JP7413854B2 - Design support system - Google Patents

Description

The present invention relates to a design support system.

In the development of products, including the development of materials, we draw up design guidelines and conduct prototype experiments based on them in order to achieve the product functions required by our customers. When developing materials, where it is necessary to select the optimal one from a huge number of combinations of compositions and process conditions, the number of prototype experiments that can be performed is limited, so it is necessary to formulate and test appropriate design policies. It is important to determine the material system, composition conditions, etc.

Developing effective design guidelines requires advanced domain knowledge regarding the development target. In this case, it is effective to take out and utilize the knowledge of other developers and the knowledge found in past design performance data.

An example of a method for disseminating developer knowledge is to organize knowledge about the relationship between product functions and product specifications/processes in the form of knowledge graphs based on quality function deployment, function block diagrams, ontology, etc. There is.

Methods for extracting knowledge from past performance information include predicting product characteristics from product specifications consisting of numerical values, natural language, molecular structure, etc. using statistics, machine learning, deep learning, etc., or product characteristics such as SHAP. There are methods to calculate the contribution rate of each variable to Here, SHAP is a method for determining the contribution of each variable (feature amount) to the prediction result of a model, and is an abbreviation for SHapley Additive exPlanations.

Furthermore, if performance information is obtained through randomized experiments, causal effects can be estimated through statistical causal inference. Causal effects derived from randomized experiments are considered reliable as a basis for decision-making. Regarding decision-making, meta-analysis, which collects multiple causal inference results and discusses causal relationships, is said to be more reliable than causal inference results.

Patent Document 1 discloses that an experimental plan or optimization method is selected from performance data including design values and performance values of a product, optimization targets, and constraints, and the selected experimental plan or optimization method is executed. An experiment support device is disclosed that calculates the effect of a plan or optimization method and outputs the effect in association with the experimental plan or optimization method.

Patent Document 2 discloses a causal relationship model and analysis results using a data analysis function for the purpose of providing technology for effectively updating a model that estimates occurring events for decision-making support through simulation. Based on this, a decision support system is disclosed that generates update data for updating a causal relationship model.

Patent Document 3 integrates a product structure information model, which is an information representation model based on product structure (E-BOM), and a process configuration information model, which is an information representation model based on process structure (M-BOM). Product/process model DB, which is a database that stores integrated models (product-process models (PPM)), and knowledge files (raw materials) related to knowledge about design specifications, standards/criteria, manufacturing methods, and knowledge/know-how in production systems. A knowledge/know-how systematization support device that has a knowledge data DB that stores (1) processes can be expressed by state transitions, (2) knowledge/know-how can be systematized, (3) knowledge - It is disclosed that know-how can be accumulated and evolved, and since the integrated model expresses processes by state transitions, it is possible to express the failure occurrence mechanism based on the causal chain relationship of failures in the production system.

Japanese Patent Application Publication No. 2019-185591 Japanese Patent Application Publication No. 2017-194730 Japanese Patent Application Publication No. 2007-241774

In order to support decision-making in complex product design in materials development, etc., it is important to provide developers with information on which product specifications are important for functional expression with a high level of comprehensiveness and accuracy.

According to the experiment support device described in Patent Document 1, it is possible to make an optimal experiment plan based on actual data and the like. However, Patent Document 1 does not describe whether information necessary for determining a design policy can be obtained, such as what kind of binary relationship exists between product specifications and characteristics.

According to the decision support system described in Patent Document 2, it is possible to support updating of a causal model using data analysis results and the like. However, since Patent Document 2 does not mention whether the data is randomized, it is difficult to determine whether the correlation is due to causality or a false correlation, and the accuracy of the causal model decreases. It is thought that problems may occur, such as problems in which the model cannot be used for decision making, or problems in which user operations are required to update the model.

According to the knowledge/know-how systematization support device described in Patent Document 3, knowledge/know-how can be systematized, accumulated, and evolved. However, Patent Document 3 does not specifically describe handling quantitative data, nor does it describe providing accurate information based on quantified knowledge.

The present invention aims to support the user by utilizing experimental data, the designer's knowledge, etc. when the user formulates a design policy.

The design support system of the present invention includes an input device capable of inputting knowledge regarding causal relationships, and integration of the knowledge regarding the causal relationships inputted into the input device with causal inference results obtained by causally inferring a plurality of experimental data for each group. and an output device that displays a result of integration by the arithmetic device.

According to the present invention, when a user formulates a design policy, it is possible to support the user by utilizing experimental data, the designer's knowledge, and the like.

FIG. 1 is a configuration diagram showing a design support system according to an embodiment. 3 is a table showing an example of an experimental data set input to the input device of the example. It is an image showing an example of causal relationship knowledge data input to the input device of the example. FIG. 2 is a configuration diagram showing an example of a schema of data handled in the design support system of the embodiment. It is a display screen showing an example of causal relationship information integrated in the design support system of the embodiment. It is a screen showing an example of details of causal relationships displayed in the design support system of the embodiment. It is a screen showing another example of details of causal relationships displayed in the design support system of the embodiment. It is an example of the screen displayed when registering causal relationship knowledge in the design support system of an example. It is an example of the screen displayed when performing optimization processing in the design support system of the embodiment.

A design support system according to an embodiment of the present invention records and presents analysis results of knowledge and design performance data necessary for formulating a product design policy.

Embodiments will be described below.

In this embodiment, the knowledge and know-how regarding the causal relationship between product functions such as materials and product specifications will be combined with product specifications (for example, material composition and process conditions) and product functions and characteristics (prototype experiment result values, etc.). Organize into interterm relationships. Furthermore, we derive causal relationships from product specifications to product functions and characteristics by making causal inferences from experimental data grouped in units of randomized experiments. We statistically integrate causal relationship knowledge/know-how and quantitative causal inference results associated with each combination of product specifications and product functions/characteristics, perform meta-analysis, and present the results to developers.

Causal relationship knowledge and know-how include the size, tendency, and direction of the causal relationship that the developer believes, texts and references that indicate the basis and principle of the causal relationship, references to external information, diagrams, and the developer himself who showed the knowledge. Contains information such as.

Experimental data grouped by randomized experiment unit refers to which randomized experiment (including experimental data obtained by experimental design methods such as orthogonal arrays, optimal designs, response surface designs, etc.) in which the experimental data was obtained. It is experimental data recorded in a form that can be identified. In other words, the group is composed of units of randomized experiments.

Causal inference results are the results of estimating causal effects by performing causal inference using statistical methods and machine learning for each group of experimental data. Causal effects include ATE (Average Treatment Effect), CATE (Conditional Average Treatment Effect), or their extension to continuous value interventions, and propensity scores. As causal inference methods, linear methods and machine learning-based methods such as Causal Forest, X-learner, and Causal Effect VAE are used as necessary. These causal inferences are performed automatically or by manual modeling by humans.

The experimental data includes experimental results in the real world and results from simulation experiments.

Finally, multiple causal relationship knowledge/know-how and causal inference results that have specific product specifications and product functions/characteristics as causes and effects are integrated using statistical operations, meta-analysis techniques, etc., to create a highly readable system for system users. Present in form.

FIG. 1 shows the configuration of a design support system according to an embodiment.

In product design work, the design support system shown in this figure typically uses the causal effect obtained by causal inference from experimental data and the product designer's own ideas when formulating product design guidelines before conducting prototype experiments. This system presents product designers with information that integrates causal relationship knowledge and supports the formulation of design guidelines. Note that this design support system may be a system that is a collection of distributed devices, or may be a design support device that has all functions integrated into one device.

In this figure, the design support system includes an input section 101 (input device), a recording section 104 (recording device), a causal inference section 105, an integration section 106, and a display section 107 (output device). There is. The causal inference unit 105 and the integration unit 106 are included in the arithmetic device. Note that the causal inference section 105 and the integration section 106 may be provided in separate devices. Furthermore, any two of the input device, arithmetic device, and output device may be configured as one device. For example, the host computer may be a computing device, and the input device and output device may be a single mobile terminal owned by the user.

The input unit 101 includes an experimental data set input unit 102 and a causal relationship knowledge input unit 103. The experimental data set input unit 102 receives input of experimental data grouped in units of randomized experiments (hereinafter referred to as "experimental data set"). The causal relationship knowledge input unit 103 receives input of knowledge/information regarding the causal relationship between product specifications and product characteristics (causal relationship knowledge) from the user. Information input at the input unit 101 is recorded in the recording unit 104.

The causal inference unit 105 performs causal inference using the data of the experimental data set recorded in the recording unit 104, and calculates causal effects and related indicators (hereinafter referred to as “causal inference results”). The calculated causal inference result is recorded in the recording unit 104.

The integrating unit 106 extracts those that have a common cause and effect of a causal relationship from among the causal relationship knowledge and causal inference results, and performs an integration process on the extracted causal relationship knowledge and causal inference results.

The display unit 107 presents the causal knowledge and causal inference results integrated by the integrating unit 106 and the raw data input by the input unit 101 to the user using a plurality of visualization methods.

Examples will be described below with reference to the drawings.

FIG. 2 shows a part of a group of experimental data sets input by the input unit 101 of FIG.

In FIG. 2, a group 201 of experimental data sets includes experimental data sets 202a and 202b. One of these, the experimental data set 202a, includes metadata 203 of the experimental data set indicating the name of the experimental data set, creation date, etc., and a plurality of experimental data 204 expressed as a table. . In this figure, the column "Material Composition 3" is shown as one of the experimental data 204, surrounded by a broken line.

The plurality of experimental data 204 is composed of columns of product specifications such as material composition, process conditions, feature quantities of constituent molecules, feature quantities characterizing the product specifications, product properties such as material properties, etc. Good too. Values in the table are represented by numerical values, character strings representing categories, and the like. The experiment data 204 also includes a column whose value is a unique key for each row, such as an experiment number. The product specifications, product characteristics, etc. of the experimental data 204 are based on experiments constructed using experimental design methods such as orthogonal arrays, optimal designs, and Latin superlattice methods, or from experiments generated using experimental design methods modified by the product designer. can get.

FIG. 3 shows part of the causal relationship knowledge data input to the input unit 101 of FIG. 1.

In FIG. 3, causal relationship knowledge data 301 includes a plurality of causal relationship knowledge 302a and 302b. The causal relationship knowledge 302a includes metadata 303 such as knowledge ID, creation date, and creator, information 304 such as product specifications/features (causes), and product characteristics (results) that are subject to the knowledge, and a graph 305. , a score that represents the size of the causal effect that the user thinks (a value standardized from -100 to 100, etc.), text that shows the basis of the finding, and additional information such as reference data for the finding, images, and links to external information. and information 307 representing the range/conditions (material system, etc.) under which the knowledge is thought to be valid.

Information such as a reason for the knowledge and a memo 306 is a blank field 308 that is not displayed in the causal relationship knowledge 302b, and can be added as an arbitrary item in the example shown in this figure.

FIG. 4 shows an example of a data model 401 of data recorded by the recording unit 104 in FIG. 1.

In this figure, a data model 401 has a structure in which an experimental data set 403 and causal knowledge data 404 refer to entity data 402. Here, the entity refers to product specifications (material composition, process conditions, material composition features, etc.) and product characteristics (test items, experiment result items, etc.).

Entity data 402 stores the name of the entity, the type of entity (category information such as material composition and material properties, information on further classification within material composition, etc.), information on parent entities, and other non-structural information. This includes metadata, etc. When there is a parent-child relationship between two entities, the child entity can be a subconcept of the parent entity (for example, parent: additive, child: hardener, etc., if the child is the parent) or a subconcept (for example, parent: material A). , child: if the child is a component of the parent, such as a monomer or additive). An entity dictionary is prepared in advance for these entity data 402 and the relationships between the entity data 402. Then, the input experimental data sets 202a, 202b (FIG. 2) or causal relationship knowledge 302a, 302b (FIG. 3) are linked to the entity data 402, or unregistered entities are linked to the experimental data sets 202a, 202b (FIG. 2). ) or by newly registering an entity when it is included in the cause/effect items of the causal relationship knowledge 302a, 302b (FIG. 3).

The experimental data 407 is composed of records corresponding to rows in the table of the experimental data 204 that constitute the experimental data sets 202a and 202b in FIG. 2, and the data included in the experimental data set 403 is referenced.

The causal inference result set 405 stores the causal inference results for the records of the experimental data set 403, and includes relationships to the causal entities and result entities of the causal inference results, and includes the causal inference results that indicate the cause and effect of the causal inference. Contains metadata, which is unstructured data that describes methods, causal effect values, distribution of causal effect values, inference accuracy, inference conditions, etc. Further, causal inference is performed automatically or manually, but when causal inference is performed manually, information of the user who made the causal inference is recorded as the creator. When performing automatic causal inference, the causal relationships from the cause column to the items in the result column are inferred using multiple methods such as linear regression-based causal inference and more complex machine learning-based causal inference, resulting in high estimation accuracy. Adopt and record things. If necessary, it is preferable to perform preprocessing such as feature generation using a molecular descriptor or the like on the cause/effect column. Further, regarding the set of causal inference results 405 and the causal knowledge data 404, evaluations of the results from users who viewed them are recorded as integer values. User data 408 includes information such as user name, affiliation, login password, and action history.

The integrated data 406 has, as records, references to multiple causal inference results and causal relationship knowledge that have a common cause entity and result entity, and includes causal effects and causal relationships of those multiple causal inference results and causal relationship knowledge. An integrated causal effect score calculated by statistical processing based on the score is stored. In other words, in the recording unit 104 of FIG. 1, it is desirable to save the causal inference results, and to record and accumulate the integrated data 406, which is the result of integration, as the causal inference results (that is, to accumulate the results of integration, learning) is desirable.

FIG. 5 shows an example of a screen of integrated causal relationship information displayed on the display unit 107 of FIG. 1.

A screen 501 shown in FIG. 5 is displayed on a desktop application, a web browser, or the like. Specifically, on the screen 501, an entity search user interface 502, search results 503, a causal relationship table 506, and the like are displayed. It is desirable that the causal relationship table 506 be in the form of a two-way table.

To display the causal relationship, first, the user inputs the product specifications, feature amounts, product characteristic names, etc. for which the causal relationship is to be confirmed on the entity search user interface 502, and executes a search. Then, the result closest to the search is displayed in the search results 503.

As the search proximity, we use the similarity when input strings, entity names to be searched, etc. are vectorized using word2vec, character-based LSTM, etc., edit distance such as Levenshtein distance, etc. By clicking the "Add to Column" button 504 or the "Add to Row" button 505, the product specifications, feature values, names of product characteristics, etc. shown in the search results 503 can be added to the column 507 of the causal relationship table 506. can be added to line 508.

In the causal relationship table 506, the column 507 corresponds to the entity name corresponding to the cause of the causal relationship, and the row 508 corresponds to the entity name corresponding to the result. As a result, in each cell 510, information on the causal relationship corresponding to "cause" in column 507 to "result" in row 508 is displayed. Inside the cell 510, an integrated causal relationship 509 is illustrated as a graph. The graph of the causal relationship 509 uses an integrated causal relationship score included in the integrated data 406 (FIG. 4). The slope of the graph rises to the right when the integrated causality score is positive, and falls to the right when it is negative.

The causal relationship breakdown display 511 can display positive causal effects, negative causal effects, the number of causal inference result sets 405 and causal knowledge data 404 (FIG. 4) in which it is determined that there is no causal relationship, and the like. Furthermore, by clicking the arrow in the breakdown display 511, causal knowledge can be easily registered and the number of records in the causal knowledge data 404 (FIG. 4) can be increased. When causal relationship knowledge is registered here, the portion corresponding to the registered causal relationship is highlighted and displayed in the breakdown display 511.

Cell 510 of causality table 506 can be selected by clicking or the like. By clicking the button 512 for displaying details of the selected causal relationship, a transition can be made to a screen for viewing detailed information on the causal relationship information corresponding to the selected cell 510.

Note that in this design support system, user management is performed. Therefore, the screen 501 displays the name 513 of the user currently using the service. The user can transition to the knowledge registration screen by clicking the knowledge registration button 514. Furthermore, by clicking the optimization execution button 515, it is possible to transition to a screen for confirming what product specifications are optimal for realizing the desired product characteristics.

FIG. 6 shows an example of a detailed confirmation screen for causal relationships that is displayed when the button 512 in FIG. 5 is clicked.

The detailed confirmation screen 601 shown in FIG. 6 includes a causal inference result list 602 and a causal knowledge list 608. On the detailed confirmation screen 601, you can view the data of the causal inference result set 405 and the causal knowledge data 404 (FIG. 4), which are the basis for the integrated result of the causal relationships corresponding to the cause and effect of the cell 510 selected in FIG. . A plurality of data of the causal inference result set 405 and the causal knowledge data 404 may be displayed.

Each frame 603 of data in the causal inference result set 405 displays the name of the experimental data set on which the causal inference result is based, an experimental data display button 604 for displaying the experimental data, accuracy and details of the inference result, etc. A detailed analysis information display button 605 and the like are provided to display the detailed analysis information.

When the user clicks on the experimental data display button 604 or analysis detailed information display button 605, the corresponding information is displayed in a modal window, file download, or the like. The causal inference results are visualized in a form such as a graph 606 with a causal effect represented by a curve and a confidence interval represented by a broken line, or a box plot 607. If there are CATE values other than the graph 606, they are displayed.

In the causal knowledge list 608, a plurality of causal knowledge data 404 (FIG. 4) corresponding to the cause and effect corresponding to the selected cell 510 (FIG. 5) are displayed. One frame 609 of the causal knowledge list 608 includes information 610 that visualizes the causal relationship score in the causal knowledge data 404 as the slope of a graph, the registrant, the reason/principle of the knowledge, image data, document data, etc. A memo column 611 and the like including icons are displayed. These are information included in the causal knowledge data 404. As shown in a frame 612, if reasons/principles, etc. are not registered, they are not displayed.

In summary, the output device displays, on the detailed confirmation screen 601, the result of integration, the causal inference result used as the basis for the result of integration, and the knowledge regarding the causal relationship.

FIG. 7 shows another example of the causal relationship details confirmation screen that is displayed when the button 512 in FIG. 5 is clicked.

A detailed confirmation screen 701 shown in FIG. 7 displays one causal inference result 702. The causal inference result 702 is an example of a case based on a data set with a large number of experimental data, that is, large-scale data.

When there is a large amount of experimental data, the visualization method is to create a machine learning model that predicts product characteristics using ensemble learning (random forest, gradient boosting, etc.), and then use LIME (Local Interpretable Model-agnostic Explanations), SHAP, etc. By using the explanatory model, the contribution of each explanatory variable can be calculated. Therefore, it is possible to employ a different visualization method than when there is little data.

The causal relationship visualization unit 703 in the causal inference result 702 visualizes the contribution of variables using SHAP values, and then uses a machine learning-based causal inference method (such as a method using S, T, X-Learner, etc.) to visualize the contribution of variables. This is an example of displaying the calculated CATE. Switching of the display method between the causal inference result list 602 and the causal inference results 702 is performed on a rule basis based on the information in the causal inference result set 405.

FIG. 8 shows a causal relationship knowledge registration screen that is displayed when the knowledge registration button 514 in FIG. 5 is clicked or when the number of findings display section of the causal relationship breakdown display 511 in FIG. 5 is clicked. be.

When the knowledge registration button 514 in FIG. 5 is clicked, the causal relationship knowledge registration screen 801 shown in FIG. You can register.

The registration screen 801 includes a causal relationship score input section 802 that allows input and visualization of the causal relationship score using a slider or numerical value. Here, the causal relationship score is determined by the user, who evaluates the magnitude of the causal relationship within a specific numerical range based on the user's subjective and experience-based evaluation or objective evaluation such as literature information. This is what you enter. Although evaluations based on user subjectivity and experience are less certain and less reliable than causal inference results or literature information, the input cost is lower than obtaining experimental data or literature information. Therefore, by inputting evaluations based on the user's subjectivity and experience, it is possible to collect causal knowledge for a large number of conditions over a wider range of conditions, although the reliability is low.

In the condition input unit 803 for establishing a causal relationship, the product specifications for which a causal relationship is established are defined by specifying constraints on the product specifications and representative product specifications using text or numerical values. In the input unit 804 for the reason/principle of the causal relationship, the cause or reason for the establishment of the corresponding causal relationship is input using rich text, an attached file, a reference to external information, or the like. The causal relationship reference material input unit 805 accepts input of other reference information such as rich text, attached files, and references to external information. By clicking the registration button 806, the input contents are reflected in the causal knowledge data 404.

When the number of findings display section of the causal relationship breakdown display 511 in FIG. 5 is clicked, in the causal relationship score input section 802 of FIG. The initial value is displayed based on the option of , no causal effect, or negative causal effect.

In summary, the output device has a configuration that allows input of additional knowledge regarding the integration results along with the integration results on the registration screen 801.

FIG. 9 shows an optimization screen displayed when the optimization execution button 515 in FIG. 5 is clicked.

In FIG. 9, an optimization screen 901 includes a target input section 902 and an optimization result 905 display section. In the target input unit 902, for each of the product specifications specified in the row 508 of the causal relationship table 506 in FIG. input the target from option 903. When the execution button 904 is clicked, optimization results 905 are displayed.

Here, the optimization result 905 is based on the integrated data 406 (FIG. 4) of causal relationship knowledge and inference results, and the product specifications in column 507 (FIG. 5) of the causal relationship table 506 that best matches the optimization goal. Combinations of feature items are calculated using optimization methods such as genetic algorithms and simplex methods. The optimization result 905 includes the name of the product specification/feature item in column 507 (FIG. 5) of the causal relationship table 506, the optimization result (whether the product specification value should be increased or decreased, or whether it should be brought closer to a certain value). etc.) are displayed as numerical values or character strings.

In summary, the output device displays the target input section and the optimization result on the optimization screen 901.

The effects of the above embodiments and examples will be described below.

According to the above embodiments, when formulating product design guidelines, the system user presents the binary relationship between each product specification and product characteristics based on knowledge and causal inference results. By doing so, appropriate design policy decisions can be made.

Additionally, because experimental data is accumulated in randomized units, statistical causal inference is possible. Therefore, the system can accumulate causal inference results, and highly reliable information can be presented as the basis for decision-making.

Meta-analysis can be facilitated by organizing the relationships into binary relationships and accumulating multiple findings and causal inference results. Even if individual knowledge or analysis results are unreliable, by integrating this information, more reliable information can be provided.

Furthermore, although the range that can be covered by causal relationships obtained from experimental data is limited, by accumulating knowledge at the same time, it is possible to supplement information in areas where data is missing.

101: Input section, 102: Experimental data set input section, 103: Causal relationship knowledge input section, 104: Recording section, 105: Causal inference section, 106: Integration section, 107: Display section, 201: Experimental data set group, 202a, 202b: Experimental data set, 203, 303: Metadata, 204: Experimental data, 301: Causal relationship knowledge data, 302a, 302b: Causal relationship knowledge, 304, 307: Information, 305: Graph, 306: Memo, 308 : Blank, 401: Data model, 402: Entity data, 403: Experimental data set, 404: Causal knowledge data, 405: Causal inference result set, 406: Integrated data, 407: Experimental data, 408: User data, 501: Screen , 502: Entity search user interface, 503: Search results, 505, 512: Button, 506: Causality table, 507: Column, 508: Row, 509: Causality, 510: Cell, 511: Breakdown display.

Claims (11)

an input device capable of inputting knowledge regarding causal relationships;
an arithmetic device that integrates the knowledge regarding the causal relationship inputted into the input device and a causal inference result obtained by causally inferring a plurality of experimental data for each group;
A design support system, comprising: an output device that displays the result of the integration by the arithmetic device. 2. The design support system according to claim 1, wherein the group is composed of units of randomized experiments. 2. The design support system according to claim 1, wherein the causal relationship includes a binary relationship between a product specification and a product function. 2. The design support system according to claim 1, wherein the integration is performed by statistical processing. Further equipped with a recording device,
The design support system according to claim 1, wherein the recording device stores the causal inference results. 6. The design support system according to claim 5, wherein the recording device records and stores the result of the integration as the result of causal inference. The design support system according to claim 1, wherein the input device is capable of inputting the causal inference result. 2. The design support system according to claim 1, wherein the output device displays the causal relationship table that is the result of the integration in a two-way table format. 2. The design support system according to claim 1, wherein the output device displays the result of the integration, the causal inference result used as a basis for the result of the integration, and the knowledge regarding the causal relationship. 2. The design support system according to claim 1, wherein the output device has a configuration that allows input of additional knowledge regarding the results of the integration together with the results of the integration. The design support system according to claim 1, wherein the output device displays a target input section and an optimization result.

Images