the extraction unit, be assigned to the selected structure.
A manufacturing support method according to a second aspect of the present invention is a manufacturing support method executed by a computer having a database storing three-dimensional design information including a plurality of partial structures, at least one of the plurality of partial structures being assigned attention information, and includes an input step of accepting data operations by a user on the three-dimensional design information, a data conversion step of converting the information of the partial structures into graph data, an extraction step of determining a similarity between the plurality of partial structures based on the graph data, and a proposal step of inputting a selected structure, which is a partial structure selected based on an input by the user in the input step, into the extraction step, and proposing to the user to assign attention information candidate, which is the attention information extracted by the extraction step, to the selected structure.

According to the present invention, by using graph data, a variety of data can be utilized to support design.

Conceptual diagram showing the effects of the manufacturing support system Manufacturing support system hardware configuration diagram FIG. 1 shows information stored in a storage unit. Welding model set configuration diagram Diagram showing the information contained in a model set Functional block diagram of the manufacturing support system A diagram showing the correspondence between three-dimensional structures and graph data Flowchart showing the process of the model set generation unit Flowchart showing the process of the model set generation unit FIG. 1 is a diagram for explaining an example of a subgraph extraction process performed by a subset generation unit. Conceptual diagram of common subset for welding Conceptual diagram showing the process of the subset generation unit Diagram explaining the prediction model Diagram explaining the search model Diagram explaining the search model FIG. 1 is a diagram for explaining a first method of the input assistance function. FIG. 2 is a diagram for explaining a second method of the input assistance function. Diagram explaining the proposed changes feature Diagram explaining the inspection assistance function

-First embodiment-
A first embodiment of a manufacturing support system S will be described below with reference to FIGS.

(System Configuration)
1 is a conceptual diagram showing the effect of the manufacturing support system S. The manufacturing support system S supports the design process and the inspection process when the process until a certain product is completed is roughly divided into three steps, namely, design, manufacturing, and inspection. In the design process, the manufacturing support system S supports reducing the input burden on the designer and preventing mistakes, and in the inspection process, supports preventing inspection omissions. In the design process, the manufacturing support system S displays candidates for input information on a liquid crystal display D1 used by a designer U1, for example. A computer P1 that outputs images to the liquid crystal display D1 has a configuration corresponding to a known CAD engine C, and a configuration corresponding to the manufacturing support system S according to the present invention.

The manufacturing support system S may function as part of the CAD engine C, or as an add-on to the CAD engine C. The manufacturing support system S works in conjunction with the CAD engine C, and input operations by the designer U1 to the computer P1 are input to both the manufacturing support system S and the CAD engine C. However, the manufacturing support system S may accept the input operations by this designer U1 and provide the necessary information to the CAD engine C, or the CAD engine C may accept the input operations by this designer U1 and provide the necessary information to the manufacturing support system S.

During the inspection process, the manufacturing support system S displays, for example, a warning regarding the inspection on a head-up display D2 or liquid crystal display D3 worn by the inspector U2. The computer P2 that outputs images to the head-up display D2 or liquid crystal display D3 has a configuration equivalent to a known CAD engine C, and a configuration corresponding to the manufacturing support system S of the present invention. However, it is preferable to use wireless communication for the connection between the computer P2 and the head-up display D2. In the following, the designer U1 and the inspector U2 are collectively referred to as "users".

(Hardware configuration)
Fig. 2 is a hardware configuration diagram of the manufacturing support system S. The computer P1 and the computer P2 described in Fig. 1 are equipped with the hardware described below. The manufacturing support system S is equipped with a CPU 41 which is a central processing unit, a ROM 42 which is a read-only storage device, a RAM 43 which is a readable and writable storage device, an input/output device 44 which is a user interface, a communication device 45, and a storage unit 46.

The CPU 41 loads a program stored in the ROM 42 into the RAM 43 and executes it to perform various calculations, which will be described later. Instead of the combination of the CPU 41, ROM 42, and RAM 43, the manufacturing support system S may be realized by a rewritable logic circuit, such as a field programmable gate array (FPGA), or an application specific integrated circuit (ASIC). Also, instead of the combination of the CPU 41, ROM 42, and RAM 43, the calculation unit may be realized by a combination of different configurations, such as a combination of the CPU 41, ROM 42, RAM 43, and an FPGA.

The input/output device 44 is an input/output interface for information with an input device operated by the user and an image display device viewed by the user. The input/output device 44 includes an input interface that accepts information from a mouse or keyboard, and an output interface that outputs image information to the liquid crystal display D1, the head-up display D2, and the liquid crystal display D3. The input/output device 44 may be a device that supports wireless communication. The communication device 45 is a communication interface that can communicate with other devices. The storage unit 46 is a non-volatile storage device, such as a hard disk drive or semiconductor memory. Below, an overview of the data configuration and functional configuration will be explained, followed by a detailed explanation.

(Data Composition)
3 is a diagram showing information stored in the storage unit 46. The storage unit 46 stores a design accumulation database (hereinafter referred to as a "design accumulation DB") 20, a total attention information graph database (hereinafter referred to as a "total attention information graph DB") 20G, and a model set 200. The total attention information graph DB 20G is a collective name for a welding graph database (hereinafter referred to as a "welding graph DB") 21G, an adhesive graph database (hereinafter referred to as a "adhesive graph DB") 22G, a screw graph database (hereinafter referred to as a "screw graph DB") 23G, and an attention graph database (hereinafter referred to as a "attention graph DB") 24G.

Model set 200 is a collective term for welding model set 201, adhesive model set 202, screw model set 203, and caution model set 204.

The design accumulation DB20 stores 3D-CAD information of past designs made by the designer and 3D-CAD information of the current design. The past 3D-CAD information may be configured to be selectable, for example, it may include only past 3D-CAD information for products on the same production line, or only past 3D-CAD information for products with the same design rules. Structure information, which is information on each structure in 3D-CAD, includes shape information and attachment information. Shape information is information that can identify the shape of an object in three-dimensional space, and there is no particular limit to the method of expressing the information.

Here, shape information is divided into (1) shape information of a structure that can be converted into a graph shape (hereinafter, simply referred to as "shape information of a structure") and (2) shape auxiliary information of a structure that cannot be converted into a graph shape (hereinafter, simply referred to as "shape auxiliary information"). The former (1) "shape information of a structure" is information that is expressed by the relationship between nodes and edges, which is a graph shape, after conversion to graph data. The latter (2) shape auxiliary information is information that is converted into numerical values or numeric vectors that nodes and edges hold after conversion to graph data. Details will be described later in the (Graph Data) section, but the former (1) "shape information of a structure" is information that expresses the connections between the faces and edges of the structure, among the shape information of the structure, and the latter (2) "shape auxiliary information" is information that expresses the properties of the faces and edges. Therefore, the latter (2) "shape auxiliary information" includes information such as the length of the edge, the curvature of the edge, the area of the face, and the angle between the faces.

It can also be said that the design accumulation DB20 stores information on multiple substructures, which are parts of a structure. For example, a structure is a single product or part, and a substructure is a single part or a part of a part. However, a substructure may be a part of multiple parts or multiple parts themselves.

The attached information is classified into component information and attention information. Attention information is information that the manufacturing support system S supports in attaching information, and includes, for example, information about the connection between parts. The attached information can be applied as long as it is information linked to a structure on 3D-CAD. Information about the connection between parts is information about welding, adhesives, screws, jigs, joints, brazing, solder, hinges, springs, sealing materials, etc. used at the connection points between parts. The component information mainly includes information about materials (hereinafter referred to as "material information"). The material information can include the material type and name, the elastic constant which is a mechanical characteristic, various physical properties such as yield stress, tensile strength, and elongation related to the stress-strain curve, thermal conductivity, melting point, etc. In addition, as material information, for example, safety factors of various measured values defined at the time of design can also be treated as part of the component information, and there are no particular restrictions on setting the information as long as it is linked to a partial structure of the design drawing in the process preceding the information attachment.

In this embodiment, the type of connection point between parts is called the "connection type." There are no restrictions on the connection type, and it can include, for example, a connection method that will be newly used in the future. In this embodiment, for the sake of simplicity, there are three connection types: welding, adhesion, and screwing, and the information related to welding included in the additional information is called "welding information," the information related to adhesion included in the additional information is called "adhesive information," and the information related to screws included in the additional information is called "screw information."

The welding information may include all or part of the Welding Procedure Specification (hereinafter referred to as WPS) that shows welding information to the welding operator. The WPS includes joint structure, shape, base material, welding material, welding conditions (welding type, current, voltage, welding speed, number of welding layers, etc.), heat treatment, etc. The welding information may also include information on the performance test record (Process Qualification Record) that supports the WPS. In addition, the process parameters of the welding robot may be used as a type of welding information linked to the structure. This is particularly effective in sites where automation of construction work using welding robots has progressed. The adhesive information includes information on the adhesive material, whether or not forced drying is performed, the drying time, and the amount of adhesive to be applied. The screw information includes information on the base material of the screw, the surface treatment of the screw, the diameter and length of the pilot hole that forms the female screw, and the pitch of the screw.

Furthermore, in this embodiment, information used to call attention during the inspection process is included in the attached information as attention information. That is, the attached information in this embodiment includes connection information and attention information. Therefore, the attached information includes welding information, adhesive information, screw information, and attention information. In the following, welding information, adhesive information, screw information, and attention information are collectively referred to as "attention information."

The total attention information graph DB20G is the structural information of the items contained in the design accumulation DB20, i.e., the shape information and attached information, converted into graph data. The graph data will be described in more detail later.

The welding model set 201 includes a plurality of combinations of any welding information included in the design accumulation DB 20, graph data converted from nearby structural information to which the welding information is attached, and a welding search model or a welding prediction model. The bonding model set 202 includes a plurality of combinations of any bonding information included in the design accumulation DB 20, graph data converted from nearby structural information to which the bonding information is attached, and a bonding search model or a bonding prediction model. The screw model set 203 includes a plurality of combinations of any screw information included in the design accumulation DB 20, graph data showing the shape of the neighborhood to which the screw information is attached, and a screw search model or a screw prediction model. The attention model set 204 includes a plurality of combinations of any attention information included in the design accumulation DB 20, graph data showing the shape of the neighborhood to which the attention information is attached, and an attention search model or an attention prediction model.

The welding model set 201, adhesive model set 202, screw model set 203, and caution model set 204 have the same configuration, so the welding model set 201 will be described in detail later as a representative of these four.

Figure 4 is a configuration diagram of the welding model set 201. The welding model set 201 includes a first welding model set 201-1, a second welding model set 201-2, a third welding model set 201-3, etc. As shown in the lower part of Figure 4, the configuration of each model set is classified into two types. The first type is shown in the lower left of Figure 4, and includes a welding search model 171, a welding common subset 211, and a welding search subset 221. The second type is shown in the lower right of Figure 4, and includes a welding prediction model 161 and a welding common subset 211.

The common subset 210 is used to create and update both the welding search model 171 and the welding prediction model 161. The search subset 220 is not included in the second type because it is used for the welding search model 171 but not for the welding prediction model 161. However, because the common subset 210 and the search subset 220 are created separately for each purpose, the data in them may overlap.

5 is a diagram showing information included in the model set 200. Although the configuration of the welding model set 201 has been described in FIG. 4, the adhesive model set 202, the screw model set 203, and the attention model set 204 have the same configuration. Therefore, it can be said that the model set 200 contains the information shown in FIG. 5. The common subset 210 is a collective term for the welding common subset 211, the adhesive common subset 212, the screw common subset 213, and the attention common subset 214. The common subset 210 is used for optimization and verification evaluation of the prediction model and the search model. Since the common subset 210 is used for machine learning as described later, it can also be called a "dataset". The search subset 220 is a collective term for the welding search subset 221, the adhesive search subset 222, the screw search subset 223, and the attention search subset 224. The search subset 220 is a database that is the search target of the search model as described later.

The welding common subset 211 and the welding search subset 221 include a portion of the welding model set 201. Specifically, the welding common subset 211 and the welding search subset 221 include several sets of combinations of welding information and graph data included in the welding model set 201. Which sets are included will be described later. The bonding common subset 212 and the bonding search subset 222 include a portion of the bonding model set 202. Specifically, the bonding common subset 212 and the bonding search subset 222 include several sets of combinations of bonding information and graph data included in the bonding model set 202. Which sets are included will be described later.

The screw common subset 213 and the screw search subset 223 include a portion of the screw model set 203. Specifically, the screw common subset 213 and the screw search subset 223 include several combinations of screw information and graph data included in the screw model set 203. Which combinations are included will be described later. The attention common subset 214 and the attention search subset 224 include a portion of the attention model set 204. Specifically, the attention common subset 214 and the attention search subset 224 include several combinations of attention information and graph data included in the attention model set 204. Which combinations are included will be described later.

The relationship between the welding common subset 211 and the welding search subset 221 and the welding model set 201 is the same as the relationship between the adhesive common subset 212 and the adhesive search subset 222 and the adhesive model set 202, the relationship between the screw common subset 213 and the screw search subset 223 and the screw model set 203, and the relationship between the attention common subset 214 and the attention search subset 224 and the attention model set 204. Therefore, the relationship between the welding common subset 211 and the welding search subset 221 and the welding model set 201 will be described in detail below as a representative of these.

(Functional configuration)
6 is a diagram showing the functions of the manufacturing support system S as functional blocks. Each function shown in FIG. 6 is realized by a combination of a CPU 41, a ROM 42, and a RAM 43. The manufacturing support system S includes a model set generation unit 12, a subset generation unit 13, an input unit 14, an extraction unit 15, a prediction model 16, a search model 17, a proposal unit 18, and a model update unit 19 as its functions. The prediction model 16 is a collective term for a welding prediction model 161, an adhesive prediction model 162, a screw prediction model 163, and an attention prediction model 164. The search model 17 is a collective term for a welding search model 171, an adhesive search model 172, a screw search model 173, and an attention search model 174.

The model set generation unit 12 generates the model set 200 using information stored in the design accumulation DB 20. The model set generation unit 12 operates the subset generation unit 13 as necessary. As described below, the design accumulation DB 20 is updated based on user operation. When the design accumulation DB 20 is updated, or when an update is instructed from the input unit 14 at the user's discretion, the model set generation unit 12 can update the model set 200 and a series of databases, prediction models 16, and search models 17 linked to the model set 200 as necessary.

A specific procedure for creating the model set 200 will be described. First, the model set generation unit 12 converts the structural information of the object contained in the design accumulation DB 20 into graph data, which will be described later, to create the all-attention-information graph DB 20G. The model set generation unit 12 then combines the graph data and the attached information contained in the all-attention-information graph DB 20G to generate the model set 200. At this time, depending on the type of attention information to be combined, it is determined whether to add the information to the welding model set 201, the adhesive model set 202, the screw model set 203, or the caution model set 204. The model set generation unit 12 converts the structural information of the object into graph data, so it can also be called a "data conversion unit."

The subset generation unit 13 generates the common subset 210 and the search subset 220 using the model set 200. Specific correspondences in the operation of the subset generation unit 13 are shown below. The subset generation unit 13 generates the common subset 211 for welding and the search subset 221 for welding using the model set 201 for welding. The subset generation unit 13 generates the common subset 212 for bonding and the search subset 222 for bonding using the model set 202 for adhesion. The subset generation unit 13 generates the common subset 213 for screws and the search subset 223 for screws using the model set 203 for screws. The subset generation unit 13 generates the common subset 214 for attention and the search subset 224 for attention using the model set 204 for attention. The subset generation unit 13 updates the common subset 210 and the search subset 220 when the model set 200 is updated or when an update is instructed from the input unit 14 at the user's discretion.

The input unit 14 accepts input operations from the user using the input/output device 44, and outputs information on the input operations by the user to the suggestion unit 18.

The extraction unit 15 extracts attention information corresponding to the input structural information of the object. Hereinafter, the attention information extracted by the extraction unit 15 is referred to as "attention information candidate". Specifically, the extraction unit 15 first converts the input structural information of the object into graph data, inputs it to an appropriate type of prediction model 16 or search model 17, and outputs the added information candidate output by the prediction model 16 or search model 17. For example, if the type of attention information is "welding information", the extraction unit 15 inputs the graph data to the welding prediction model 161 or welding search model 171.

The prediction model 16 calculates and outputs one or more pieces of attention information corresponding to the input graph data. The search model 17 searches the search subset 220 for one or more pieces of graph data similar to the input graph data, and outputs attachment information corresponding to the similar graph data. Specifically, the welding search model 171 uses the welding search subset 221 for searches, the adhesive search model 172 uses the adhesive search subset 222 for searches, the screw search model 173 uses the screw search subset 223 for searches, and the attention search model 174 uses the attention search subset 224 for searches.

The suggestion unit 18 identifies a selected structure based on an input to the input unit 14 by the user, inputs the selected structure to the extraction unit 15, and suggests to the user that the extraction unit 15 assigns the candidate attention information, which is the attention information extracted by the selection unit 15, to the selected structure. In this embodiment, the partial structure selected based on the input to the input unit 14 by the user is called a "selected structure". The selected structure is a partial structure selected by the user, or a partial structure similar to the partial structure selected by the user. The suggestion unit 18 may determine which of the partial structure selected by the user or the partial structure similar to the partial structure selected by the user is to be the selected structure according to the operation mode of the manufacturing support system S, or may determine which of the above is to be the selected structure according to the size of the area specified by the user to extract the selected structure. The suggestion unit 18 also updates the design accumulation DB 20 based on the user's response when the suggestion unit 18 suggests the candidate attention information to the user.

The model update unit 19 updates the prediction model 16 and the search model 17 by referring to the search subset 220. Specifically, the model update unit 19 updates the welding prediction model 161 and the welding search model 171 by referring to the welding search subset 221. The model update unit 19 updates the adhesive prediction model 162 and the adhesive search model 172 by referring to the adhesive search subset 222. The model update unit 19 updates the screw prediction model 163 and the screw search model 173 by referring to the screw search subset 223. The model update unit 19 updates the attention prediction model 164 and the attention search model 174 by referring to the attention search subset 224. Note that in this embodiment, the model update unit 19 updates the prediction model 16 and the search model 17, but the data scientist may update the prediction model 16 and the search model 17.

(Graph data)
Before describing the welding model set 201 and the like, graph data will be described. In this embodiment, graph data is used to abstract parts to which additional information is added and handle them in a common format. The graph data can represent shape information indicating connections between physical lines and faces that are components of a structure in a graph form. In addition, shape auxiliary information and additional information of a structure, which are pieces of shape information that cannot be represented in a graph form, can be held on a graph in the form of a feature vector.

Figure 7 is a diagram showing the correspondence between a three-dimensional structure and graph data. The upper part of Figure 7 is a diagram showing the three-dimensional structure as a Boundary Representation (Brep). In Brep, an object in three-dimensional space is represented using physical vertices, which are the vertices of the three-dimensional structure, physical edges formed by the physical vertices, physical surfaces formed by the physical edges, and solids, which are solid 3D structures surrounded by physical surfaces. Furthermore, an object made up of one or more solids is considered a part, and a collection of parts makes up a single product.

As shown in the lower part of Figure 7, graph data is a data structure made up of vertices V called nodes and lines E that show the connections between the vertices called edges. In this embodiment, the vertices and edges on the 3D-CAD are called "physical vertices" and "physical edges", and the vertices and edges of the graph structure are called "nodes" and "edges". In addition, each node and edge can hold multiple feature amounts in vector format. The vectors are used as feature vectors to express the characteristics of each node and each edge.

The components of the Brep structure and the components of the graph data can be associated, for example, with physical surfaces as nodes and physical edges as edges. The physical surfaces V701-V706 shown in the upper part of Figure 7 correspond to the nodes V701-V706 of the graph structure shown in the lower part of Figure 7. Similarly, each physical edge of the three-dimensional structure corresponds to edges E707-E718.

The edge E of the graph data can hold auxiliary shape information such as the length of the edge, the curvature of the edge, and the angle between the faces that share the edge, while the node V can hold auxiliary shape information such as the area and curvature of the face. This information is held as the feature amount of each component of the vector. Such a vector is called a feature amount vector. If the auxiliary shape information is not a constant value on the physical edge or surface corresponding to the edge E or node V, the average and variance, which are characteristics of the distribution of values indicating this information, can be held in the feature amount vector. In addition, when converting information on the components that make up an item, which is one of the additional information, into graph data, it is desirable to hold material information such as the material type and its physical properties as the feature amount vector of the node or edge. Converting into graph data including the component information is particularly effective when treating the joining conditions as the information of interest, because the joining conditions depend on the material properties and material type.

The handling of categorical variables in this embodiment will be described. A categorical variable refers to a variable of information that cannot be directly expressed by a numerical value, such as a material type. When handling a variable of information that cannot be directly expressed by a numerical value, it is necessary to perform a quantification process of the categorical variable as a preprocessing. For example, by performing one-hot encoding process using a material type as an example of a categorical variable, the categorical variable of each data can be converted into a vector notation. The vector obtained after the quantification process of the categorical variable indicates the material type of each data by setting a specific vector component corresponding to a certain material type to 1 and the remaining vector components to 0, and can be retained as part of the feature vector of the node or edge. However, in the one-hot encoding process, if the dimension of the feature vector finally obtained becomes too large, the accuracy of prediction and search may decrease. In such a case, it is desirable to perform label encoding, which is an operation that directly converts each categorical variable into a numerical value. There are no particular restrictions on the method of converting the categorical variable as long as it is within the range of the notation method that can be retained in the feature vector format for the node or edge, and a general conversion method can be used instead.

As an example of secondarily generated features, consider the features of a structure joined at the interface (physical surface) of dissimilar materials. Faces and edges shared by two structures are at the boundary between structures with different types of material information, and therefore cannot be defined by the features of a single type of material. Therefore, the average, variance, etc. of the features of both materials between the interface can be stored as secondarily generated features in the components of vectors of nodes and edges. In addition, the values of each component of the feature vectors stored in nodes and edges are generally graph data containing features obtained secondarily by performing a standardization process on the model set 200.

(Model set generation unit)
8 and 9 are flowcharts showing the processing of the model set generation unit 12. As described below, the model set generation unit 12 generates a welding model set 201, a bonding model set 202, a screw model set 203, and a caution model set 204 by using the attached information contained in the design accumulation DB 20 and the total attention information graph DB 20G.

First, in step S301, the model set generation unit 12 determines the target attention information (hereinafter referred to as "target attention information"). A model set corresponding to the target attention information determined here is generated, so in order to generate all four of the welding model set 201, adhesive model set 202, screw model set 203, and caution model set 204, the processing from step S302 onwards is executed four times while changing the attention information in step S301. Here, the explanation will continue using as a representative example the case where "welding information" is determined as the target attention information in step S301.

In the next step S302, the model set generation unit 12 generates a database (graph DB) containing graph data corresponding to the target attention information determined in step S301, based on the design accumulation DB 20. For example, if the target attention information is "welding information", a welding graph DB 21G is generated. Each graph DB has different information in the nodes and edges of the graph data. Of all the structure information, only the structure information to which each target attention information is assigned is converted into graph data and converted into each database 21G to 24G.

In step S303, the model set generation unit 12 selects a model optimization method. Specifically, the model set generation unit 12 selects a machine learning method and whether to create a search model or a prediction model. This selection is made, for example, in an order specified in advance by the user. In this case, the user specifies in advance options for when step S303 is executed for the first time, options for when step S303 is executed for the second time after returning from step S309, and the like. In the following step S304, the model set generation unit 12 selects one from the graph extraction algorithm candidates. The selection here may be made in an order specified in advance by the user, or may be made randomly. However, when step S304 is executed after returning from step S308, an algorithm that has not been selected so far is selected.

In the following step S305, the model set generation unit 12 generates a model set according to the extraction algorithm selected in step S304. In the following step S306, the model set generation unit 12 extracts a common subset for optimizing the prediction model and the search model using the model set generated in step S305.

In the next step S307, the model set generation unit 12 optimizes the model using the common subset and verifies the prediction accuracy. A test dataset (not shown) is used to verify the prediction accuracy. The test dataset is data in the same input/output format as the common subset, and is a collection of data that was created in the same flow as the common subset but is not included in the common subset. In the next step S308, the model set generation unit 12 determines whether or not to end the consideration of the candidate extraction algorithms. If the model set generation unit 12 determines that the consideration can be ended, it proceeds to step S309, and if it determines that the consideration should be continued, it returns to step S304.

In step S309, the model set generation unit 12 judges whether the prediction accuracy of the generated model is sufficient. In this step, the calculation speed may be further evaluated to judge whether the performance of the prediction model is sufficient. If the model set generation unit 12 judges that the prediction accuracy is sufficient, it proceeds to step S310 in FIG. 9 via circled A, and if it judges that the prediction accuracy is insufficient, it returns to step S303. In the operation of step S309, the candidate models obtained by executing the operations from S303 to S308 with various parameters may be evaluated collectively in S309 to select the best prediction model. In step S310, the model set generation unit 12 judges whether the generated model is a search model, and if it is judged to be a search model, it proceeds to step S311, and if it is not a search model, i.e., it is a prediction model, it proceeds to step S312.

In step S311, the model set generation unit 12 selects a search subset to be searched by the search model, and proceeds to step S312. In this case, the model set generation unit 12 selects a search model set based on the search speed and search accuracy for a test data set that is configured in the same format as the model set. In step S312, the model set generation unit 12 selects a model, extraction algorithm, model set, and each subset for the selected attention information. In the following step S313, the model set generation unit 12 determines whether or not model creation and updating for other attention information is necessary, and if it is determined that it is necessary, returns to step S301 via circled B, and if it is determined that it is not necessary, ends the processing shown in Figures 8 and 9.

(Subset Generation Unit)
Fig. 10 is a diagram for explaining an example of the subgraph extraction process by the subset generation unit 13. The graph data shown in the upper part of Fig. 10 shows graph data of parts A and B that are joined by welding, where the nodes correspond to the physical faces of the 3D structure and the edges correspond to the physical sides of the 3D structure. Parts A and B are joined by edge E indicated by weld lines 801 and 802. Here, it is assumed that the information on the weld line 801 was the attention information selected in step S302.

In this case, in the second proximity extraction algorithm, which is an example of the extraction algorithm, the graph data including the nodes adjacent to the adjacent node of the position related to the attention information selected in step S302 by the edge, i.e., the second proximity node, is extracted as the subgraph data. By this extraction process, the subgraph data shown in the lower part of FIG. 10 is obtained. In addition, at this time, the elements of some dimensions of the feature vectors F _V,total and F _E,total held by the nodes and edges of the graph data are extracted as F _V,partial and F _E,partial . Here, the graph data including the second proximity node is extracted as the subgraph data, but it may be extended to a third proximity extraction algorithm that further includes adjacent nodes in the subgraph data. This may be further extended to adopt the nth proximity extraction algorithm expressed using the natural number n. Similarly, a partial feature vector including the natural number n and m elements of some dimensions of the feature vector is selected. The selection of the nth proximity and the combination of the number of m elements and the element type is determined by trial and error from the evaluation of the prediction accuracy according to the processing of the flowcharts shown in FIG. 8 and FIG. For example, as n and m increase, the degree of freedom of the model created for the common subset increases, so at a certain point, the prediction accuracy for the test data will no longer improve with increasing n and m. Therefore, as an example of a method for determining the upper limit of the natural n and m to be considered, a method of stopping consideration of higher n and m is effective.

In addition, if the information of interest is linked to a solid rather than a physical edge, an algorithm may be used that extracts a graph structure including the nearest edge and face of the solid up to the n-th nearest node. This concludes the explanation of the subgraph data extraction algorithm.

11 is a conceptual diagram of the welding common subset 211. The welding common subset 211 includes a plurality of data (j). "j" is an index indicating the data in the welding common subset 211. Data (j) 1004A is converted from the original data (j) 1000. The original data includes structure information (j) that is the source of input information to the prediction model 16, and welding information (j) corresponding to output information to the prediction model 16. A vector t j (1006) is obtained, which is a vector for identifying the welding information obtained by performing one-hot encoding processing on the graph data (j) 1005 obtained by the processing of step _S305 and the welding information.

The subset generation unit 13 generates a common subset 210 and a search subset 220 from each model set. Specifically, the subset generation unit 13 uses the welding model set 201 to create a welding common subset 211 and a welding search subset 221. The subset generation unit 13 uses the adhesive model set 202 to create a bonding common subset 212 and an adhesive search subset 222. The subset generation unit 13 uses the screw model set 203 to create a screw common subset 213 and a screw search subset 223. The subset generation unit 13 uses the attention model set 204 to create an attention common subset 214 and an attention search subset 224.

A method for creating a welding common subset 211 and a welding search subset 221 will be shown using the welding model set 201 as an example. The subset generation unit 13 extracts multiple pairs of graph data and attention information included in the welding common subset and the welding search subset from the attention information paired with the graph data included in the welding model set 201, for each subset.

This extraction process can be considered as a sampling process of the data contained in the model set. After pointing out the necessity of a sampling method for creating a common welding subset and a search subset and the problems that need to be solved, an example of a sampling method is explained using Figure 12.

First, sampling methods will be explained using a common subset as an example when used to optimize a predictive model. A common subset is used for learning when optimizing a predictive model using machine learning methods. Therefore, if sampling when creating a common subset is biased towards specific data, problems may arise in which the predictive accuracy of the predictive model for new data decreases, or the predictive accuracy of each class becomes biased. To prevent these problems, it is necessary to correct the bias in the learning data by creating a common subset using appropriate sampling processing. Here, the classification result of the attention information is called label data, and the type of label data is called a class. Attention information obtained at product production sites often has biased numbers of data for each class and their distribution, even within the same product.

Therefore, there are two problems that sampling can cause that need to be solved. Problem (1): In general random sampling, label data belonging to a particular class is excessively sampled, resulting in the generation of a prediction model that prioritizes the prediction accuracy of classes with a large amount of data at the expense of the prediction accuracy of classes with a small amount of data. Problem (2): By repeatedly sampling homogeneous data in an excessive manner, not only does the prediction accuracy not improve, but the computational costs of training also increase. One example of a solution to these problems is to randomly sample each class so that the number of data for each class is roughly the same.

The effect of sampling is explained using the upper diagram in Figure 12. The upper diagram shows a schematic of the distribution of subgraph data of the model set in the data space, and the differences in each figure indicate the differences in the classes of attention information corresponding to the subgraph data. In particular, the data of the classes indicated by triangles is much more abundant than the other classes, and problem (1) occurs when all the data indicated by triangles is used. Therefore, only sampled data, i.e., extracted data, is used. This reduces the amount of data and at the same time avoids unnecessarily increasing the computational costs of learning, addressing problem (2).

Here is an example of a solution using another sampling method. In the solution of using only the extracted data mentioned above, the bias in the predictive model may become large due to the bias in random sampling for each class. Therefore, it is also effective to combine the sampling method for each class mentioned above with ensemble learning. A further solution is to create multiple predictive models using multiple common subsets obtained from multiple samplings, and then treat the average of the prediction results of each of these predictive models as the output of a new predictive model.

The purpose of the sampling method mentioned above is to reduce the learning cost of the predictive model and improve the prediction accuracy for unknown data. Therefore, any general sampling method can be used as long as a common subset that corrects the bias in the dataset and satisfies the desired evaluation indicators of computational cost and prediction accuracy is sampled. Note that while the predictive model has been described as a classification model, a similar sampling method can also be used to extract a common subset for a regression model. A similar sampling method can also be used to extract a common subset when using a common subset to optimize a search model.

Next, the necessity of a specific sampling process in creating a search subset and the problem to be solved are described, and an example of a sampling method as a solution is given. As with the creation of a common subset, the search subset needs to reduce the computational cost of the search model and the bias in search accuracy for unknown data of each class. In addition, since the search model is greatly related to the speed at which information is presented to the user by the suggestion unit 18, the number of data in the search subset to be searched needs to be carefully selected in order to maintain a sufficient search speed. Therefore, in addition to the problem related to the bias of problem data when creating a common subset, a unique problem may also arise when creating a search subset. Since data is searched and acquired according to the similarity of graph data in the subset during a search, there is a problem that the search results are significantly affected by outliers in the search subset. To solve this problem, it is necessary to remove the outliers by sampling. For example, an example of a sampling method is given in the reference [D.L. Wilson, IEEE Trans. Syst. Man. Cybern. SMC-2, 408 (1972).]. By using such a method, outlier data of different classes of peripheral data can be excluded from sampling.

An example of sampling to remove outliers will be explained using the lower diagram in Figure 12. In Figure 12, it is preferable to remove outlier data 1201, indicated by triangles in the rectangular class area, and data indicated by reference symbol 1202 that is mixed near the boundaries of the class areas. This makes it possible to stabilize search results. The method in the above reference is a method for removing outlier data.

The sampling process to remove outliers in the above-mentioned references requires the distance between data in feature vector space, but it is generally difficult to define the distance between graph data in feature space. Furthermore, it is more desirable to use the sampling method mentioned in the common subset in combination to correct the bias in the number of data between classes. Therefore, it is desirable to use it in combination with the method of embedding graph data in feature vector space, which will be described later in the retrieval model section. With this embedding method, graph data can be sampled based on the obtained distance in feature space.

The evaluation of whether or not sampling has been performed appropriately is based on the search response speed and search accuracy for the search subset by the search model. As long as this evaluation is satisfied, there are no particular restrictions on the sampling method, and multiple combinations may be used. As an example of a sampling method, a method of removing data from closely related classes is also effective (see Ivan Tomek. Two modifications of cnn. IEEE Trans. Systems, Man and Cybernetics, 6:769-772, 1976.). This method is particularly effective when removing data where data from different classes is mixed at the boundaries of each class, as shown by reference numeral 1202 in FIG. 12.

The subset generation unit 13 obtains a common subset for welding 211 using the sampling method for the common subset described above based on the welding information paired with each piece of graph data. The subset generation unit 13 obtains a search subset for welding 221 using the sampling method for the search subset described above based on the welding information paired with each piece of graph data.

The set of noteworthy information paired with the graph data obtained according to the sampling process of the subset generation unit 13 may be included in both the common subset and the search subset, or may be included in only one of the subsets.

(Prediction model)
FIG. 13 is a diagram for explaining the prediction model 16. Input data 903 to the prediction model 16 is information created from the structure information (i) 901 in FIG. 13. Here, i is an index for distinguishing each piece of structure information. In FIG. 13, as an example, the information of interest is shown as welding information, and the structure information (i) 901 is shown as a structure in the vicinity of a weld line. Here, since the structure information (i) cannot be directly input to the prediction model 16, it is converted into sub-graph data (i) 903. The conversion process from this structure information to sub-graph data is as described above. The sub-graph data 903 is obtained according to the extraction algorithm selected by the process shown in the flowcharts of FIG. 8 and FIG. 9.

The prediction model 16 receives subgraph data (i) 903 as a direct input and obtains output information 905. Information obtained as a direct output from the prediction model 16 is an output vector t _i 906 indicating the probability of each piece of assigned information. Each component of t _i indicates the probability of assigning each piece of candidate attention information 907-909 to the structure information (i). Therefore, the sum of each component of t _i is 1. As shown in FIG. 13, each component, component 1, component 2, component 3, ... indicates the probability of assigning candidate welding information (1), candidate welding information (2), candidate welding information (3), ..., respectively. In the example of FIG. 10, since the probability of component 3 is the highest, welding information (3) (907) is proposed as the assigned information to the structure information (i) from the output result of the prediction model 16.

When displaying attention information using the output vector t _i , candidates for attention information are listed in order of the probability indicated by each component of t _i . In addition, attention information with the highest attachment probability may be provisionally attached to structure information as in attention information 910. Note that here, a classifier that estimates appropriate attachment information from a group of attachment information candidates when structure information is input is used as the prediction model 16. However, the prediction model 16 is not limited to a classifier, and a regression model that predicts attention information in the form of a numerical value or a numerical vector may be used as the prediction model 16. However, the prediction model 16 is characterized in that it receives graph data converted from structure information as an input, and the output of the prediction model 16 indicates information to be attached to the structure.

A machine learning method for creating the prediction model 16 will be described. The machine learning method used in the present invention can create a prediction model 16 such as a classifier using graph data as input. For example, among the methods that use graph data as input, it is desirable to use a method classified as a GNN (Graph Neural Network) that can create a prediction model 16 with high generalization performance. In particular, it is desirable to use a GCN (Graph Convolutional Network) that reflects local information around the nodes of the graph data in the prediction model 16. Each process in the GNN, including the propagation process, sampling process, and pooling process, can be examined by trial and error for the common subset 210, and there are no particular restrictions as long as a process with high prediction accuracy is selected.

In optimizing the classification model (a type of prediction model 16) illustrated in FIG. 10, cross entropy error is often used as a loss function, but there are no particular limitations as long as it is possible to optimize which candidate of attention information is desirable for the input of structural information with respect to the common subset 210. Similarly, if the prediction model 16 is a regression model, a general loss function such as sum of squares error is used, but is not limited to this. In addition to GNN, a graph kernel that defines the similarity between graph data may be used as an example of a machine learning method that uses graph data as input.

Types of graph kernels include the random walk kernel [H. Kashima, K. Tsuda, and A. Inokuchi, in Proc. Twent. Int. Conf. Int. Conf. Mach. Learn. (AAAI Press, 2003), pp. 321-328.], the shortest kernel [K.M. Borgwardt and H.-P. Kriegel, in Proc. Fifth IEEE Int. Conf. Data Min. (IEEE Computer Society, USA, 2005), pp. 74-81.], and the subgraph matching kernel [N. Kriege and P. Mutzel, in Proc. 29th Int. Conference Int. Conf. Mach. Learn. (Omnipress, Madison, WI, USA, 2012), pp. 291-298.]. Machine learning methods that can be used in combination with the graph kernel include SVM (Support Vector Machine), kernel perceptron, Gaussian Process, linear discriminant analysis, etc., but there are no particular restrictions as long as it is an algorithm for creating a predictive model 16 to be used in combination with the kernel method.

However, a drawback of using a graph kernel to create a predictive model 16 is that the computational cost of the graph kernel can become a bottleneck when creating and using the predictive model 16. Therefore, although it is necessary to implement some ingenuity in the kernel calculation, such as limiting the amount of training data used, the graph kernel may be used to create the predictive model 16 as long as the predictive model 16 can achieve a computation speed that does not cause dissatisfaction to users.

(Search model)
14 and 15 are diagrams for explaining the search model 17. Input data 1103 to the search model 17 is subgraph data (i) 1103 based on structure information (i) 1102. Here, i is an index for distinguishing each piece of structure information. The structure information (i) is converted into subgraph data (i) 1103 and then input to the search model 17. The series of processes for creating the subgraph data (i) 1103 shown here is the same as the example of the prediction model 16 shown in FIG. 13, and a search model is obtained by optimizing the model for the input/output results for the common subset 210.

The purpose of the search model 17 is to identify subgraph data similar to the subgraph data (i) that is the search query from among the graph data (j) included in the database 1105. Here, the subgraph data (j) is data converted from the structure information (j), similar to the subgraph data (i). j is an index that identifies the subgraph data (j) and the original structure information (j) in the database, and is included in the search subset 220.

The search model 17 calculates similarity S _ij between the graph data (i) 1105 to be compared and each graph data (j) in the search subset 220. As shown by 1107A, 1108A, and 1109A in Fig. 14, the similarities calculated for subgraph data (1), subgraph data (2), and subgraph data (j) are indicated by similarity S _i1 , similarity S _i2 , ..., similarity S _ij . Also, attention (1) 1107B, attention information (2) 1108B, ..., attention information (j) 1109B are associated in advance with each graph data.

The extraction unit 15 uses the similarity S _ij as an evaluation index to extract subgraph data (j) 1109A having a high similarity to the subgraph data (i). In the example shown in FIG. 14 and FIG. 15, attention information (j) 1109B of the original information (j) linked to the subgraph data (j) is extracted because of its large similarity S _ij . Since the structure information (i) and the structure information (j) are similar through the calculation result of the similarity S _ij , attention information (j) 1109B is linked to the structure information (i) to estimate attention information (i) 1200. By the operation using the search model 17 shown here, attention information (i) 1200 to be linked to the structure information 1103 of the input data (i) 1103 can be obtained. In addition, by acquiring a plurality of promising attention information (j) from similar data (j) having a high similarity in order of similarity S _i1 , these can be treated as candidates for information to be added to the structure information (i).

Note that in FIG. 14, for the sake of abstraction, a search model 17 and a search subset 220 are used, but in practice, a specific search model 17 and a search subset 220 are used according to the information of interest. For example, when welding information is the information of interest, a search model 171 for welding and a search subset 221 for welding are used.

The following points should be noted when calculating similarity using the search model 17. In theory, it is possible to calculate similarity for all data in a model set, but in actual operation, it is difficult to calculate similarity S _ij for all the large amount of data in a model set within a desired time. Therefore, as will be explained next, it is desirable to adopt a method for lowering the calculation cost of search processing by similarity calculation and reducing the required time.

For example, the calculation cost can be reduced by calculating the similarity of a data set (j) of a model set only for a search subset obtained by sampling a model set based on attention information. As another method, the search subset can be divided into hierarchical groups, and the similarity S _ij can be calculated for representative data of each group for each hierarchy, so that the same attention information can be linked to the representative structure information (j) of the group to which it ultimately belongs. In this latter method, the similarity calculation is performed for each hierarchy like a decision tree, so that the number of times the overall similarity calculation is performed can be reduced. In this way, various known search algorithms can be used as long as the search method is characterized in that the similarity S _ij of graph data is used as an evaluation index.

A method for creating the search model 17 will be described from two perspectives: the common subset 210 and the machine learning method. In the search model 17, the similarity S _ij between the subgraph data (i) of the input structure information (i) and the subgraph data of the search target structure information (j) is preferably an evaluation index that indicates high similarity when the attention information is the same in the data set (j) and low similarity when the attention information is different. In order to create the search model 17 _{that calculates S ij that} satisfies the requirements, the method for calculating the similarity can be optimized in advance using the common subset 210 so that data can be distinguished for each attention information label. The optimization method follows the machine learning method described later. Here, the common subset 210 is composed of a set of subgraph data (j) and the corresponding attention information (j), and has the same format as the data set (j) used to optimize the prediction model 16.

A method for creating a search model 17 that performs similarity calculations according to each machine learning method will be described. The machine learning method must be able to evaluate the similarity S _ij of each graph data by using graph data as input, and is preferably determined using a graph kernel or a GNN (graph neural network). When using a graph kernel, the distance between graph data (i) and graph data (j) can be calculated from the definition of the graph kernel function. This calculated value is treated as the similarity between graph data (i) and graph data (j), and data with a closer distance is treated as a higher similarity. Although the similarity between two graph data can be directly calculated by the graph kernel, the distance in a secondary feature vector space obtained by clustering or dimensional reduction using the search subset 220 and a kernel function may be calculated as the similarity. The secondary feature can be created by using the clustering or dimensional reduction method of a general kernel method in the same way.

The definition of distance can be cosine distance, Mahalanobis distance, Euclidean distance, Hausdorff distance, etc. in feature space, but there are no particular restrictions as long as the desired information can be added by using it as a similarity measure. However, one thing to note about calculations using graph kernels is that the information to be extracted from graph data changes depending on the attention information (label data) added to the graph data, so the graph kernel must be designed depending on the type of attention information. For this reason, it is not easy to extract information from graph data in a general purpose manner, and there is a drawback in that the kernel must be redesigned to improve prediction and search models.

On the other hand, when using GNN, which is a method suitable for constructing a model with high generalization performance, a search model 17 for calculating similarity is constructed. To optimize the search model, the method for calculating the similarity between graph data can be optimized through a process of optimizing the parameters of the GNN using the data set (j) of the subgraph data of the common subset 210 and the information of interest, so as to improve the accuracy of identifying the subgraph data of the common subset 210 as the information of interest. For this reason, it is particularly desirable to use GNN as a method for creating the search model 17.

Among the search model creation methods using GNNs, for example, a model for calculating the similarity between graph data can be created by using the method described in the literature "P. Riba, A. Fischer, J. Llados, and A. Fornns, in 2018 24th Int. Conf. Pattern Recognit. (2018), pp. 2239-2244." In this case, the similarity scale can be optimized by calculating subgraph data to which label data of the search subset 220 is assigned as a high similarity, and calculating subgraph data to which different label data is assigned as a low similarity.

In addition, the method of the literature "Y. Li, C. Gu, T. Dullien, O. Vinyals, and P. Kohli, CoRR abs/1904.12787, (2019)" makes it possible to embed graph data of the same label data at close distances in the feature vector space through optimization based on label data. In addition, the similarity between graph data can be determined based on the distance in this feature vector space. As in the literature, the search model 17 that calculates the similarity can be optimized using a loss function based on the label data of the common subset 210 by comparing the similarity between graphs between two bodies and three bodies. Here, too, the similarity calculation method (search model 17) is optimized so that graph data with common attention information (label data) have high similarity and graph data with different attachment information have low similarity.

The method described in the aforementioned document can map graph data to a general feature vector space based on optimization for the common subset 210. This has the advantage that the search model can be used in combination with a general search algorithm that targets feature vectors corresponding to each graph data. Furthermore, the method described in the aforementioned document can be used in combination with a clustering process that groups feature vector data, or a general sampling method that focuses on the data distribution in feature vector space, so that the search subset 220 to be searched can be easily created by obtaining it using each of the aforementioned methods as appropriate.

The machine learning technique for creating the search model 17 is not limited to the above example, as long as it can calculate the similarity S _ij between the subgraph data (i) and the subgraph data (j) in the search subset 220, which is the database to be searched, by inputting the subgraph data (i). However, it is preferable that the machine learning technique for creating the search model 17 can optimize the method of calculating the similarity according to the common subset 210.

The prediction model 16 and the search model 17 are distinguished based on their intended use. The prediction model 16 is created using the common subset 210, and therefore has advance knowledge of candidates for attention information (label data), making it desirable in cases where one of a limited range of label data is to be assigned to new structure information (i). This is because the prediction model 16 predicts attention information to be assigned to the input graph data (i) from the range of label data learned in advance.

On the other hand, the search model 17 can learn the similarity of the graph data by the common subset 210 and obtain the similarity S _ij which is an evaluation criterion for the search. At this time, the search model 17 can calculate the similarity even for the label data of the attention information which is not included in the common subset 210 at the time of learning. Therefore, in cases where the range of candidates for attention information cannot be determined in advance, such as when the label data of new attention information is updated in the database to be searched, linking of the attached information by the search model 17 is effective.

(Example of operation)
The manufacturing support system S has two input assistance functions, a change suggestion function, and an inspection assistance function. These functions are executed when the user specifies each function. However, each function may be executed without the user specifying it by setting conditions in advance. Each function will be described below in order with reference to Figs. 16 to 19.

Designer U1 assigns welding information to the weld lines that indicate the positions to be welded in the 3D-CAD structure. Although weld lines on 3D-CAD are generally expressed as sketch lines (one-dimensional lines that are structural components) or parts (3D structure), this system can be applied as long as the weld lines are expressed as structural components. However, it is desirable to have a unified method of expressing structures for each site where information management is performed. Here, an example of setting welding information for a circular metal plate 101 and a square metal plate 102 is explained with reference to Figures 16 and 17. If designer U1 were to assign information without using this system, it would be necessary to manually input all of the information for the weld lines 103 to 106 on the top surface of the square metal plate 102.

In this embodiment, the system complements the work of inputting the added information, thereby reducing the workload of designer U1. Two complementation methods are described below. In the first method, welding information that has already been input as welding information for a selected location is also added to other locations on the same 3D-CAD file. In the second method, welding information that has not been input in the selected range is estimated and added. The following process is provided as a function of the 3D-CAD software or an add-on function.

Figure 16 is a diagram illustrating a first method of the input assistance function. In the first method, the following four steps, steps 1-1 to 1-4, are executed in order. First, in step 1-1, the input unit 14 accepts an input operation by the designer U1, that is, a selection operation of a weld line 103 for which welding information has been input. However, multiple welds may be selected in this step. In the next step 1-2, the input unit 14 accepts an input from the designer U1 of a range in which there is a target to which welding information is to be assigned. In Figure 16, the selection range 305 selected by the designer U1 is indicated by a dashed rectangular parallelepiped. When the designer U1 selects the selection range 305, the suggestion unit 18 extracts the weld lines included in the selection range 305.

If the 3D-CAD structural elements have information indicating weld lines as targets to which attention information should be added, the suggestion unit 18 uses this information to set the weld lines 104, 105, and 106 as selected structures selected based on user input. If the 3D-CAD structural elements do not have information distinguishing weld lines, the suggestion unit 18 obtains the weld lines 104, 105, and 106 by extracting structures similar to the welds selected in step 1-1 based on the similarity of the graph data using the search model 17. The suggestion unit 18 uses the extraction unit 15 to extract welding information corresponding to the weld lines 104, 105, and 106. At this time, the extraction unit 15 uses the prediction model 16 or search model 17 specified by the designer U1 or set in advance.

In the next step 1-3, the suggestion unit 18 displays on a screen that the welding information of the welding line selected in step 1-1 will be added to the welding line extracted by the extraction unit 15 in step 1-2. In FIG. 16, the welding line to which the welding information is to be added is shown as a dashed oval. Here, an exclamation point displayed next to the welding line indicates that information has been added to the welding line, and a dashed speech bubble indicates that the welding information has not yet been finalized. The suggestion unit 18 accepts approval or correction by the designer U1. If approved, the suggestion unit 18 finalizes the addition of the welding information to the extracted welding line, and if corrected, the suggestion unit 18 adds the corrected welding information.

The above process determines the information to be added as shown in steps 1-4. Regardless of whether or not the welding information is corrected, it is reflected in the design accumulation DB 20 when it is determined. The total attention information graph DB 20G and each model set and subset are updated as appropriate according to a criterion predetermined by the user or when the user determines that the performance of the prediction model or the search model needs to be improved. The timing of the update may be, for example, every time a predetermined time has passed, or when the number of data items with corrections to the user-added information by the designer U1 reaches a certain level. Here, data that has been corrected from the attention information proposed by the user or the proposal unit 18 indicates that there is room for improvement in the system's proposal, and is important data that should be learned by the model as a priority in order to improve the performance of each prediction and search model as each model set and subset are updated. Therefore, it is desirable for the data reflected in the design accumulation DB 20 to hold information on whether or not the corrections have been made.

Figure 17 is a diagram explaining the second method of the input assistance function. In the second method, the four steps 2-1 to 2-4 described below are executed in order. First, in step 2-1, the input unit 14 accepts a designation from the designer U1 as to whether to use the prediction model 16 or the search model 17. However, step 2-1 may be omitted by completing the designation in advance. The designation by the designer U1 in step 2-1 may be made by selecting an option displayed on the screen, or by inputting a predetermined key from the keyboard.

In the next step 2-2, the input unit 14 accepts input from the designer U1 of the range in which the target to which welding information is to be assigned exists. In FIG. 17, the range specified by the designer U1 is indicated by a dashed line as the specified range 301. In the next step 2-3, the suggestion unit 18 first extracts welding lines from the specified range 301. The suggestion unit 18 then calculates welding information candidates for each welding line using the prediction model 16 or search model 17 specified by the designer U1 in step 2-1. The suggestion unit 18 then displays the calculated welding information candidates and waits for the designer U1 to make a selection. Note that in FIG. 17, for convenience of drawing, candidates are displayed only for the welding line 106, but in reality, candidates are displayed for all welding lines.

At this time, the suggestion unit 18 visually displays the corresponding probability. For example, the suggestion unit 18 may arrange the welding information from top to bottom in order of the probability, or may display the welding information with a higher probability larger. The suggestion unit 18 may also output the welding information with the highest probability as the first candidate. The "first candidate" is, for example, a candidate selected by default, and is a candidate that the designer U1 can select with the least operations. Note that although the table format is shown in FIG. 17, it is also possible to handle json or xml formats, which are highly flexible data structures, and the data may be displayed in a general tree format using a data structure visualization method. When the designer U1 selects the welding information for all the welding lines, the process proceeds to step 2-4, and all the welding information is confirmed. The confirmed welding information is reflected in the design accumulation DB 20, and is also reflected in the all-attention information graph DB 20G, the model set, and the subset.

Figure 18 is a diagram explaining the change suggestion function. In the change suggestion function, the following four steps, steps 3-1 to 3-4, are executed in order. Steps 3-1 and 3-2 are the same as steps 2-1 and 2-2 described above, so their explanation is omitted. However, in this example, some welding information has already been added to the weld lines 103 to 106.

In step 3-3, the suggestion unit 18 first extracts welding lines from the specified range 301. Next, the extraction unit 15 calculates welding information candidates for each welding line using the prediction model 16 or search model 17 specified by the designer U1 in step 2-1. Unlike step 2-3, the welding information candidate calculated here is only one with the highest probability. Then, for each extracted welding line, the suggestion unit 18 calculates the difference between the calculated welding information candidate and the welding information already assigned, and displays it on the screen as a change proposal 501. In FIG. 18, two exclamation marks are added to the welding lines that include the change proposal 501.

The change proposal 501 is preferably one that allows the user to check the information before and after the change to the welding information. However, there are no particular restrictions as long as the information allows the user to check the reasons why the welding information needs to be rechecked. The designer U1 judges whether the change proposal 501 is appropriate or not, and inputs the judgment result to the input unit 14. At this time, the designer U1 may input other correct values that are not written in the change proposal 501. In step 3-4, the suggestion unit 18 confirms each of the attached information based on the input by the designer U1 in step 3-3. The confirmed welding information is reflected in the design accumulation DB 20, and is also reflected in the total attention information graph DB 20G, the model set, and the subset.

The change suggestion function explained above has the effect of prompting the process designer to recheck the welding information in light of similar past cases when the process designer adds welding information to a weld line, thereby preventing errors from occurring when entering the welding information. In order to demonstrate the need for a system that achieves this effect, we will show the issues at product production sites. Even when it seems that the welding information has been properly specified in the upstream process of process design, it is often found that the construction conditions are difficult to implement in the manufacturing process. For example, it may become clear at the construction stage that a problem with the product structure makes it difficult to perform construction that satisfies the torch angle specified in the welding information.

When such problems occur, the manufacturing process staff or manager may deal with the problem by deciding on and implementing alternative processing conditions. However, there is a problem in that the data for changes to processing conditions is not managed in a way that allows the process designer to use the changes to improve future process condition instructions. Therefore, the change suggestion function in this embodiment links data indicating changes to process conditions to the target structure, thereby bringing the process designer to their attention and preventing errors from occurring when entering process conditions.

Figure 19 is a diagram explaining the inspection assistance function. Here, caution information is mainly used. The inspection assistance function executes the following four steps, procedure 4-1 to procedure 4-4, in order. When inspector U2 uses liquid crystal display D3, the image information shown in Figure 19 is displayed as is. When inspector U2 uses head-up display D2, the image information shown in Figure 19 may be displayed transparently, or the image information shown in Figure 19 excluding the part shape may be displayed.

First, in step 4-1, the suggestion unit 18 uses the CAD engine C to output image information of the structural information of the inspection point of the product targeted by the inspector U2 from the same viewpoint as the inspector U2. In the next step 4-2, the proposing unit 18 further displays the weld lines to which the information of the inspection items is linked, and check boxes for managing the progress of the inspection. Specifically, the weld lines 103-106 in the figure are marked with check boxes indicating the status of the end of the welding inspection, and the exclamation marks next to the check boxes are icons for displaying details of the information to be inspected. In step 4-2, none of the check boxes are checked. For example, in a case where the weld line 106 is similar to the structure near the weld line and was overlooked during the weld line inspection, the exclamation mark 401 is displayed in a different manner from the others to draw attention.

When the inspection is completed, the inspector U2 checks the checkbox to manage the progress. In step 4-3, the inspection of the weld line 103 and the weld line 105 is completed. The inspector U2 can also check the detailed information to be inspected as shown by the symbol 403 by selecting the exclamation mark icon. Furthermore, items that were overlooked in the past inspection may be displayed in a different manner from the others to further draw attention. For example, in the example shown in step 4-3, the item "Number of welding layers" is underlined to draw attention. In step 4-4, if the attention information given to the weld line is inappropriate, the suggestion unit 18 accepts corrections to the attention information from the inspector U2. The attention information corrected by the inspector U2 is reflected in the design accumulation DB 20.

The inspection assistance function described above has the effect of bringing to the attention of inspector U2 and preventing the occurrence of similar defects or information input errors by showing information related to defects or information input errors that have occurred in structure information similar to the target structure. The current issue with inspector U2 is that in order to prevent inspection oversights, multiple inspectors inspect the same inspection points multiple times. This increases the amount of work required, and there is a demand for both reducing the amount of work required and preventing inspection oversights.

In general, oversights during inspections are influenced by on-site conditions and human cognitive tendencies, and it is believed that there is a probabilistic bias in places where oversights are likely to occur. When considering accidents caused by improper inspections, as mentioned in Heinrich's rule of thumb, which is an empirical rule for workplace accidents, it is said that behind one serious accident, there are 29 minor accidents and 300 abnormalities. Similarly to this rule, behind the occurrence of serious oversights that occur during duplicate inspections, there are many minor oversights. Therefore, this function is useful as it prevents inspectors from oversights by linking inspection information that has been overlooked in the past to similar target structures and raising awareness.

In order to handle a state in which no attention information is added in the prediction model 16 and the search model 17, information indicating that no attention information is added (label data) is treated as one type of label data during the system processing of Fig. 19, or attention information is linked to target structure information (i) only when the addition probability or similarity S _ij , which is a criterion for acquiring candidates for added information, exceeds a certain threshold. The threshold value is also determined from an evaluation index of the accuracy of linking attention information using the common subset 210.

According to the above-described first embodiment, the following advantageous effects can be obtained.
(1) The manufacturing support system S includes a design accumulation DB 20 that stores three-dimensional design information including a plurality of partial structures, at least one of which is assigned with attention information, an input unit 14 that accepts a user's data operation on the three-dimensional design information, a model set generation unit 12 that converts the information of the partial structures into graph data, an extraction unit 15 that extracts attention information corresponding to the input graph data, and a suggestion unit 18 that inputs a selected structure, which is a partial structure selected based on an input by the user to the input unit 14, to the extraction unit 15 and suggests to the user that the attention information candidate, which is the attention information extracted by the extraction unit 15, be assigned to the selected structure. Therefore, the system can be used to support design, targeting not only welding information but also various data to be linked to structures that constitute a product.

(2) The suggestion unit 18 sets the partial structure including the part selected by the user via the input unit 14 as the selected structure. Therefore, the manufacturing support system S can suggest candidates for attention information to be assigned to the partial structure selected by the user.

(3) The suggestion unit 18 sets a partial structure similar to a partial structure including a part selected by the user via the input unit 14 as a selected structure. Therefore, the manufacturing support system S can suggest candidates for attention information to be assigned to other partial structures similar to the partial structure selected by the user.

(4) The manufacturing support system S includes a prediction model 16 that identifies and outputs candidates for attention information corresponding to the input graph data of the selected structure based on prior learning. The extraction unit 15 outputs the candidates for attention information output by the prediction model 16.

(5) The extraction unit 15 calculates likelihood information for each of the multiple candidate attention information. The extraction unit outputs the candidate attention information with the highest likelihood as the first candidate from among the multiple candidate attention information output by the prediction model 16. Therefore, the manufacturing support system S can prioritize the multiple candidate attention information and present them to the user.

(6) The manufacturing support system S includes a common subset 210 including multiple combinations of graph data and attention information, and a search model 17 that searches the common subset 210 for similar graph data, which is graph data similar to the graph data of the selected structure, and outputs the attention information combined with the similar graph data. The extraction unit 15 outputs the attention information output by the search model 17.

(7) The search model 17 calculates likelihood information for each of the multiple candidate attention information. The extraction unit 15 outputs the candidate attention information with the highest likelihood as the first candidate from among the multiple candidate attention information output by the prediction model. Therefore, the manufacturing support system S can prioritize the multiple candidate attention information and present them to the user.

(8) The proposal unit 18 updates the design accumulation DB 20 by recording the user's response to the proposal in the design accumulation DB 20. The manufacturing support system S includes a model update unit 19 that has a function of updating the parameters of the search model using the design accumulation DB 20 updated by the proposal unit 18.

(9) The proposal unit 18 updates the design accumulation DB 20 by recording the user's response to the proposal in the design accumulation DB 20. The manufacturing support system S includes a model update unit 19 that has a function of updating the parameters of the prediction model using the design accumulation DB 20 updated by the proposal unit 18.

(10) The attention information includes at least one of the following regarding the welding of the partial structure to which the attention information is attached: construction conditions, inspection information, welding construction defects, information related to accident cases, and jig information.

(11) The attention information includes at least one of information on the screw holes corresponding to the screws, information on the type of screw, inspection information, construction defects, and information related to accident cases, regarding the fastening of the partial structure to which the attention information is attached using screws.

(12) The attention information includes at least one of information related to the adhesive used for bonding the partial structure to which the attention information is attached, inspection information, defects, and accident cases.

As described in this embodiment, constructing a system that collects and manages data in a common format is highly compatible with information addition using machine learning, and data scientists can continuously improve the prediction model 16 and the search model 17 using the collected database. The subgraph data that is input to the prediction model 16 and the search model 17 is obtained by partially extracting the feature vectors and graph data held by the nodes and edges of the graph data, and this extraction method also follows the state of the updated database, making it easy for data scientists to improve it based on the evaluation results of the prediction accuracy of the information of interest.

Manufacturing support system S manages information storage, search, and utilization of data at production and development sites. Manufacturing support system S converts structural information on 3D CAD, which is the product design drawing, into graph data, and uses it as data in a common format for storage, search, and utilization. As a result, manufacturing support system S promotes the reuse of data that has not been utilized until now by horizontally deploying past cases and similar cases, and achieves data-based search and prediction of a wide range of information to be added to structures. In this way, manufacturing support system S improves work efficiency by supporting users' awareness to prevent human error and complementing the user's information addition work.

In addition, by adding data obtained from the user's feedback on the success or failure of the focus information on the structure information to the database, the prediction model 16 and the search model 17 can be continuously improved. At this time, the newly created prediction model 16 and search model 17 also use graph data as input, and the prediction model 16 and search model 17 can be updated.

As described above, the predictive model 16 and the search model 17 can be continuously updated, and a general-purpose predictive model 16 and search model 17 can be generated in response to changes in the search subset 220 due to the following two points: (1) the data within the system is unified into graph data, which is a flexible and general-purpose data format, and (2) a method similar to GNN is used that uses graph data as input to the predictive model 16 and the search model 17 and generates a model with high generalization performance for a variety of input data.

In the above-described embodiment, the functional block configuration is merely an example. Several functional configurations shown as separate functional blocks may be configured as one, or a configuration shown in one functional block diagram may be divided into two or more functions. In addition, some of the functions of each functional block may be provided by other functional blocks. The present invention is not limited to the contents of the above-described embodiment. Other aspects conceivable within the scope of the technical concept of the present invention are also included in the scope of the present invention.

12... Model set generation unit 13... Subset generation unit 14... Input unit 15... Extraction unit 16... Prediction model 17... Search model 18... Proposal unit 19... Model update unit 20... Design accumulation database 20G... Total attention information graph database 200... Model set 210... Common subset 220... Search subset

Claims (13)

an accumulation database for storing three-dimensional design information including a plurality of partial structures, at least one of the plurality of partial structures being assigned attention information;
an input unit that accepts data operations by a user on the three-dimensional design information;
a data conversion unit that converts information about the partial structure into graph data;
an extraction unit that extracts information of interest corresponding to the input graph data;
a proposal unit that inputs a selected structure, which is the partial structure selected based on an input by the user to the input unit, to the extraction unit, and proposes to the user that candidate attention information, which is the attention information extracted by the extraction unit, be assigned to the selected structure. 2. The manufacturing support system according to claim 1,
The proposal unit sets the partial structure including the part selected by the user via the input unit as the selected structure, in this manufacturing support system. 2. The manufacturing support system according to claim 1,
The suggestion unit sets the partial structure similar to the partial structure including a portion selected by the user via the input unit as the selected structure, in this manufacturing support system. 2. The manufacturing support system according to claim 1,
The method further includes a prediction model that identifies and outputs candidates for the attention information corresponding to the input graph data of the selected structure based on prior learning,
The extraction unit outputs candidates for the attention information output by the prediction model. The manufacturing support system according to claim 4,
The prediction model further calculates likelihood information for each of the plurality of candidates for the attention information,
The extraction unit outputs, as a first candidate, the candidate for attention information with the highest likelihood from among the multiple candidates for attention information output by the predictive model. 2. The manufacturing support system according to claim 1,
a search database including a plurality of combinations of the graph data and the attention information;
a search model that searches the search database for similar graph data that is similar to the graph data of the selected structure, and outputs the attention information that is combined with the similar graph data;
The extraction unit outputs the attention information output by the search model. 7. The manufacturing support system according to claim 6,
The search model further calculates likelihood information for each of the plurality of candidates of the attention information,
The extraction unit outputs, as a first candidate, the candidate of interest information having the highest likelihood from among the multiple candidates of interest information output by the search model. The manufacturing support system according to claim 4,
The input unit updates the accumulation database by recording a response of the user to the suggestion by the suggestion unit in the accumulation database;
The manufacturing support system further comprises a model updating unit having a function of updating parameters of the prediction model using the accumulated database updated by the input unit. 7. The manufacturing support system according to claim 6,
The input unit updates the accumulation database by recording a response of the user to the suggestion by the suggestion unit in the accumulation database;
The manufacturing support system further comprises a model update unit having a function of updating parameters of the search model using the accumulation database updated by the input unit. 2. The manufacturing support system according to claim 1,
A manufacturing support system, wherein the attention information includes at least one of construction conditions, inspection information, welding construction defects, information related to accident cases, and jig information regarding the welding of the partial structure to which the attention information is attached. 2. The manufacturing support system according to claim 1,
A manufacturing support system in which the attention information includes at least one of information regarding the fastening of the partial structure to which the attention information is attached using a screw, information regarding the screw hole corresponding to the screw, information regarding the type of the screw, inspection information, construction defects, and information regarding accident cases. 2. The manufacturing support system according to claim 1,
A manufacturing support system, wherein the attention information includes at least one of information relating to the adhesive used in bonding the partial structure to which the attention information is attached, inspection information, defects, and accident cases. A manufacturing support method executed by a computer having a database storing three-dimensional design information including a plurality of partial structures, at least one of the plurality of partial structures being assigned attention information, the method comprising:
an input step of accepting a user's data operation for the three-dimensional design information;
a data conversion step of converting information of the partial structure into graph data;
an extraction step of determining a similarity between the plurality of partial structures based on the graph data;
a proposal step of inputting a selected structure, which is the partial structure selected based on the user's input in the input step, into the extraction step, and proposing to the user that candidate attention information, which is the attention information extracted by the extraction step, be assigned to the selected structure.

Images