WO2020225959A1 - Design assistance system, design assistance method, and design assistance program - Google Patents

Design assistance system, design assistance method, and design assistance program Download PDF

Info

Publication number
WO2020225959A1
WO2020225959A1 PCT/JP2020/006914 JP2020006914W WO2020225959A1 WO 2020225959 A1 WO2020225959 A1 WO 2020225959A1 JP 2020006914 W JP2020006914 W JP 2020006914W WO 2020225959 A1 WO2020225959 A1 WO 2020225959A1
Authority
WO
WIPO (PCT)
Prior art keywords
design
group
learning model
unit
design support
Prior art date
Application number
PCT/JP2020/006914
Other languages
French (fr)
Japanese (ja)
Inventor
野中　紀彦
Original Assignee
株式会社日立製作所
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 株式会社日立製作所 filed Critical 株式会社日立製作所
Publication of WO2020225959A1 publication Critical patent/WO2020225959A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Definitions

  • the present invention relates to a design support system, a design support method, and a design support program.
  • Patent Document 1 discloses "a design support technique for generating an analysis model for CAE (Computer Aided Engineering) from CAD design data using a learning model”.
  • Patent Document 2 discloses "a design support technique for narrowing down the classification of similar design data based on the feature data extracted from the design data and displaying it as related design information".
  • Patent Document 1 a data group in which a user's quality evaluation is given as a teacher value to past design data and an analysis model is prepared as learning data. Machine learning of the learning model is performed using this learning data. In such a learning model, good output results can be easily obtained if machine learning is performed using a sufficient amount of learning data. However, in the initial stage where the training data is insufficient, good output results cannot be obtained because the machine learning is not sufficient. No specific disclosure of such a problem or its solution is found in Patent Document 1.
  • design data is classified into a group of similar design data in the past by a similarity search of feature data.
  • the groups classified as having high similarity may not be appropriate groups.
  • an object of the present invention is to provide a new technique for supporting the design.
  • one of the representative design support systems of the present invention is from the requirement specification input unit for inputting the requirement specifications of the design and the group of groups that classify the past design history. It is equipped with a classification unit that obtains a group in which required specifications are classified (hereinafter referred to as "affiliation group”) and a calculation result display unit that acquires information about the affiliation group and makes it possible to present it as design support information. Acquires information about a group in which design results different from the required specifications are classified (hereinafter referred to as "related groups”) for design candidates designed based on the required specifications, and makes it possible to present them as design support information.
  • a classification unit that obtains a group in which required specifications are classified
  • a calculation result display unit that acquires information about the affiliation group and makes it possible to present it as design support information.
  • Acquires information about a group in which design results different from the required specifications are classified hereinafter referred to as "related groups” for design candidates designed based on the required specifications, and makes it possible to present them
  • FIG. 1 is an overall configuration diagram of the first embodiment.
  • FIG. 2 is a diagram showing a processing procedure (phase 1) of the first embodiment.
  • FIG. 3 is a diagram showing the processing procedure (Phase 2) of the first embodiment.
  • FIG. 4 is a diagram showing the processing procedure (Phase 3) of the first embodiment.
  • FIG. 5 is a diagram showing an example of an input screen for calculation conditions.
  • FIG. 6 is a diagram showing an example of an analysis result display screen.
  • FIG. 7 is a diagram showing an example of an input screen of the required specifications.
  • FIG. 8 is a diagram showing an example of a design support display screen.
  • FIG. 9 is a diagram showing an example of a design result input screen.
  • FIG. 10 is a diagram showing an example of a design result display screen.
  • FIG. 10 is a diagram showing an example of a design result display screen.
  • FIG. 11 is a diagram showing an example of a display screen of a network structure without networking.
  • FIG. 12 is a diagram showing an example of a display screen of a network structure in the case of networking.
  • FIG. 13 is a diagram showing an example of a design support display screen after networking.
  • FIG. 1 is a diagram showing the configuration of the design support system of the first embodiment.
  • the design support system includes a requirement specification input unit 101, a classification unit 102, a calculation result display unit 103, a design result input unit 104, and a network structure calculation unit 105. More specifically, a calculation condition input unit 111, a database 112, a data analysis unit 113, a learning model generation unit 114, and a machine learning unit 115 are provided inside the classification unit 102.
  • the requirement specification input unit 101 provides an operator with an input means for inputting a design requirement specification.
  • the "required specification” means a specification such as an item, data, or a specification that an object created by design should satisfy, and a specification that should be specified in order to receive design support.
  • dimensions, shapes, components, characteristics, etc. are required specifications, but the specifications are not limited to these.
  • the classification unit 102 searches for a group to which the required specifications should be classified from the group of groups that have classified the past design history, and sets it as the "affiliation group".
  • the calculation result display unit 103 acquires the information of the belonging group and provides it to the operator as design support information. Based on this design support information, the operator temporarily or automatically determines a design candidate.
  • the operator performs measurement experiments of prototypes, cost calculation, spec calculation by simulation, etc. for these design candidates, and obtains design results that are effective in determining the suitability of adoption of the design candidates. At least a part of this design result is data different from the above required specifications.
  • the manufacturing cost of the design candidate, the test result (hardness), the test result (tensile), and the like are the design results, but the design results are not limited to these.
  • the design result input unit 104 provides the operator with means for inputting these design results.
  • the network structure calculation unit 105 finds a group in which the input design result should be classified and sets it as a related group. Preferably, the network structure calculation unit 105 acquires the design result of the design history belonging to the belonging group and determines the deviation from the design result of the design candidate. The network structure calculation unit 105 obtains a group in which the dissociated design results are classified and sets it as a related group.
  • the calculation result display unit 103 makes it possible to provide the operator with information based on the related group as design support information in addition to the information based on the group to which the calculation result is displayed.
  • the calculation result display unit 103 creates a synthetic group (hereinafter referred to as “networking group”) in which the belonging group and the related group are combined, and presents the information based on the networking group as the design support information.
  • the calculation condition input unit 111 provides the operator with an input means for accepting the calculation conditions of the "data item used as the requirement specification" and the "number of groups to be classified".
  • a plurality of data sets including requirement specifications and design results used or collected in the past are accumulated in the database 112 as a design history.
  • the data analysis unit 113 classifies the set of these data sets into a plurality of groups by a clustering technique such as a self-organization map based on the calculation conditions (requirement specifications, number of groups) of the calculation condition input unit 111.
  • the learning model generation unit 114 gives the requirement specifications to the input layer based on the calculation conditions (requirement specifications, number of groups) of the calculation condition input unit 111, and uses the classification result of the number of groups as the output layer for the learning model in the initial state. Generate.
  • the learning model generation unit 114 adds the network structure of the related group to the output layer of the learning model.
  • the machine learning unit 115 creates learning data by assigning the classification result of the group of the data analysis unit 113 as a teacher value to the past required specifications based on the design history in the database 112.
  • the machine learning unit 115 adds the network structure of the related group as a teacher value to the learning data.
  • the machine learning unit 115 uses the created learning data to perform machine learning of the learning model.
  • the classification unit 102 gives the requirement specifications input by the requirement specification input unit 101 to the input of the learning model that has undergone machine learning.
  • the classification unit 102 estimates the group to which the required specifications should be classified based on the processing result of this learning model.
  • the classification unit 102 estimates the network structure of the related group in addition to the estimation of the belonging group based on the processing result of this learning model. You can also do it together.
  • Such a design support system is realized by a computer system 120 equipped with a CPU (Central Processing Unit), a memory, and the like as hardware.
  • a CPU Central Processing Unit
  • the computer system 120 functions as various configurations of the design support system described above.
  • a dedicated device For a part or all of the computer system 120, a dedicated device, a general-purpose machine learning machine, a DSP (Digital Signal Processor), an FPGA (Field-Programmable Gate Array), a GPU (Graphics Processing Unit), and a PLD (programmable logic). It may be replaced with device) or the like.
  • DSP Digital Signal Processor
  • FPGA Field-Programmable Gate Array
  • GPU Graphics Processing Unit
  • PLD programmable logic
  • one or more client terminals may use the design support system exclusively or jointly via the network by centralizing or distributing a part or all of the hardware to the servers on the network and arranging them in the cloud. ..
  • FIG. 2 is a diagram of inputting calculation conditions necessary for machine learning and analyzing past design information for machine learning.
  • the second phase shown in FIG. 3 is a phase in which required specifications are input and design support is provided.
  • the third phase shown in FIG. 4 is a phase in which the result of the design designed in Phase 2 is input, group networking is performed, and additional learning of the learning model is performed accordingly.
  • Phase 1 processing procedure The operation of Phase 1 will be described along with the step numbers shown in FIG. First, in steps S101 to S104, work of determining calculation conditions for performing machine learning is performed.
  • Step S101 The calculation condition input unit 111 acquires the past design information (design history) from the database 112.
  • Step S102 The calculation condition input unit 111 creates an input screen of the calculation conditions necessary for machine learning based on the data items of the design history, and provides the input screen to the operator.
  • FIG. 5 is a diagram showing an example of an input screen for calculation conditions.
  • the operator inputs (or selects) "metal material" as the learning model name on this input screen.
  • the input screen has two configurations, the first is “selection of required specification items (variables)" and the second is "the number of groups for dividing the design history”.
  • step S101 the item name of the past design history obtained in step S101 is displayed. For example, “vertical dimension”, “horizontal dimension”, “radius”, “component”, “melting drawing NO”, “hot drawing NO”, “cold drawing NO”, “test result (hardness)”, “ “Test result (tensile)” and “Manufacturing cost” are displayed.
  • the operator selects the required specification items (variables) for data analysis of the past design history.
  • Step S103 The calculation condition input unit 111 acquires the calculation condition information input in step S102 via the input screen.
  • Step S104 The calculation condition input unit 111 registers the information acquired in step S103 in the database 112.
  • steps S201 to S207 data analysis as a preparation for machine learning and machine learning based on the data analysis are performed.
  • Step S201 The data analysis unit 113 acquires the past design history from the database 112.
  • This design history is a set of data sets including data values of required specifications, manufacturing costs of design products, and data values of design results.
  • Step S202 The data analysis unit 113 acquires the calculation conditions required for machine learning input in step S103 from the database 112.
  • the data analysis unit 113 acquires the item names of "vertical dimension”, “horizontal dimension”, “radius”, and “component”, and the number of divisions "10" of the group as calculation conditions.
  • Step S203 The data analysis unit 113 analyzes the data based on the acquired information and groups the past design history into groups.
  • a plurality of methods for grouping are known as clustering techniques and are not particularly limited, but here, a method called a self-organizing map is used.
  • a self-organizing map is a type of machine learning technique such as a neural network that models the visual cortex of the cerebral cortex.
  • weight vectors are randomly initially arranged on the map as many as the number of group divisions, and one input vector is prepared based on the requirement specifications included in the design history data set. Calculate the similarity to the input vector for all weight vectors on the map. Euclidean distance is used for similarity. Each time the distance between each vector is found, the weight vector in the vicinity is changed by the following equation.
  • Wu is a weight vector
  • ⁇ is a neighborhood radius
  • ⁇ is a learning coefficient
  • U is an input vector
  • n is the number of repetitions
  • t is the processing number of the input vector.
  • Step S204 The calculation result display unit 103 provides the operator with the analysis result of the grouping in step S203.
  • FIG. 6 shows an example of a display screen of the analysis result of grouping.
  • the past design history of "vertical dimension”, “horizontal dimension”, “radius”, and “component” shows highly similar data such as "A” group to "J” group, respectively. It is divided into 10 groups and displayed. If the result is acceptable, the operator presses the "OK” button, and if the result is acceptable, the operator presses the "Redo” button to review the calculation conditions from step S101.
  • Step S205 The learning model generation unit 114 generates a learning model before performing machine learning.
  • the learning model is not particularly limited, but a neural network is used here.
  • a neural network is a mathematical model that aims to represent the characteristics of a brain consisting of a large number of nerve cells by computer simulation.
  • Neural network placing the i-th layer of one or more artificial neuron and X i, given by a recurrence formula, such as the following equation.
  • a i and Bi are weight parameters and bias parameters, respectively.
  • f is an activation function.
  • X 1 is an input layer
  • X 2 is an intermediate layer
  • X 3 is an output layer.
  • a network with multiple intermediate layers is called a deep neural network.
  • Step S206 The machine learning unit 115 creates learning data for machine learning the learning model.
  • step S203 by assigning the result of grouping by step S203 as a teacher value to the data values of the required specifications (“vertical dimension”, “horizontal dimension”, “radius”, “component”) extracted from the past design history. , Generate a data group for training. Learning data is generated by collecting these data groups for learning.
  • the machine learning unit 115 gives learning data to the learning model created in step S205, performs machine learning such as an error back propagation method, and determines weight parameters and bias parameters for each artificial neuron.
  • Step S207 The machine learning unit 115 registers the information generated in Phase 1 in the database 112.
  • Phase 1 the learning model used by the classification unit 102 is ready.
  • Step S301 The requirement specification input unit 101 provides the operator with a requirement specification input screen.
  • the operator sets a rough target and inputs the desired required specifications.
  • FIG. 7 shows an example of an input screen of the required specifications.
  • the operator inputs the required specifications for manufacturing the metal material.
  • the vertical dimension "400.0” the horizontal dimension "600.0"
  • the radius "0.0” of the metal steel piece the carbon "0.12%” which is a component of the metal material
  • the chromium "4.1%” which is a component of the metal material
  • Molybdenum "4.25%” is input.
  • Step S302 The requirement specification input unit 101 acquires the requirement specification input in step S301.
  • Step S303 The classification unit 102 gives the requirement specifications input by the operator to the input layer of the learning model created in Phase 1. Based on the processing result of the learning model, the classification unit 102 estimates the belonging groups similar to the required specifications from the groups A to J.
  • Step S304 If the learning model has already learned about the network structure (for example, a related group, a group of related groups, and their cooperative relationship), the classification unit 102 estimates the network structure based on the processing result of the learning model. .. The details of this operation will be described later. In the first stage, the learning model does not learn about the network structure, so the learning model does not estimate the network structure.
  • the network structure for example, a related group, a group of related groups, and their cooperative relationship
  • Step S305 The calculation result display unit 103 acquires and displays the design support information about the belonging group estimated in step S303 and, if possible, the design support information about the related group related to the network from the database 112.
  • FIG. 8 shows an example of a design support information display screen.
  • the affiliation group having a high degree of similarity to the requirement specifications input by the operator is presumed to be "B" here, and "affiliation group: B” is displayed.
  • a scatter plot with past design results as coordinate axes is displayed for the group to which it belongs.
  • data plots are performed on the data group of the belonging group "B” with, for example, the X-axis "manufacturing cost” and the Y-axis "test result (hardness)" as coordinate axes. The operator selects a plot close to the desired design result by clicking on the scatter plot.
  • the calculation result display unit 103 acquires the information of the past design history corresponding to the selected plot from the database 112 and displays it on the display screen.
  • the melting drawing NO “A144”, the hot drawing NO “B247”, the cold drawing NO “C369”, the manufacturing cost “142”, the test result (hardness) “244”, the test result (tensile) “ 724 "and the like are displayed.
  • the required specifications collected in S301 are also displayed. The operator once determines the required specifications of the desired design candidate by comparing or referring to the design support information and the required specifications.
  • Step S306 The classification unit 102 registers the information obtained in Phase 2 in the database 112.
  • Phase 3 processing procedure Subsequently, the operation of Phase 3 will be described along with the step numbers shown in FIG. In steps S401 to S409, further design support for the design candidate and additional learning of the network structure for the learning model are carried out.
  • Step S401 The operator is effective in determining the suitability of adopting the design candidate by performing measurement experiments of prototypes, cost calculation, spec calculation by simulation, etc. for the design candidate once determined in Phase 2. Obtain data (hereinafter referred to as "design result").
  • the design result input unit 104 provides the operator with means for inputting these design results.
  • FIG. 9 shows an example of a design result input screen.
  • the design result input unit 104 acquires information (required specifications, design result) input by the operator on the input screen.
  • Step S403 The classification unit 102 inputs the input requirement specifications to the learning model, and estimates the group to which the design candidate requirement specifications should be classified based on the processing result of the learning model.
  • Step S404 If the learning model has already learned about the network structure, the classification unit 102 estimates the network structure based on the processing result of the learning model. The details of this operation will be described later. At this stage of explanation, the learning model does not learn about the network structure, so the learning model does not estimate the network structure.
  • Step S405 The calculation result display unit 103 acquires the design support information about the belonging group estimated in step S403, and if possible, the design support information about the related group related to the network from the database 112, and is input in step S401. Display with the design result.
  • FIG. 10 shows an example of a design result display screen. Here, duplicate description of the same display element as in FIG. 8 will be omitted.
  • plots " ⁇ " corresponding to the design results of the belonging group B plots " ⁇ " corresponding to the newly input design results of the design candidates are additionally displayed. Will be done.
  • FIG. 10 shows a result (hereinafter referred to as “unexpected”) in which the manufacturing cost data value is far from the manufacturing cost data group of the belonging group B among the design results of the design candidates.
  • the network structure calculation unit 105 automatically detects unexpected design results from the belonging group such as this manufacturing cost by calculating the distance on the scatter plot. In this case, the calculation result display unit 103 adds a “check mark” to the manufacturing cost column shown in FIG. 10 to highlight it, and notifies the operator that there is an unexpected design result.
  • the operator can visually judge the unexpected design result by visually observing the scatter plot and input a "check mark" as shown in the manufacturing cost column shown in FIG.
  • the calculation result display unit 103 acquires information on the unexpected design result determined by the operator via the display screen.
  • the operator determines whether or not further design support (design support by network learning) is necessary for unexpected design results, and the buttons on the display screen (“learn by networking”, “learn without networking”, Selectively operate "Cancel”).
  • Step S406 When "learning without networking" is selected in the previous step S405, the calculation result display unit 103 displays the network structure of the belonging group estimated in step S403.
  • FIG. 11 shows an example of a display screen of the network structure without networking. Since there is no networking, only the belonging group "B" is highlighted as shown in FIG.
  • the network structure calculation unit 105 determines the unexpected design result (manufacturing cost in FIG. 10) selected in step S405 and the design result within the group (in FIG. 10). For example, a group having a small deviation from the average value, mode value, median value, deviation, etc. of manufacturing costs is specified as a related group.
  • the network structure calculation unit 105 may determine the deviation of the design results based on the Euclidean distance or the weighted evaluation value.
  • the network structure calculation unit 105 may set one group having the smallest deviation from the unexpected design result as the related group. Further, the network structure calculation unit 105 may specify a plurality of groups whose deviation from the unexpected design result is smaller than a predetermined value and form a group of related groups.
  • the calculation result display unit 103 displays the network structure of the belonging group and the related group obtained in this way.
  • FIG. 12 shows an example of a display screen of a network structure in the case of networking.
  • the group “I” closest to the manufacturing cost, which is an unexpected design result is calculated, and the related group is specified as "I".
  • the calculation result display unit 103 displays the cooperation relationship of this network structure by the arrow mark from the belonging group "B” to the related group "I".
  • the design support screen shown in FIG. 13 may be displayed based on the processing result of the network structure calculation unit 105.
  • the calculation result display unit 103 displays the information of the networked groups “B, I”, which is a combination of the belonging group B and the related group I, as the design support information at once.
  • Step S407 The learning model generation unit 114 reflects the operation of the design support up to now in the learning model. (In the case of no networking, the same as steps S205 to S206 described above except that the latest training data is used, and thus duplicate description is omitted here.)
  • the learning model generation unit 114 adds an artificial neuron for estimating the network structure to the existing learning model.
  • the learning model generation unit 114 uses the required specifications (“vertical dimension”, “horizontal dimension”, “radius”, “component”) as the input layer of the learning model. Further, the learning model generation unit 114 adds a network structure of related groups (for example, a related group, a group of related groups, a cooperative relationship, etc.) to the output layer in addition to the output layer showing the classification result of the group to which the learning model belongs. This is a new output layer for the learning model. Depending on the newly added output layer, a network of artificial neurons is added to the middle layer.
  • a network structure of related groups for example, a related group, a group of related groups, a cooperative relationship, etc.
  • the machine learning unit 115 further adds the network structure of the related group as a teacher value to the learning model similar to step S206 described above.
  • Step S408 The machine learning unit 115 gives the latest learning data to the latest learning model, performs machine learning such as an error back propagation method, and determines weight parameters and bias parameters for each artificial neuron.
  • Step S409 The machine learning unit 115 registers the information generated in Phase 3 in the database 112.
  • the learning model can perform machine learning about the network structure of the related group.
  • Step S304 The classification unit 102 inputs the requirement specifications input to the requirement specification input unit 101 to the input layer of the learning model that has learned the network structure.
  • the classification unit 102 estimates the classification of the belonging group and the network structure of the related group based on the processing result of the learning model.
  • Step S305 The calculation result display unit 103 acquires and displays the design support information regarding the belonging group estimated in the previous step and the design support information based on the network structure of the related group from the database 112.
  • FIG. 13 is a diagram showing an example of a display screen of design support by the networking group. In the figure, duplicate description will be omitted for the display elements similar to those in FIG.
  • the belonging group having a high degree of similarity to the requirement specifications input by the operator is estimated to be “B” by the learning model.
  • the related group in which the design results are estimated to be similar based on the required specifications is estimated to be "I”.
  • the calculation result display unit 103 displays "affiliation group: B" and "related group: I" as the estimation result of the learning model.
  • the operator selects a plot close to the desired design result by clicking on this scatter plot.
  • the calculation result display unit 103 acquires the information of the past design history corresponding to the selected plot from the database 112 and displays it on the display screen.
  • melting drawing NO “A154”, hot drawing NO “B264”, cold drawing NO “C144”, manufacturing cost “240”, test result (hardness) “228”, test result (tensile) “ 710 "and the like are displayed.
  • the requirement specifications collected earlier are also displayed. The operator determines the required specifications of the desired design candidate from a wider range of options by referring to the design support information by these networking groups "B, I" and the individual required specifications.
  • Example 1 Effect of Example 1 (1)
  • the classification unit 102 obtains a group in which the required specifications are classified as a belonging group from the group of groups in which the past design history is classified.
  • the calculation result display unit 103 makes it possible to present information based on the belonging group as design support information. Further, the calculation result display unit 103 acquires information about the design candidates designed based on the required specifications and the group in which the design results different from the required specifications are classified (hereinafter referred to as “related groups”), and designs the design. Make it possible to present it as support information. Therefore, not only the information of the group to which the employee belongs but also the information of the related group can be utilized as the design support information. Therefore, in the first embodiment, the oversight of the design support information is reduced by the amount of the information of the related group as compared with the conventional technique, and it becomes possible to provide the operator with a wider range of design support information.
  • the calculation result display unit 103 creates a networked group by synthesizing the belonging group and the related group, and provides the operator with the information based on the networked group as the design support information ((2). See FIG. 13). Therefore, the operator can list the design support information of the entire networked group (that is, obtain it all at once) without being aware of the troublesome things such as the difference between the affiliated group and the related group.
  • the network structure calculation unit 105 obtains the related group based on the input design result. At this stage, the learning model is not needed yet. Therefore, even in the initial stage of system operation in which the learning data cannot be sufficiently collected and the learning model cannot be sufficiently machine-learned, the network structure calculation unit 105 makes it possible to support the design using the related groups.
  • the network structure calculation unit 105 acquires the design result of the design history belonging to the belonging group and determines the deviation from the design result of the design candidate.
  • the network structure calculation unit 105 obtains a group showing a design result that is close to (that is, does not deviate) from the design result determined to be dissociated as a related group. Therefore, the related group is searched only when there is a design result that deviates from the belonging group, and the search of the related group does not expand unnecessarily. Therefore, it is possible to narrow down the necessary design support information and provide it to the operator.
  • the data analysis unit 113 clusters a set of data sets including past requirement specifications and design results based on the requirement specifications and classifies them into a plurality of groups.
  • the learning model generation unit 114 generates a learning model for performing this group classification.
  • the machine learning unit 115 creates learning data using the classification result of the data analysis unit 113 as a teacher value and the past required specifications as an input value, and performs machine learning of the learning model. Therefore, the classification unit 102 can estimate the group to which the requirement specifications should be classified based on the processing result of the learning model by giving the requirement specifications acquired by the requirement specification input unit 101 to the input of the learning model. become.
  • the learning model generation unit 114 generates a learning model in which the required specifications are input and the network structure of the belonging group and the related group is output.
  • the machine learning unit 115 further adds the network structure of the related group by the network structure calculation unit as a teacher value to the learning data, and performs machine learning of the learning model. Therefore, by giving the requirement specifications acquired by the requirement specification input unit 101 to the input of the learning model, the classification unit 102 can estimate the affiliation group to which the requirement specifications should be classified and the network structure of the related group. Become.
  • a self-organizing map is used for grouping, but the present invention is not limited to this.
  • another clustering technique such as K-means may be used.
  • the design support work may be shared and carried out by a plurality of hardware and a plurality of software.
  • one or more artificial neurons for estimating the design result input in step S402 may be arranged in the middle layer of the learning model. Such an artificial neuron may perform machine learning using the required specifications as inputs and the design result input in step S402 as a teacher value. By arranging the artificial neurons that estimate the design result in the middle layer in this way, it becomes possible to add the estimation result of the design result to the estimation of the network structure of the related group, and the estimation accuracy of the network structure of the related group. It becomes possible to increase.
  • a case where one related group "I" is required is described as a network structure of related groups in order to simplify the explanation.
  • the present invention is not limited to this.
  • a plurality of related groups spreading horizontally may be obtained as a network structure for the belonging group.
  • the related group may be further requested.
  • a group of related groups having a tree structure may be obtained for the belonging group. It is preferable to appropriately select these according to the circumstances of the design target and the like.
  • a neural network was used as a learning model.
  • learning models include decision tree learning, correlation rule learning, genetic programming, inductive logic programming, support vector machines, clustering, Basian networks, reinforcement learning, expression learning, principal component analysis, and extreme learning machines.
  • a learning model based on at least one of the other machine learning techniques may be adopted.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Geometry (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Hardware Design (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The objective of the present invention is to provide a technique for assisting design. One representative design assistance system according to the present invention comprises: a requested specification input unit for inputting a requested specification for a design; a classification unit that determines a group (hereafter "affiliation group") into which the requested specification is classified, from among a group of groups in which past design history is classified; and a calculation result display unit that acquires information about the affiliation group and makes it possible to present the information as design assistance information, the calculation result display unit acquiring information about a group (hereafter "relationship group") into which design results other than the requested specification are classified, for design candidates designed on the basis of the requested specification, and making it possible to present the information as design assistance information.

Description

設計支援システム、設計支援方法および設計支援プログラムDesign support system, design support method and design support program
　本発明は、設計支援システム、設計支援方法および設計支援プログラムに関する。 The present invention relates to a design support system, a design support method, and a design support program.
　従来、設計作業を支援するための技術が提案されている。 Conventionally, technologies for supporting design work have been proposed.
　例えば、特許文献１は、「学習モデルを用いてＣＡＤの設計データからＣＡＥ（Computer Aided Engineering）用の解析モデルを生成する設計支援の技術」を開示する。 For example, Patent Document 1 discloses "a design support technique for generating an analysis model for CAE (Computer Aided Engineering) from CAD design data using a learning model".
　また、特許文献２は、「設計データから抽出される特徴データに基づいて、類似する設計データの分類絞込みを行い、関連のある設計情報として表示する設計支援の技術」を開示する。 Further, Patent Document 2 discloses "a design support technique for narrowing down the classification of similar design data based on the feature data extracted from the design data and displaying it as related design information".
特開２０１９－００３３２４号公報JP-A-2019-003324 特開２００８－１７６４６４号公報JP-A-2008-176464
　特許文献１では、過去の設計データおよび解析モデルに対して、ユーザの良否評価を教師値として付与したデータ群を学習データとして準備する。この学習データを用いて学習モデルの機械学習が行われる。このような学習モデルでは、十分な量の学習データを用いて機械学習を行えば、良好な出力結果を得やすい。しかしながら、学習データが不足する初期段階においては、機械学習が十分ではないため良好な出力結果は得られない。特許文献１には、このような課題やその解決策について具体的な開示は見当たらない。 In Patent Document 1, a data group in which a user's quality evaluation is given as a teacher value to past design data and an analysis model is prepared as learning data. Machine learning of the learning model is performed using this learning data. In such a learning model, good output results can be easily obtained if machine learning is performed using a sufficient amount of learning data. However, in the initial stage where the training data is insufficient, good output results cannot be obtained because the machine learning is not sufficient. No specific disclosure of such a problem or its solution is found in Patent Document 1.
　特許文献２では、特徴データの類似検索により設計データを過去の類似する設計データのグループに分類する。しかし、類似性が高いと分類されたグループが、適切なグループでない場合がある。特に、特徴データに基づく類似検索では、特徴データ以外の関係において参照すべきグループが存在する場合があり、それらグループを設計者は見落としてしまうという問題がある。特許文献２には、このような課題やその解決策について具体的な開示は見当たらない。 In Patent Document 2, design data is classified into a group of similar design data in the past by a similarity search of feature data. However, the groups classified as having high similarity may not be appropriate groups. In particular, in the similar search based on the feature data, there may be groups to be referred to in relations other than the feature data, and there is a problem that the designer overlooks those groups. No specific disclosure is found in Patent Document 2 regarding such a problem and its solution.
　そこで、本発明は、設計を支援するための新たな技術を提供することを目的とする。 Therefore, an object of the present invention is to provide a new technique for supporting the design.
　上記課題を解決するために、代表的な本発明の設計支援システムの一つは、設計の要求仕様を入力するための要求仕様入力部と、過去の設計履歴を分類したグループの群の中から要求仕様が分類されるグループ（以下「所属グループ」という）を求める分類部と、所属グループについて情報を取得して設計支援情報として提示可能にする計算結果表示部とを備え、この計算結果表示部は要求仕様に基づいて設計される設計候補について要求仕様とは別の設計結果が分類されるグループ（以下「関連グループ」という）について情報を取得して設計支援情報として提示可能にする。 In order to solve the above problems, one of the representative design support systems of the present invention is from the requirement specification input unit for inputting the requirement specifications of the design and the group of groups that classify the past design history. It is equipped with a classification unit that obtains a group in which required specifications are classified (hereinafter referred to as "affiliation group") and a calculation result display unit that acquires information about the affiliation group and makes it possible to present it as design support information. Acquires information about a group in which design results different from the required specifications are classified (hereinafter referred to as "related groups") for design candidates designed based on the required specifications, and makes it possible to present them as design support information.
　本発明によれば、設計を支援するための技術が提供される。
　上記した以外の課題、構成及び効果は、以下の実施形態の説明により明らかにされる。 According to the present invention, a technique for supporting the design is provided.
Issues, configurations and effects other than those described above will be clarified by the description of the following embodiments.
図１は、実施例１の全体構成図である。FIG. 1 is an overall configuration diagram of the first embodiment. 図２は、実施例１の処理手順（フェーズ１）を表す図である。FIG. 2 is a diagram showing a processing procedure (phase 1) of the first embodiment. 図３は、実施例１の処理手順（フェーズ２）を表す図である。FIG. 3 is a diagram showing the processing procedure (Phase 2) of the first embodiment. 図４は、実施例１の処理手順（フェーズ３）を表す図である。FIG. 4 is a diagram showing the processing procedure (Phase 3) of the first embodiment. 図５は、計算条件の入力画面の一例を表す図である。FIG. 5 is a diagram showing an example of an input screen for calculation conditions. 図６は、分析結果の表示画面の一例を表す図である。FIG. 6 is a diagram showing an example of an analysis result display screen. 図７は、要求仕様の入力画面の一例を表す図である。FIG. 7 is a diagram showing an example of an input screen of the required specifications. 図８は、設計支援の表示画面の一例を表す図である。FIG. 8 is a diagram showing an example of a design support display screen. 図９は、設計結果の入力画面の一例を表す図である。FIG. 9 is a diagram showing an example of a design result input screen. 図１０は、設計結果の表示画面の一例を表す図である。FIG. 10 is a diagram showing an example of a design result display screen. 図１１は、ネットワーク化無しの場合のネットワーク構造の表示画面の一例を表す図である。FIG. 11 is a diagram showing an example of a display screen of a network structure without networking. 図１２は、ネットワーク化する場合のネットワーク構造の表示画面の一例を表す図である。FIG. 12 is a diagram showing an example of a display screen of a network structure in the case of networking. 図１３は、ネットワーク化した後の設計支援の表示画面の一例を表す図である。FIG. 13 is a diagram showing an example of a design support display screen after networking.
　以下、実施例を、図面を用いて説明する。 Hereinafter, examples will be described with reference to the drawings.
［１］実施例１の構成
　図１は、実施例１の設計支援システムの構成を示す図である。
　同図において、設計支援システムは、要求仕様入力部１０１、分類部１０２、計算結果表示部１０３、設計結果入力部１０４、およびネットワーク構造計算部１０５を備える。さらに詳しくは、分類部１０２の内部には、計算条件入力部１１１、データベース１１２、データ分析部１１３、学習モデル生成部１１４、および機械学習部１１５が設けられる。 [1] Configuration of the first embodiment FIG. 1 is a diagram showing the configuration of the design support system of the first embodiment.
In the figure, the design support system includes a requirement specification input unit 101, a classification unit 102, a calculation result display unit 103, a design result input unit 104, and a network structure calculation unit 105. More specifically, a calculation condition input unit 111, a database 112, a data analysis unit 113, a learning model generation unit 114, and a machine learning unit 115 are provided inside the classification unit 102.
　要求仕様入力部１０１は、設計上の要求仕様を入力するための入力手段を操作者に提供する。ここで、「要求仕様」とは、設計により作り出した対象物が満たすべき項目やデータやスペックなどの仕様であって、設計支援を受けるために特定すべき仕様を意味する。具体的な一例としては、寸法、形状、成分、特性などが要求仕様になるが、これに限定されるものではない。 The requirement specification input unit 101 provides an operator with an input means for inputting a design requirement specification. Here, the "required specification" means a specification such as an item, data, or a specification that an object created by design should satisfy, and a specification that should be specified in order to receive design support. As a specific example, dimensions, shapes, components, characteristics, etc. are required specifications, but the specifications are not limited to these.
　分類部１０２は、過去の設計履歴を分類したグループの群の中から、要求仕様が分類されるべきグループを求めて「所属グループ」とする。 The classification unit 102 searches for a group to which the required specifications should be classified from the group of groups that have classified the past design history, and sets it as the "affiliation group".
　計算結果表示部１０３は、所属グループの情報を取得して、設計支援情報として操作者に提供する。
　操作者は、この設計支援情報に基づいて、手動または自動によって設計候補を一旦決定する。 The calculation result display unit 103 acquires the information of the belonging group and provides it to the operator as design support information.
Based on this design support information, the operator temporarily or automatically determines a design candidate.
　操作者は、この設計候補について、試作品の計測実験や、コスト計算や、シミュレーションによるスペック計算などを実施し、設計候補の採用の適否判断に有効な設計結果を求める。この設計結果の少なくとも一部は、先の要求仕様とは別種のデータである。具体的な一例としては、設計候補の製造コスト、試験結果（硬さ）、試験結果（引張）などが設計結果になるが、これに限定されるものではない。
　設計結果入力部１０４は、これら設計結果を入力するための手段を操作者に提供する。 The operator performs measurement experiments of prototypes, cost calculation, spec calculation by simulation, etc. for these design candidates, and obtains design results that are effective in determining the suitability of adoption of the design candidates. At least a part of this design result is data different from the above required specifications. As a specific example, the manufacturing cost of the design candidate, the test result (hardness), the test result (tensile), and the like are the design results, but the design results are not limited to these.
The design result input unit 104 provides the operator with means for inputting these design results.
　ネットワーク構造計算部１０５は、入力された設計結果が分類されるべきグループを求めて関連グループとする。好ましくは、ネットワーク構造計算部１０５は、所属グループに属する設計履歴の設計結果を取得して設計候補の設計結果との乖離を判定する。ネットワーク構造計算部１０５は、この乖離した設計結果が分類されるグループを求めて、関連グループとする。 The network structure calculation unit 105 finds a group in which the input design result should be classified and sets it as a related group. Preferably, the network structure calculation unit 105 acquires the design result of the design history belonging to the belonging group and determines the deviation from the design result of the design candidate. The network structure calculation unit 105 obtains a group in which the dissociated design results are classified and sets it as a related group.
　計算結果表示部１０３は、所属グループに基づく情報に加えて、関連グループに基づく情報も設計支援情報として操作者に提供可能にする。好ましくは、計算結果表示部１０３は、所属グループと関連グループとを組み合わせた合成グループ（以下「ネットワーク化グループ」という）を作成し、ネットワーク化グループに基づく情報を設計支援情報として提示する。 The calculation result display unit 103 makes it possible to provide the operator with information based on the related group as design support information in addition to the information based on the group to which the calculation result is displayed. Preferably, the calculation result display unit 103 creates a synthetic group (hereinafter referred to as “networking group”) in which the belonging group and the related group are combined, and presents the information based on the networking group as the design support information.
　計算条件入力部１１１は、「要求仕様として使用するデータ項目」および「分類するグループ数」の計算条件を受け付けるための入力手段を操作者に提供する。 The calculation condition input unit 111 provides the operator with an input means for accepting the calculation conditions of the "data item used as the requirement specification" and the "number of groups to be classified".
　データベース１１２には、設計履歴として、過去に使用または収集された要求仕様および設計結果を含むデータセットが複数蓄積される。 A plurality of data sets including requirement specifications and design results used or collected in the past are accumulated in the database 112 as a design history.
　データ分析部１１３は、これらデータセットの集合を、計算条件入力部１１１の計算条件（要求仕様，グループ数）に基づいて、自己組織化マップなどのクラスタリング技法により複数のグループに分類する。 The data analysis unit 113 classifies the set of these data sets into a plurality of groups by a clustering technique such as a self-organization map based on the calculation conditions (requirement specifications, number of groups) of the calculation condition input unit 111.
　学習モデル生成部１１４は、計算条件入力部１１１の計算条件（要求仕様，グループ数）に基づいて、要求仕様を入力層に与え、グループ数の分類結果を出力層とする初期状態の学習モデルを生成する。 The learning model generation unit 114 gives the requirement specifications to the input layer based on the calculation conditions (requirement specifications, number of groups) of the calculation condition input unit 111, and uses the classification result of the number of groups as the output layer for the learning model in the initial state. Generate.
　なお、所属グループと関連グループのネットワーク構造が既知の場合、学習モデル生成部１１４は、関連グループのネットワーク構造を学習モデルの出力層に追加する。 If the network structure of the affiliated group and the related group is known, the learning model generation unit 114 adds the network structure of the related group to the output layer of the learning model.
　機械学習部１１５は、データベース１１２内の設計履歴に基づいて、過去の要求仕様に対し、データ分析部１１３のグループの分類結果を教師値として付与することにより、学習データを作成する。 The machine learning unit 115 creates learning data by assigning the classification result of the group of the data analysis unit 113 as a teacher value to the past required specifications based on the design history in the database 112.
　なお、所属グループと関連グループのネットワーク構造が既知の場合、機械学習部１１５は、関連グループのネットワーク構造を教師値として学習データに追加する。
　機械学習部１１５は、作成した学習データを用いて、学習モデルの機械学習を実施する。 When the network structure of the affiliated group and the related group is known, the machine learning unit 115 adds the network structure of the related group as a teacher value to the learning data.
The machine learning unit 115 uses the created learning data to perform machine learning of the learning model.
　分類部１０２は、機械学習を行った学習モデルの入力に、要求仕様入力部１０１で入力された要求仕様を与える。分類部１０２は、この学習モデルの処理結果に基づいて、要求仕様が分類されるべき所属グループの推定を行う。 The classification unit 102 gives the requirement specifications input by the requirement specification input unit 101 to the input of the learning model that has undergone machine learning. The classification unit 102 estimates the group to which the required specifications should be classified based on the processing result of this learning model.
　なお、関連グループのネットワーク構造について学習モデルの機械学習が済んでいる場合、分類部１０２は、この学習モデルの処理結果に基づいて、所属グループの推定に加えて、関連グループのネットワーク構造の推定を一緒に行うこともできる。 When the machine learning of the learning model has been completed for the network structure of the related group, the classification unit 102 estimates the network structure of the related group in addition to the estimation of the belonging group based on the processing result of this learning model. You can also do it together.
　このような設計支援システムは、ハードウェアとしてＣＰＵ（Central Processing Unit）やメモリなどを備えたコンピュータシステム１２０により実現される。このハードウェアが設計支援プログラムを実行することにより、コンピュータシステム１２０は、上述した設計支援システムの各種構成として機能する。 Such a design support system is realized by a computer system 120 equipped with a CPU (Central Processing Unit), a memory, and the like as hardware. When the hardware executes the design support program, the computer system 120 functions as various configurations of the design support system described above.
　なお、コンピュータシステム１２０の一部または全部については、専用の装置、汎用の機械学習マシン、ＤＳＰ（Digital Signal Processor）、ＦＰＧＡ（Field-Programmable Gate Array）、ＧＰＵ（Graphics Processing Unit）、ＰＬＤ(programmable logic device）などで代替してもよい。 For a part or all of the computer system 120, a dedicated device, a general-purpose machine learning machine, a DSP (Digital Signal Processor), an FPGA (Field-Programmable Gate Array), a GPU (Graphics Processing Unit), and a PLD (programmable logic). It may be replaced with device) or the like.
　また、ハードウェアの一部または全部をネットワーク上のサーバに集中または分散してクラウド配置することにより、一つ以上のクライアント端末がネットワークを介して設計支援システムを専用または共同で使用してもよい。 In addition, one or more client terminals may use the design support system exclusively or jointly via the network by centralizing or distributing a part or all of the hardware to the servers on the network and arranging them in the cloud. ..
［２］実施例１の動作
　上述のように構成される実施例１の設計支援システムについて、動作を説明する。
　図２～図４は、設計支援システムの処理手順を３つのフェーズに分けて説明する図である。
　図２に示す一つ目のフェーズは、機械学習に必要な計算条件の入力と、過去の設計情報を分析して機械学習するフェーズである。
　図３に示す二つ目のフェーズは、要求仕様を入力し、設計支援を行うフェーズである。
　図４に示す三つ目のフェーズは、フェーズ２で設計した設計の結果を入力し、グループのネットワーク化とそれに伴う学習モデルの追加学習を行うフェーズである。 [2] Operation of the first embodiment The operation of the design support system of the first embodiment configured as described above will be described.
2 to 4 are diagrams for explaining the processing procedure of the design support system by dividing it into three phases.
The first phase shown in FIG. 2 is a phase of inputting calculation conditions necessary for machine learning and analyzing past design information for machine learning.
The second phase shown in FIG. 3 is a phase in which required specifications are input and design support is provided.
The third phase shown in FIG. 4 is a phase in which the result of the design designed in Phase 2 is input, group networking is performed, and additional learning of the learning model is performed accordingly.
　以下、機械構造物を構成する金属材料の製造設計を例に取り、金属材料の製造コスト、使用した図面番号、性能などを算出する方法について説明する。 Hereinafter, the method of calculating the manufacturing cost of the metal material, the drawing number used, the performance, etc. will be described by taking the manufacturing design of the metal material constituting the mechanical structure as an example.
［２－１］フェーズ１の処理手順
　図２に示すステップ番号に沿って、フェーズ１の動作を述べる。
　まず、ステップＳ１０１～Ｓ１０４では、機械学習を行うための計算条件の確定作業が行われる。 [2-1] Phase 1 processing procedure The operation of Phase 1 will be described along with the step numbers shown in FIG.
First, in steps S101 to S104, work of determining calculation conditions for performing machine learning is performed.
ステップＳ１０１：　計算条件入力部１１１は、データベース１１２から過去の設計情報（設計履歴）を取得する。 Step S101: The calculation condition input unit 111 acquires the past design information (design history) from the database 112.
ステップＳ１０２：　計算条件入力部１１１は、設計履歴のデータ項目に基づいて機械学習に必要な計算条件の入力画面を作成し、操作者に提供する。
　図５は、計算条件の入力画面の一例を示す図である。 Step S102: The calculation condition input unit 111 creates an input screen of the calculation conditions necessary for machine learning based on the data items of the design history, and provides the input screen to the operator.
FIG. 5 is a diagram showing an example of an input screen for calculation conditions.
　例えば、操作者は、この入力画面において、学習モデル名として「金属材料」を入力（または選択）する。さらに、入力画面は二つの構成となっており、一つ目は「要求仕様の項目（変数）の選択」、二つ目は「設計履歴を分割するグループ数」となっている。 For example, the operator inputs (or selects) "metal material" as the learning model name on this input screen. Furthermore, the input screen has two configurations, the first is "selection of required specification items (variables)" and the second is "the number of groups for dividing the design history".
　「要求仕様の項目（変数）の選択」では、ステップＳ１０１で入手した過去の設計履歴の項目名が表示される。例えば、「縦寸法」、「横寸法」、「半径」、「成分」、「溶解図面ＮＯ」、「熱間図面ＮＯ」、「冷間図面ＮＯ」、「試験結果（硬さ）」、「試験結果（引張）」、「製造コスト」が表示される。 In "Selection of required specification item (variable)", the item name of the past design history obtained in step S101 is displayed. For example, "vertical dimension", "horizontal dimension", "radius", "component", "melting drawing NO", "hot drawing NO", "cold drawing NO", "test result (hardness)", " "Test result (tensile)" and "Manufacturing cost" are displayed.
　操作者は、過去の設計履歴について、データ分析するための要求仕様の項目（変数）を選択する。 The operator selects the required specification items (variables) for data analysis of the past design history.
　ここでは、「縦寸法」、「横寸法」、「半径」、「成分」が選択されている。
　「グループ数」では、過去の設計履歴をグループ分けするグループの数を入力する。例えば、「１０」が入力される。 Here, "vertical dimension", "horizontal dimension", "radius", and "component" are selected.
In "Number of groups", enter the number of groups for grouping the past design history. For example, "10" is input.
ステップＳ１０３：　計算条件入力部１１１は、ステップＳ１０２で入力された計算条件の情報を入力画面を介して取得する。 Step S103: The calculation condition input unit 111 acquires the calculation condition information input in step S102 via the input screen.
ステップＳ１０４：　計算条件入力部１１１は、ステップＳ１０３で取得した情報を、データベース１１２に登録する。 Step S104: The calculation condition input unit 111 registers the information acquired in step S103 in the database 112.
　続いて、ステップＳ２０１～Ｓ２０７では、機械学習の準備としてのデータ分析と、それに基づく機械学習が行われる。 Subsequently, in steps S201 to S207, data analysis as a preparation for machine learning and machine learning based on the data analysis are performed.
ステップＳ２０１：　データ分析部１１３は、データベース１１２から過去の設計履歴を取得する。この設計履歴は、要求仕様のデータ値、設計品の製造コスト、および設計結果のデータ値などを含むデータセットの集合である。 Step S201: The data analysis unit 113 acquires the past design history from the database 112. This design history is a set of data sets including data values of required specifications, manufacturing costs of design products, and data values of design results.
ステップＳ２０２：　データ分析部１１３は、ステップＳ１０３で入力された機械学習に必要な計算条件をデータベース１１２から取得する。
　ここでは、データ分析部１１３は、計算条件として、「縦寸法」、「横寸法」、「半径」、「成分」の項目名、およびグループの分割数「１０」を取得する。 Step S202: The data analysis unit 113 acquires the calculation conditions required for machine learning input in step S103 from the database 112.
Here, the data analysis unit 113 acquires the item names of "vertical dimension", "horizontal dimension", "radius", and "component", and the number of divisions "10" of the group as calculation conditions.
ステップＳ２０３：　データ分析部１１３は、取得した情報に基づいてデータ分析を行い、過去の設計履歴をグループ分けする。
　グループ分けする方法は、クラスタリング技法として複数知られており特に限定されないが、ここでは自己組織化マップと呼ばれる方法を使用する。
　自己組織化マップは、ニューラルネットワークなどの機械学習技法の一種であり、大脳皮質の視覚野をモデル化したものである。 Step S203: The data analysis unit 113 analyzes the data based on the acquired information and groups the past design history into groups.
A plurality of methods for grouping are known as clustering techniques and are not particularly limited, but here, a method called a self-organizing map is used.
A self-organizing map is a type of machine learning technique such as a neural network that models the visual cortex of the cerebral cortex.
　自己組織化マップは、グループ分割の数だけ重みベクトルをマップ上にランダムに初期配置し、設計履歴のデータセットに含まれる要求仕様に基づいて入力ベクトルを一つ用意する。マップ上の全ての重みベクトルに対して、入力ベクトルとの類似度を計算する。類似度にはユークリッド的な距離を用いる。
　各ベクトル間の距離が最も小さいものを見つけるたびに、その近傍の重みベクトルを下式により変更する。
Figure JPOXMLDOC01-appb-M000001
 In the self-organizing map, weight vectors are randomly initially arranged on the map as many as the number of group divisions, and one input vector is prepared based on the requirement specifications included in the design history data set. Calculate the similarity to the input vector for all weight vectors on the map. Euclidean distance is used for similarity.
Each time the distance between each vector is found, the weight vector in the vicinity is changed by the following equation.
Figure JPOXMLDOC01-appb-M000001
　ここでＷｕは重みベクトル、θは近傍半径、αは学習係数、Ｕは入力ベクトル、ｎは繰り返し回数、ｔは入力ベクトルの処理番号を表す。
　このように自己組織化マップを用いて、過去の設計履歴のデータセットを、要求仕様の類似性の高いグループに分ける。 Here, Wu is a weight vector, θ is a neighborhood radius, α is a learning coefficient, U is an input vector, n is the number of repetitions, and t is the processing number of the input vector.
In this way, the self-organizing map is used to divide the data set of the past design history into groups with high similarity of the required specifications.
ステップＳ２０４：　計算結果表示部１０３は、ステップＳ２０３におけるグループ分けの分析結果を操作者に提供する。 Step S204: The calculation result display unit 103 provides the operator with the analysis result of the grouping in step S203.
　図６は、グループ分けの分析結果の表示画面の一例を示す。
　同図において、表示画面には、「縦寸法」、「横寸法」、「半径」、「成分」の過去の設計履歴が、「Ａ」グループから「Ｊ」グループといった類似性の高いデータをそれぞれ集めた１０個のグループに分けて表示される。
　操作者は、この結果で良ければ「決定」ボタンを押し、やり直す場合は「やり直し」ボタンを押してステップＳ１０１から計算条件の見直しを行う。 FIG. 6 shows an example of a display screen of the analysis result of grouping.
In the figure, on the display screen, the past design history of "vertical dimension", "horizontal dimension", "radius", and "component" shows highly similar data such as "A" group to "J" group, respectively. It is divided into 10 groups and displayed.
If the result is acceptable, the operator presses the "OK" button, and if the result is acceptable, the operator presses the "Redo" button to review the calculation conditions from step S101.
ステップＳ２０５：　学習モデル生成部１１４は、機械学習を行う前の学習モデルを生成する。
　学習モデルは特に限定されないが、ここではニューラルネットワークを用いる。 Step S205: The learning model generation unit 114 generates a learning model before performing machine learning.
The learning model is not particularly limited, but a neural network is used here.
　ニューラルネットワークは、多数の神経細胞からなる脳の特性を計算機上のシミュレーションで表現することを目的とした数学モデルである。ニューラルネットワークは、一つ以上の人工ニューロンからなるｉ番目の層をＸｉと置くと、下式のような漸化式で与えられる。
Figure JPOXMLDOC01-appb-M000002
 A neural network is a mathematical model that aims to represent the characteristics of a brain consisting of a large number of nerve cells by computer simulation. Neural network, placing the i-th layer of one or more artificial neuron and X i, given by a recurrence formula, such as the following equation.
Figure JPOXMLDOC01-appb-M000002
　ここでＡｉ、Ｂｉはそれぞれ重みパラメータ、バイアスパラメータである。ｆは活性化関数である。なお３層の場合はＸ１が入力層、Ｘ２が中間層、Ｘ３が出力層となる。中間層が複数あるものをディープニューラルネットと呼ぶ。 Here, A i and Bi are weight parameters and bias parameters, respectively. f is an activation function. In the case of three layers, X 1 is an input layer, X 2 is an intermediate layer, and X 3 is an output layer. A network with multiple intermediate layers is called a deep neural network.
　例えば、「縦寸法」、「横寸法」、「半径」、「成分」を要求仕様として入力層とし、データ分析部１１３で得られたグループ分けと同様の結果を出力層とし、その間に中間層を複数設けた学習モデルが作成される。 For example, "vertical dimension", "horizontal dimension", "radius", and "component" are set as input layers, and the same result as the grouping obtained by the data analysis unit 113 is used as an output layer, and an intermediate layer is provided between them. A learning model with a plurality of is created.
ステップＳ２０６：　機械学習部１１５は、学習モデルを機械学習するための学習データを作成する。 Step S206: The machine learning unit 115 creates learning data for machine learning the learning model.
　例えば、過去の設計履歴から抽出した要求仕様のデータ値（「縦寸法」、「横寸法」、「半径」、「成分」）に、ステップＳ２０３によるグループ分けの結果を教師値として付与することにより、学習用のデータ群を生成する。これら学習用のデータ群を収集することにより、学習データが生成される。 For example, by assigning the result of grouping by step S203 as a teacher value to the data values of the required specifications (“vertical dimension”, “horizontal dimension”, “radius”, “component”) extracted from the past design history. , Generate a data group for training. Learning data is generated by collecting these data groups for learning.
　機械学習部１１５は、ステップＳ２０５において作成された学習モデルに対して学習データを与えて誤差逆伝搬法などの機械学習を実施し、人工ニューロンそれぞれについて重みパラメータ、バイアスパラメータを決定する。 The machine learning unit 115 gives learning data to the learning model created in step S205, performs machine learning such as an error back propagation method, and determines weight parameters and bias parameters for each artificial neuron.
ステップＳ２０７：　機械学習部１１５は、フェーズ１で生成された情報をデータベース１１２に登録する。 Step S207: The machine learning unit 115 registers the information generated in Phase 1 in the database 112.
　上述したフェーズ１の動作により、分類部１０２が使用する学習モデルの準備が整う。 By the operation of Phase 1 described above, the learning model used by the classification unit 102 is ready.
［２－２］フェーズ２の処理手順
　続いて、図３に示すステップ番号に沿って、フェーズ２の動作を述べる。
　まず、ステップＳ３０１～Ｓ３０６では、操作者の入力する要求仕様に基づいて設計支援を実施する。 [2-2] Phase 2 processing procedure Subsequently, the operation of Phase 2 will be described along with the step numbers shown in FIG.
First, in steps S301 to S306, design support is provided based on the required specifications input by the operator.
ステップＳ３０１：　要求仕様入力部１０１は、要求仕様の入力画面を操作者に提供する。操作者は、所望する要求仕様について大まかな目標を定めて入力する。
　図７は、要求仕様の入力画面の一例を示す。
　同図において、操作者は、金属材料を製造するための要求仕様を入力する。ここでは、金属鋼片の縦寸法「４００．０」、横寸法「６００．０」、半径「０．０」、金属材料の成分である炭素「０．１２％」、クロム「４．１％」、モリブデン「４．２５％」が入力される。 Step S301: The requirement specification input unit 101 provides the operator with a requirement specification input screen. The operator sets a rough target and inputs the desired required specifications.
FIG. 7 shows an example of an input screen of the required specifications.
In the figure, the operator inputs the required specifications for manufacturing the metal material. Here, the vertical dimension "400.0", the horizontal dimension "600.0", the radius "0.0" of the metal steel piece, the carbon "0.12%" which is a component of the metal material, and the chromium "4.1%". , Molybdenum "4.25%" is input.
ステップＳ３０２：　要求仕様入力部１０１は、ステップＳ３０１で入力された要求仕様を取得する。 Step S302: The requirement specification input unit 101 acquires the requirement specification input in step S301.
ステップＳ３０３：　分類部１０２は、フェーズ１で作成された学習モデルの入力層に操作者が入力した要求仕様を与える。分類部１０２は、学習モデルの処理結果に基づいて、この要求仕様に類似する所属グループをグループＡ～Ｊの中から推定する。 Step S303: The classification unit 102 gives the requirement specifications input by the operator to the input layer of the learning model created in Phase 1. Based on the processing result of the learning model, the classification unit 102 estimates the belonging groups similar to the required specifications from the groups A to J.
ステップＳ３０４：　学習モデルがネットワーク構造（例えば関連グループや関連グループの群やその連携関係など）について学習済みであれば、分類部１０２は、学習モデルの処理結果に基づいて、ネットワーク構造の推定を行う。この動作の詳細については後述する。
　最初の段階では、学習モデルはネットワーク構造について学習していないため、学習モデルではネットワーク構造は推定されない。 Step S304: If the learning model has already learned about the network structure (for example, a related group, a group of related groups, and their cooperative relationship), the classification unit 102 estimates the network structure based on the processing result of the learning model. .. The details of this operation will be described later.
In the first stage, the learning model does not learn about the network structure, so the learning model does not estimate the network structure.
ステップＳ３０５：　計算結果表示部１０３は、ステップＳ３０３で推定した所属グループに関する設計支援情報、また可能であればネットワーク関係のある関連グループに関する設計支援情報をデータベース１１２から取得して表示する。 Step S305: The calculation result display unit 103 acquires and displays the design support information about the belonging group estimated in step S303 and, if possible, the design support information about the related group related to the network from the database 112.
　図８は、設計支援情報の表示画面の一例を示す。
　同図において、操作者が入力した要求仕様と類似度の高い所属グループは、ここでは「Ｂ」と推定され、「所属グループ：Ｂ」が表示される。さらに、所属グループについて、過去の設計結果を座標軸にした散布図が表示される。この散布図には、所属グループ「Ｂ」のデータ群について、例えばＸ軸「製造コスト」およびＹ軸「試験結果（硬さ）」を座標軸にしたデータプロットが行われる。操作者は、この散布図の上で、所望する設計結果に近いプロットをクリック操作などにより選択する。計算結果表示部１０３は、選択されたプロットに相当する過去の設計履歴の情報をデータベース１１２から取得し、表示画面に表示する。図８においては、熔解図面ＮＯ「Ａ１４４」、熱間図面ＮＯ「Ｂ２４７」、冷間図面ＮＯ「Ｃ３６９」、製造コスト「１４２」、試験結果（硬さ）「２４４」、試験結果（引張）「７２４」などが表示される。なお、Ｓ３０１で収集された要求仕様も表示される。操作者は、これら設計支援情報と要求仕様を対比したり参考にしたりすることにより、所望する設計候補の要求仕様を一旦決定する。 FIG. 8 shows an example of a design support information display screen.
In the figure, the affiliation group having a high degree of similarity to the requirement specifications input by the operator is presumed to be "B" here, and "affiliation group: B" is displayed. In addition, a scatter plot with past design results as coordinate axes is displayed for the group to which it belongs. In this scatter plot, data plots are performed on the data group of the belonging group "B" with, for example, the X-axis "manufacturing cost" and the Y-axis "test result (hardness)" as coordinate axes. The operator selects a plot close to the desired design result by clicking on the scatter plot. The calculation result display unit 103 acquires the information of the past design history corresponding to the selected plot from the database 112 and displays it on the display screen. In FIG. 8, the melting drawing NO “A144”, the hot drawing NO “B247”, the cold drawing NO “C369”, the manufacturing cost “142”, the test result (hardness) “244”, the test result (tensile) “ 724 "and the like are displayed. The required specifications collected in S301 are also displayed. The operator once determines the required specifications of the desired design candidate by comparing or referring to the design support information and the required specifications.
ステップＳ３０６：　分類部１０２は、フェーズ２において得られた情報をデータベース１１２に登録する。 Step S306: The classification unit 102 registers the information obtained in Phase 2 in the database 112.
　上述したフェーズ２の動作により、設計支援システムによる第一段階の設計支援と、操作者による設計候補の要求仕様が一旦決定される。 By the operation of Phase 2 described above, the design support of the first stage by the design support system and the required specifications of the design candidate by the operator are once determined.
［２－３］フェーズ３の処理手順
　続いて、図４に示すステップ番号に沿って、フェーズ３の動作を述べる。
　ステップＳ４０１～Ｓ４０９では、設計候補に対する更なる設計支援および学習モデルに対するネットワーク構造の追加学習が実施される。 [2-3] Phase 3 processing procedure Subsequently, the operation of Phase 3 will be described along with the step numbers shown in FIG.
In steps S401 to S409, further design support for the design candidate and additional learning of the network structure for the learning model are carried out.
ステップＳ４０１：　操作者は、フェーズ２で一旦決定した設計候補について、試作品の計測実験や、コスト計算や、シミュレーションによるスペック計算などを実施し、設計候補の採用の適否を判断するのに有効なデータ（以下「設計結果」という）を求める。設計結果入力部１０４は、これら設計結果を入力するための手段を操作者に提供する。 Step S401: The operator is effective in determining the suitability of adopting the design candidate by performing measurement experiments of prototypes, cost calculation, spec calculation by simulation, etc. for the design candidate once determined in Phase 2. Obtain data (hereinafter referred to as "design result"). The design result input unit 104 provides the operator with means for inputting these design results.
　図９は、設計結果の入力画面の一例を示す。
　同図において、入力画面には、設計候補の縦寸法、横寸法、半径、成分といった要求仕様の入力欄と共に、熔解図面ＮＯ、熱間図面ＮＯ、冷間図面ＮＯ、製造コスト、試験結果（硬さ）、試験結果（引張）といった設計結果の入力欄が設けられる。
ステップＳ４０２：　設計結果入力部１０４は、操作者が入力画面に入力した情報（要求仕様，設計結果）を情報取得する。 FIG. 9 shows an example of a design result input screen.
In the figure, on the input screen, along with input fields of required specifications such as vertical dimension, horizontal dimension, radius, and component of design candidates, melting drawing NO, hot drawing NO, cold drawing NO, manufacturing cost, and test result (hardness) An input field for design results such as test results (tensile) is provided.
Step S402: The design result input unit 104 acquires information (required specifications, design result) input by the operator on the input screen.
ステップＳ４０３：　分類部１０２は、入力された要求仕様を学習モデルに入力し、学習モデルの処理結果に基づいて設計候補の要求仕様が分類されるべき所属グループを推定する。 Step S403: The classification unit 102 inputs the input requirement specifications to the learning model, and estimates the group to which the design candidate requirement specifications should be classified based on the processing result of the learning model.
ステップＳ４０４：　学習モデルがネットワーク構造について学習済みであれば、分類部１０２は、学習モデルの処理結果に基づいて、ネットワーク構造の推定を行う。この動作の詳細については後述する。
　この説明の段階では、学習モデルはネットワーク構造について学習していないため、学習モデルではネットワーク構造は推定されない。 Step S404: If the learning model has already learned about the network structure, the classification unit 102 estimates the network structure based on the processing result of the learning model. The details of this operation will be described later.
At this stage of explanation, the learning model does not learn about the network structure, so the learning model does not estimate the network structure.
ステップＳ４０５：　計算結果表示部１０３は、ステップＳ４０３で推定された所属グループに関する設計支援情報、可能であればネットワーク関係のある関連グループに関する設計支援情報をデータベース１１２から取得し、ステップＳ４０１で入力された設計結果と共に表示する。 Step S405: The calculation result display unit 103 acquires the design support information about the belonging group estimated in step S403, and if possible, the design support information about the related group related to the network from the database 112, and is input in step S401. Display with the design result.
　図１０は、設計結果の表示画面の一例を示す。
　ここでは、図８と同一の表示要素についての重複説明を省略する。
　図１０に示す散布図には、所属グループＢの設計結果に対応するプロット「○」の群に加えて、新たに入力された設計候補の設計結果に対応するプロット「●」が追加的に表示される。 FIG. 10 shows an example of a design result display screen.
Here, duplicate description of the same display element as in FIG. 8 will be omitted.
In the scatter plot shown in FIG. 10, in addition to the group of plots "○" corresponding to the design results of the belonging group B, plots "●" corresponding to the newly input design results of the design candidates are additionally displayed. Will be done.
　図１０では、設計候補の設計結果の中で、製造コストのデータ値が、所属グループＢの製造コストのデータ群からかけ離れた結果（以下「予想外」という）を示す。 FIG. 10 shows a result (hereinafter referred to as “unexpected”) in which the manufacturing cost data value is far from the manufacturing cost data group of the belonging group B among the design results of the design candidates.
　ネットワーク構造計算部１０５は、この製造コストのように所属グループからは予想外の設計結果を散布図上の距離計算などにより自動的に検出する。この場合、計算結果表示部１０３は、図１０に示す製造コスト欄に「チェックマーク」を付与して強調表示し、予想外の設計結果があったことを操作者に報知する。 The network structure calculation unit 105 automatically detects unexpected design results from the belonging group such as this manufacturing cost by calculating the distance on the scatter plot. In this case, the calculation result display unit 103 adds a “check mark” to the manufacturing cost column shown in FIG. 10 to highlight it, and notifies the operator that there is an unexpected design result.
　また、操作者が、予想外の設計結果を散布図を目視して判断し、図１０に示す製造コスト欄のように「チェックマーク」を入力することもできる。この場合は、計算結果表示部１０３は、操作者の判断した予想外の設計結果を表示画面を介して情報取得する。 In addition, the operator can visually judge the unexpected design result by visually observing the scatter plot and input a "check mark" as shown in the manufacturing cost column shown in FIG. In this case, the calculation result display unit 103 acquires information on the unexpected design result determined by the operator via the display screen.
　操作者は、予想外の設計結果について更なる設計支援（ネットワーク学習による設計支援）が必要か否かを判断し、表示画面のボタン（「ネットワーク化して学習」、「ネットワーク化無しで学習」、「取消し」）を選択的に操作する。 The operator determines whether or not further design support (design support by network learning) is necessary for unexpected design results, and the buttons on the display screen (“learn by networking”, “learn without networking”, Selectively operate "Cancel").
ステップＳ４０６： 一つ前のステップＳ４０５において「ネットワーク化無しで学習」が選択された場合、計算結果表示部１０３は、ステップＳ４０３で推定した所属グループのネットワーク構造を表示する。 Step S406: When "learning without networking" is selected in the previous step S405, the calculation result display unit 103 displays the network structure of the belonging group estimated in step S403.
　図１１は、ネットワーク化無しの場合のネットワーク構造の表示画面の一例を示す。
　ネットワーク化無しなので、図１１に示すように所属グループ「Ｂ」のみが強調表示されている。 FIG. 11 shows an example of a display screen of the network structure without networking.
Since there is no networking, only the belonging group "B" is highlighted as shown in FIG.
　一方、ステップＳ４０５において「ネットワーク化して学習」が選択された場合、ネットワーク構造計算部１０５は、ステップＳ４０５で選択された予想外の設計結果（図１０では製造コスト）と、グループ内の設計結果（例えば製造コストの平均値、最頻値、中間値、偏差など）との乖離が小さいグループを特定して関連グループとする。予想外の設計結果が複数ある場合、ネットワーク構造計算部１０５は、設計結果の乖離をユークリッド的な距離や重み付きの評価値により判断してもよい。なお、ネットワーク構造計算部１０５は、予想外の設計結果との乖離が最も小さい一つのグループを関連グループとしてもよい。また、ネットワーク構造計算部１０５は、予想外の設計結果との乖離が所定値よりも小さい複数のグループを特定して関連グループの群としてもよい。 On the other hand, when "networking and learning" is selected in step S405, the network structure calculation unit 105 determines the unexpected design result (manufacturing cost in FIG. 10) selected in step S405 and the design result within the group (in FIG. 10). For example, a group having a small deviation from the average value, mode value, median value, deviation, etc. of manufacturing costs is specified as a related group. When there are a plurality of unexpected design results, the network structure calculation unit 105 may determine the deviation of the design results based on the Euclidean distance or the weighted evaluation value. The network structure calculation unit 105 may set one group having the smallest deviation from the unexpected design result as the related group. Further, the network structure calculation unit 105 may specify a plurality of groups whose deviation from the unexpected design result is smaller than a predetermined value and form a group of related groups.
　計算結果表示部１０３は、このように求めた所属グループと関連グループのネットワーク構造を表示する。 The calculation result display unit 103 displays the network structure of the belonging group and the related group obtained in this way.
　図１２は、ネットワーク化する場合のネットワーク構造の表示画面の一例を示す。
　ここでは、予想外の設計結果である製造コストに最も近いグループ「Ｉ」が計算され、関連グループが「Ｉ」に特定される。計算結果表示部１０３は、このネットワーク構造の連携関係を、所属グループ「Ｂ」から関連グループ「Ｉ」への矢印マークにより表示する。 FIG. 12 shows an example of a display screen of a network structure in the case of networking.
Here, the group "I" closest to the manufacturing cost, which is an unexpected design result, is calculated, and the related group is specified as "I". The calculation result display unit 103 displays the cooperation relationship of this network structure by the arrow mark from the belonging group "B" to the related group "I".
　なお、この段階で、ネットワーク構造計算部１０５の処理結果に基づいて、図１３に示す設計支援の画面を表示してもよい。
　図１３では、計算結果表示部１０３は、所属グループＢと関連グループＩとを合成したネットワーク化グループ「Ｂ，Ｉ」の情報を設計支援情報として一度に表示する。 At this stage, the design support screen shown in FIG. 13 may be displayed based on the processing result of the network structure calculation unit 105.
In FIG. 13, the calculation result display unit 103 displays the information of the networked groups “B, I”, which is a combination of the belonging group B and the related group I, as the design support information at once.
ステップＳ４０７：　学習モデル生成部１１４は、現在までの設計支援の動作を、学習モデルに反映させる。（なお、ネットワーク化無しの場合は、最新の学習データを使用する点を除いて、上述したステップＳ２０５～Ｓ２０６と同様のため、ここでの重複説明を省略する。） Step S407: The learning model generation unit 114 reflects the operation of the design support up to now in the learning model. (In the case of no networking, the same as steps S205 to S206 described above except that the latest training data is used, and thus duplicate description is omitted here.)
　ここでは、ステップＳ４０６の処理により、所属グループおよび関連グループのネットワーク構造が既知である。そこで、学習モデル生成部１１４は、そのネットワーク構造を推定するための人口ニューロンを既存の学習モデルに追加する。 Here, the network structure of the belonging group and the related group is known by the process of step S406. Therefore, the learning model generation unit 114 adds an artificial neuron for estimating the network structure to the existing learning model.
　その結果、学習モデル生成部１１４は、要求仕様（「縦寸法」、「横寸法」、「半径」、「成分」）を学習モデルの入力層とする。さらに、学習モデル生成部１１４は、所属グループの分類結果を示す出力層に加え、関連グループのネットワーク構造（例えば、関連グループ・関連グループの群・連携関係などのいずれか）を出力層に追加して、学習モデルの新たな出力層とする。新たに追加された出力層に応じて、中間層には人口ニューロンによるネットワークが追加される。 As a result, the learning model generation unit 114 uses the required specifications (“vertical dimension”, “horizontal dimension”, “radius”, “component”) as the input layer of the learning model. Further, the learning model generation unit 114 adds a network structure of related groups (for example, a related group, a group of related groups, a cooperative relationship, etc.) to the output layer in addition to the output layer showing the classification result of the group to which the learning model belongs. This is a new output layer for the learning model. Depending on the newly added output layer, a network of artificial neurons is added to the middle layer.
　一方、機械学習部１１５は、上述したステップＳ２０６と同様の学習モデルに対して、関連グループのネットワーク構造を教師値として更に追加する。 On the other hand, the machine learning unit 115 further adds the network structure of the related group as a teacher value to the learning model similar to step S206 described above.
ステップＳ４０８：　機械学習部１１５は、最新の学習モデルに対して最新の学習データを与えて誤差逆伝搬法などの機械学習を実施し、人工ニューロンそれぞれについて重みパラメータ、バイアスパラメータを決定する。 Step S408: The machine learning unit 115 gives the latest learning data to the latest learning model, performs machine learning such as an error back propagation method, and determines weight parameters and bias parameters for each artificial neuron.
ステップＳ４０９：　機械学習部１１５は、フェーズ３で生成された情報をデータベース１１２に登録する。 Step S409: The machine learning unit 115 registers the information generated in Phase 3 in the database 112.
［２－４］ネットワーク構造を学習した学習モデルの動作 [2-4] Operation of the learning model that learned the network structure
　上述したフェーズ３では、関連グループのネットワーク構造を機械学習するのに十分な量の学習データを収集することにより、学習モデルは関連グループのネットワーク構造について機械学習が可能になる。 In Phase 3 described above, by collecting a sufficient amount of learning data for machine learning the network structure of the related group, the learning model can perform machine learning about the network structure of the related group.
　このようにネットワーク構造を学習した学習モデルの動作について、先に説明したステップＳ３０４～Ｓ３０５（Ｓ４０４～Ｓ４０５）について説明する。 The operation of the learning model that has learned the network structure in this way will be described in steps S304 to S305 (S404 to S405) described above.
ステップＳ３０４（Ｓ４０４）：　分類部１０２は、ネットワーク構造を学習した学習モデルの入力層に対して、要求仕様入力部１０１に入力された要求仕様を入力する。分類部１０２は、学習モデルの処理結果に基づいて、所属グループの分類、および関連グループのネットワーク構造を推定する。 Step S304 (S404): The classification unit 102 inputs the requirement specifications input to the requirement specification input unit 101 to the input layer of the learning model that has learned the network structure. The classification unit 102 estimates the classification of the belonging group and the network structure of the related group based on the processing result of the learning model.
ステップＳ３０５（Ｓ４０５）：　計算結果表示部１０３は、前ステップで推定した所属グループに関する設計支援情報、関連グループのネットワーク構造に基づく設計支援情報をデータベース１１２から取得して表示する。 Step S305 (S405): The calculation result display unit 103 acquires and displays the design support information regarding the belonging group estimated in the previous step and the design support information based on the network structure of the related group from the database 112.
　図１３は、ネットワーク化グループによる設計支援の表示画面の一例を表す図である。
　同図において、図８と同様の表示要素については、重複説明を省略する。
　図１３において、操作者が入力した要求仕様と類似度の高い所属グループは、学習モデルにより「Ｂ」と推定される。さらに、学習モデルにより、要求仕様に基づいて設計結果が類似するであろうと推定される関連グループは「Ｉ」と推定される。
　計算結果表示部１０３は、学習モデルの推定結果として、「所属グループ：Ｂ」および「関連グループ：Ｉ」を表示する。 FIG. 13 is a diagram showing an example of a display screen of design support by the networking group.
In the figure, duplicate description will be omitted for the display elements similar to those in FIG.
In FIG. 13, the belonging group having a high degree of similarity to the requirement specifications input by the operator is estimated to be “B” by the learning model. Furthermore, according to the learning model, the related group in which the design results are estimated to be similar based on the required specifications is estimated to be "I".
The calculation result display unit 103 displays "affiliation group: B" and "related group: I" as the estimation result of the learning model.
　さらに、所属グループＢと関連グループＩとを合成したネットワーク化グループ「Ｂ，Ｉ」について、過去の設計結果を座標軸にした散布図が表示される。この散布図には、ネットワーク化グループ「Ｂ，Ｉ」のデータ群について、例えばＸ軸「製造コスト」およびＹ軸「試験結果（硬さ）」を座標軸にしたデータプロットが行われる。 Furthermore, for the networked group "B, I" that combines the belonging group B and the related group I, a scatter diagram with the past design results as the coordinate axes is displayed. In this scatter plot, data plots are made for the data groups of the networked groups "B, I", for example, with the X-axis "manufacturing cost" and the Y-axis "test result (hardness)" as coordinate axes.
　操作者は、この散布図の上で、所望する設計結果に近いプロットをクリック操作などにより選択する。計算結果表示部１０３は、選択されたプロットに相当する過去の設計履歴の情報をデータベース１１２から取得し、表示画面に表示する。 The operator selects a plot close to the desired design result by clicking on this scatter plot. The calculation result display unit 103 acquires the information of the past design history corresponding to the selected plot from the database 112 and displays it on the display screen.
　図１３においては、熔解図面ＮＯ「Ａ１５４」、熱間図面ＮＯ「Ｂ２６４」、冷間図面ＮＯ「Ｃ１４４」、製造コスト「２４０」、試験結果（硬さ）「２２８」、試験結果（引張）「７１０」などが表示される。なお、先に収集された要求仕様も表示される。操作者は、これらネットワーク化グループ「Ｂ，Ｉ」による設計支援情報と個々の要求仕様を参考にすることにより、所望する設計候補の要求仕様を更に広範囲の選択肢から決定する。 In FIG. 13, melting drawing NO “A154”, hot drawing NO “B264”, cold drawing NO “C144”, manufacturing cost “240”, test result (hardness) “228”, test result (tensile) “ 710 "and the like are displayed. The requirement specifications collected earlier are also displayed. The operator determines the required specifications of the desired design candidate from a wider range of options by referring to the design support information by these networking groups "B, I" and the individual required specifications.
［３］実施例１の効果
（１）実施例１では、分類部１０２は、過去の設計履歴を分類したグループの群の中から、要求仕様が分類されるグループを所属グループとして求める。計算結果表示部１０３は、所属グループに基づく情報を設計支援情報として提示可能にする。さらに、計算結果表示部１０３は、要求仕様に基づいて設計される設計候補について、要求仕様とは別の設計結果が分類されるグループ（以下「関連グループ」という）について情報を取得して、設計支援情報として提示可能にする。
　そのため、所属グループの情報のみならず、関連グループの情報についても設計支援情報として活用することが可能になる。したがって、実施例１では、関連グループの情報分だけ、従来技術に比べて設計支援情報の見落としが低減し、より広範囲の設計支援情報を操作者に提供することが可能になる。 [3] Effect of Example 1 (1) In Example 1, the classification unit 102 obtains a group in which the required specifications are classified as a belonging group from the group of groups in which the past design history is classified. The calculation result display unit 103 makes it possible to present information based on the belonging group as design support information. Further, the calculation result display unit 103 acquires information about the design candidates designed based on the required specifications and the group in which the design results different from the required specifications are classified (hereinafter referred to as “related groups”), and designs the design. Make it possible to present it as support information.
Therefore, not only the information of the group to which the employee belongs but also the information of the related group can be utilized as the design support information. Therefore, in the first embodiment, the oversight of the design support information is reduced by the amount of the information of the related group as compared with the conventional technique, and it becomes possible to provide the operator with a wider range of design support information.
（２）実施例１では、計算結果表示部１０３は、所属グループと関連グループとを合成してネットワーク化グループを作成し、このネットワーク化グループに基づく情報を設計支援情報として操作者に提供する（図１３参照）。
　したがって、操作者は、所属グループと関連グループとの違いなどの面倒なことを意識することなく、ネットワーク化グループ全体の設計支援情報を一覧する（つまり一度にまとめて得る）ことが可能になる。 (2) In the first embodiment, the calculation result display unit 103 creates a networked group by synthesizing the belonging group and the related group, and provides the operator with the information based on the networked group as the design support information ((2). See FIG. 13).
Therefore, the operator can list the design support information of the entire networked group (that is, obtain it all at once) without being aware of the troublesome things such as the difference between the affiliated group and the related group.
（３）実施例１では、入力される設計結果に基づいて、ネットワーク構造計算部１０５が関連グループを求める。この段階では、学習モデルはまだ必要ない。
　したがって、学習データの収集が不足し、学習モデルを十分に機械学習できないシステム運用の初期段階であっても、ネットワーク構造計算部１０５を備えることにより、関連グループを用いた設計支援が可能になる。 (3) In the first embodiment, the network structure calculation unit 105 obtains the related group based on the input design result. At this stage, the learning model is not needed yet.
Therefore, even in the initial stage of system operation in which the learning data cannot be sufficiently collected and the learning model cannot be sufficiently machine-learned, the network structure calculation unit 105 makes it possible to support the design using the related groups.
（４）実施例１では、ネットワーク構造計算部１０５が、所属グループに属する設計履歴の設計結果を取得して設計候補の設計結果との乖離を判定する。ネットワーク構造計算部１０５は、乖離すると判定された設計結果と近い（つまり乖離しない）設計結果を示すグループを関連グループとして求める。
　そのため、所属グループに対して乖離する設計結果がある場合に限って関連グループが探索されることになり、関連グループの探索が不必要に拡大しない。したがって、より必要な設計支援情報に絞って操作者に提供することが可能になる。 (4) In the first embodiment, the network structure calculation unit 105 acquires the design result of the design history belonging to the belonging group and determines the deviation from the design result of the design candidate. The network structure calculation unit 105 obtains a group showing a design result that is close to (that is, does not deviate) from the design result determined to be dissociated as a related group.
Therefore, the related group is searched only when there is a design result that deviates from the belonging group, and the search of the related group does not expand unnecessarily. Therefore, it is possible to narrow down the necessary design support information and provide it to the operator.
（５）実施例１では、データ分析部１１３は、過去の要求仕様および設計結果を含むデータセットの集合を、要求仕様に基づいてクラスタリングして複数のグループに分類する。学習モデル生成部１１４は、このグループ分類を行うための学習モデルを生成する。機械学習部１１５は、データ分析部１１３の分類結果を教師値として、過去の要求仕様を入力値とする学習データを作成して、学習モデルの機械学習を行う。
　したがって、分類部１０２は、要求仕様入力部１０１が取得した要求仕様を学習モデルの入力に与えることにより、要求仕様が分類されるべき所属グループを学習モデルの処理結果に基づいて推定することが可能になる。 (5) In the first embodiment, the data analysis unit 113 clusters a set of data sets including past requirement specifications and design results based on the requirement specifications and classifies them into a plurality of groups. The learning model generation unit 114 generates a learning model for performing this group classification. The machine learning unit 115 creates learning data using the classification result of the data analysis unit 113 as a teacher value and the past required specifications as an input value, and performs machine learning of the learning model.
Therefore, the classification unit 102 can estimate the group to which the requirement specifications should be classified based on the processing result of the learning model by giving the requirement specifications acquired by the requirement specification input unit 101 to the input of the learning model. become.
（６）実施例１では、学習モデル生成部１１４は、要求仕様を入力とし、所属グループおよび関連グループのネットワーク構造を出力とする学習モデルを生成する。機械学習部１１５は、学習データに対して、ネットワーク構造計算部による関連グループのネットワーク構造を教師値として更に付与して、学習モデルの機械学習を行う。
　したがって、分類部１０２は、要求仕様入力部１０１が取得した要求仕様を学習モデルの入力に与えることにより、要求仕様が分類されるべき所属グールプと、関連グループのネットワーク構造を推定することが可能になる。
　特に、実施例１では、設計結果が未知の状態でも関連グループのネットワーク構造を推定することが可能になる。その結果、設計結果を得るための操作者の手間を省くことが可能になる。 (6) In the first embodiment, the learning model generation unit 114 generates a learning model in which the required specifications are input and the network structure of the belonging group and the related group is output. The machine learning unit 115 further adds the network structure of the related group by the network structure calculation unit as a teacher value to the learning data, and performs machine learning of the learning model.
Therefore, by giving the requirement specifications acquired by the requirement specification input unit 101 to the input of the learning model, the classification unit 102 can estimate the affiliation group to which the requirement specifications should be classified and the network structure of the related group. Become.
In particular, in the first embodiment, it is possible to estimate the network structure of the related group even when the design result is unknown. As a result, it is possible to save the operator time and effort for obtaining the design result.
［実施形態の補足事項］
　なお、実施形態では、グループ分けに自己組織化マップを用いたが、本発明はこれに限定されない。例えば、Ｋ平均法などの別のクラスタリング技法を用いてもよい。 [Supplementary matters of the embodiment]
In the embodiment, a self-organizing map is used for grouping, but the present invention is not limited to this. For example, another clustering technique such as K-means may be used.
　さらに、実施形態では、グループ分けや機械学習などを、一台のコンピュータで実施するように記載しているが、本発明はこれに限定されない。ネットワーク環境を利用することによって、複数のハードウェアや複数のソフトウェアにより、設計支援作業を分担して実施してもよい。 Further, in the embodiment, it is described that grouping, machine learning, etc. are performed by one computer, but the present invention is not limited to this. By using the network environment, the design support work may be shared and carried out by a plurality of hardware and a plurality of software.
　また、学習モデルの中間層に、ステップＳ４０２で入力された設計結果を推定する人口ニューロンを一つ以上配置してもよい。このような人口ニューロンは、要求仕様を入力として、ステップＳ４０２で入力される設計結果を教師値として機械学習を行えばよい。このように設計結果を推定する人口ニューロンを中間層に配置することにより、関連グループのネットワーク構造の推定に、設計結果の推定結果を加味することが可能になり、関連グループのネットワーク構造の推定精度を高めることが可能になる。 Further, one or more artificial neurons for estimating the design result input in step S402 may be arranged in the middle layer of the learning model. Such an artificial neuron may perform machine learning using the required specifications as inputs and the design result input in step S402 as a teacher value. By arranging the artificial neurons that estimate the design result in the middle layer in this way, it becomes possible to add the estimation result of the design result to the estimation of the network structure of the related group, and the estimation accuracy of the network structure of the related group. It becomes possible to increase.
　さらに、実施形態では、関連グループのネットワーク構造として、説明を簡単にするため、１つの関連グループ「Ｉ」を求めるケースについて説明した。しかしながら、本発明はこれに限定されない。所属グループに対して、横に広がる複数の関連グループをネットワーク構造として求めてもよい。また、関連グループに対して、さらに先の関連グループを求めてもよい。さらに、所属グループに対してツリー構造の関連グループの群を求めてもよい。これらは、設計対象の事情などに応じて適宜に選択することが好ましい。 Further, in the embodiment, a case where one related group "I" is required is described as a network structure of related groups in order to simplify the explanation. However, the present invention is not limited to this. A plurality of related groups spreading horizontally may be obtained as a network structure for the belonging group. In addition, the related group may be further requested. Further, a group of related groups having a tree structure may be obtained for the belonging group. It is preferable to appropriately select these according to the circumstances of the design target and the like.
　また、実施形態では、学習モデルとして、ニューラルネットワークを使用した。しかしながら、本発明はこれに限定されない。例えば、このような学習モデルとしては、決定木学習、相関ルール学習、遺伝的プログラミング、帰納論理プログラミング、サポートベクターマシン、クラスタリング、ベイジアンネットワーク、強化学習、表現学習、主成分分析、エクストリーム・ラーニング・マシン、およびその他の機械学習技法の少なくとも一つの技法に基づく学習モデルを採用してもよい。 Also, in the embodiment, a neural network was used as a learning model. However, the present invention is not limited to this. For example, such learning models include decision tree learning, correlation rule learning, genetic programming, inductive logic programming, support vector machines, clustering, Basian networks, reinforcement learning, expression learning, principal component analysis, and extreme learning machines. , And a learning model based on at least one of the other machine learning techniques may be adopted.
１０１…要求仕様入力部、１０２…分類部、１０３…計算結果表示部、１０４…設計結果入力部、１０５…ネットワーク構造計算部、１１１…計算条件入力部、１１２…データベース、１１３…データ分析部、１１４…学習モデル生成部、１１５…機械学習部 101 ... Requirement specification input unit, 102 ... Classification unit, 103 ... Calculation result display unit, 104 ... Design result input unit, 105 ... Network structure calculation unit, 111 ... Calculation condition input unit, 112 ... Database, 113 ... Data analysis unit, 114 ... Learning model generator, 115 ... Machine learning unit

Claims (12)

  1. 　設計の要求仕様を入力するための要求仕様入力部と、
    　過去の設計履歴を分類したグループの群の中から、前記要求仕様が分類される前記グループ（以下「所属グループ」という）を求める分類部と、
    　前記所属グループについて情報を取得して、設計支援情報として提示可能にする計算結果表示部とを備え、
    　前記計算結果表示部は、
    　　前記要求仕様に基づいて設計される設計候補について、前記要求仕様と別の設計結果が分類される前記グループ（以下「関連グループ」という）について情報を取得して、設計支援情報として提示可能にする
    　ことを特徴とする設計支援システム。     Requirement specification input section for inputting design requirement specifications,
    From the group of groups that classify the past design history, the classification unit that obtains the group (hereinafter referred to as "affiliation group") in which the required specifications are classified, and the classification unit.
    It is equipped with a calculation result display unit that can acquire information about the group to which it belongs and present it as design support information.
    The calculation result display unit is
    Regarding design candidates designed based on the required specifications, information about the group (hereinafter referred to as "related group") in which design results different from the required specifications are classified can be acquired and presented as design support information. A design support system characterized by this.
  2. 　請求項１に記載の設計支援システムにおいて、
    　前記計算結果表示部は、
    　前記所属グループと、前記関連グループとを組み合わせた合成グループ（以下「ネットワーク化グループ」という）を作成し、前記ネットワーク化グループに基づく情報を前記設計支援情報として提示可能にする
    　ことを特徴とする設計支援システム。     In the design support system according to claim 1,
    The calculation result display unit is
    A design characterized in that a synthetic group (hereinafter referred to as "networking group") that combines the belonging group and the related group is created, and information based on the networking group can be presented as the design support information. Support system.
  3. 　請求項１～２のいずれか一項に記載の設計支援システムにおいて、
    　前記要求仕様に基づいて設計された設計候補について、前記設計結果を入力するための設計結果入力部と、
    　入力された前記設計結果が分類される前記関連グループを求めるネットワーク構造計算部と
    　を備えることを特徴とする設計支援システム。     In the design support system according to any one of claims 1 and 2.
    For the design candidates designed based on the required specifications, a design result input unit for inputting the design results, and
    A design support system including a network structure calculation unit for obtaining the related group in which the input design result is classified.
  4. 　請求項３に記載の設計支援システムにおいて、
    　前記ネットワーク構造計算部は、
    　前記所属グループに属する前記設計履歴の前記設計結果を取得して前記設計候補の前記設計結果との乖離を判定し、乖離する前記設計結果が分類される前記グループを前記関連グループとして求める
    　ことを特徴とする設計支援システム。     In the design support system according to claim 3,
    The network structure calculation unit
    It is characterized in that the design result of the design history belonging to the belonging group is acquired, the deviation of the design candidate from the design result is determined, and the group in which the dissociated design result is classified is obtained as the related group. Design support system.
  5. 　請求項１～４のいずれか一項に記載の設計支援システムにおいて、
    　前記分類部は、
    　前記設計履歴として、過去の前記要求仕様および前記設計結果を含むデータセットの集合を保持するデータベースと、
    　前記データセットの集合を、前記要求仕様に基づいて、自己組織化マップなどのクラスタリング技法により複数の前記所属グループに分類するデータ分析部と、
    　前記要求仕様を入力とし、前記所属グループの分類を出力とする学習モデルを生成する学習モデル生成部と、
    　前記データセットの前記要求仕様に対して、前記データ分析部による前記所属グループの分類を教師値として付与した学習データを作成して、前記学習モデルの機械学習を行う機械学習部とを備え、
    　前記要求仕様入力部が取得した前記要求仕様を前記学習モデルの入力に与え、前記学習モデルの処理結果に基づいて、前記要求仕様が分類される前記所属グループを推定する
    　ことを特徴とする設計支援システム。     In the design support system according to any one of claims 1 to 4,
    The classification unit
    As the design history, a database that holds a set of data sets including the past required specifications and the design results, and
    A data analysis unit that classifies the set of data sets into a plurality of the belonging groups by a clustering technique such as a self-organizing map based on the required specifications.
    A learning model generation unit that generates a learning model that inputs the required specifications and outputs the classification of the group to which it belongs.
    It is provided with a machine learning unit that creates learning data in which the classification of the belonging group by the data analysis unit is added as a teacher value to the required specifications of the data set and performs machine learning of the learning model.
    Design support characterized in that the requirement specifications acquired by the requirement specification input unit are given to the input of the learning model, and the belonging group to which the requirement specifications are classified is estimated based on the processing result of the learning model. system.
  6. 　請求項５に記載の設計支援システムにおいて、
    　前記学習モデル生成部は、前記要求仕様を入力とし、前記所属グループおよび前記関連グループのネットワーク構造を出力とする前記学習モデルを生成し、
    　前記機械学習部は、前記学習データに対して、前記関連グループのネットワーク構造を教師値として更に付与して、前記学習モデルの機械学習を行い、
    　前記分類部は、前記要求仕様入力部が取得した前記要求仕様を前記学習モデルの入力に与え、前記学習モデルの処理結果に基づいて、前記要求仕様が分類される前記所属グループおよび前記関連グループのネットワーク構造を推定する
    　ことを特徴とする設計支援システム。     In the design support system according to claim 5,
    The learning model generation unit generates the learning model that inputs the required specifications and outputs the network structure of the belonging group and the related group.
    The machine learning unit further adds the network structure of the related group as a teacher value to the learning data, and performs machine learning of the learning model.
    The classification unit gives the requirement specifications acquired by the requirement specification input unit to the input of the learning model, and based on the processing result of the learning model, the affiliation group and the related group to which the requirement specifications are classified are classified. A design support system characterized by estimating the network structure.
  7. 　設計の要求仕様を入力するための要求仕様入力ステップと、
    　過去の設計履歴を分類したグループの群の中から、前記要求仕様が分類される前記グループ（以下「所属グループ」という）を求める分類ステップと、
    　前記所属グループについて情報を取得して、設計支援情報として提示可能にする計算結果表示ステップとを備え、
    　前記計算結果表示ステップは、
    　　前記要求仕様に基づいて設計される設計候補について、前記要求仕様と別の設計結果が分類される前記グループ（以下「関連グループ」という）について情報を取得して、設計支援情報として提示可能にする
    　ことを特徴とする設計支援方法。     Requirement input steps for inputting design requirements and
    A classification step for obtaining the group (hereinafter referred to as "affiliation group") in which the required specifications are classified from the group of groups in which the past design history is classified, and
    It is provided with a calculation result display step that acquires information about the group to which it belongs and makes it possible to present it as design support information.
    The calculation result display step is
    Regarding design candidates designed based on the required specifications, information about the group (hereinafter referred to as "related group") in which design results different from the required specifications are classified can be acquired and presented as design support information. A design support method characterized by this.
  8. 　請求項７に記載の設計支援方法において、
    　前記計算結果表示ステップは、
    　前記所属グループと、前記関連グループとを組み合わせた合成グループ（以下「ネットワーク化グループ」という）を作成し、前記ネットワーク化グループに基づく情報を前記設計支援情報として提示可能にする
    　ことを特徴とする設計支援方法。     In the design support method according to claim 7,
    The calculation result display step is
    A design characterized in that a synthetic group (hereinafter referred to as "networking group") that combines the belonging group and the related group is created, and information based on the networking group can be presented as the design support information. Support method.
  9. 　請求項７～８のいずれか一項に記載の設計支援方法において、
    　前記要求仕様に基づいて設計された設計候補について、前記設計結果を入力するための設計結果入力ステップと、
    　入力された前記設計結果が分類される前記関連グループを求めるネットワーク構造計算ステップと
    　を備えることを特徴とする設計支援方法。     In the design support method according to any one of claims 7 to 8,
    For a design candidate designed based on the required specifications, a design result input step for inputting the design result and a design result input step.
    A design support method comprising: a network structure calculation step for obtaining the related group in which the input design result is classified.
  10. 　請求項７～９のいずれか一項に記載の設計支援方法において、
    　前記分類ステップは、
    　前記設計履歴として、過去の前記要求仕様および前記設計結果を含むデータセットの集合を保持するデータ保持ステップと、
    　前記データセットの集合を、前記要求仕様に基づいて、自己組織化マップなどのクラスタリング技法により複数の前記所属グループに分類するデータ分析ステップと、
    　前記要求仕様を入力とし、前記所属グループの分類を出力とする学習モデルを生成する学習モデル生成ステップと、
    　前記データセットの前記要求仕様に対して、前記データ分析ステップによる前記所属グループの分類を教師値として付与した学習データを作成して、前記学習モデルの機械学習を行う機械学習ステップとを備え、
    　前記要求仕様入力ステップが取得した前記要求仕様を前記学習モデルの入力に与え、前記学習モデルの処理結果に基づいて、前記要求仕様が分類される前記所属グループを推定する
    　ことを特徴とする設計支援方法。     In the design support method according to any one of claims 7 to 9,
    The classification step
    As the design history, a data holding step for holding a set of data sets including the past required specifications and the design result, and
    A data analysis step of classifying the set of data sets into a plurality of the belonging groups by a clustering technique such as a self-organizing map based on the required specifications.
    A learning model generation step that generates a learning model that inputs the required specifications and outputs the classification of the group to which the member belongs.
    It is provided with a machine learning step of creating learning data in which the classification of the belonging group by the data analysis step is added as a teacher value to the required specifications of the data set and performing machine learning of the learning model.
    Design support characterized in that the requirement specifications acquired by the requirement specification input step are given to the input of the learning model, and the belonging group to which the requirement specifications are classified is estimated based on the processing result of the learning model. Method.
  11. 　請求項１０に記載の設計支援方法において、
    　前記学習モデル生成ステップは、前記要求仕様を入力とし、前記所属グループおよび前記関連グループのネットワーク構造を出力とする前記学習モデルを生成し、
    　前記機械学習ステップは、前記学習データに対して、前記関連グループのネットワーク構造を教師値として更に付与して、前記学習モデルの機械学習を行い、
    　前記分類ステップは、前記要求仕様入力ステップが取得した前記要求仕様を前記学習モデルの入力に与え、前記学習モデルの処理結果に基づいて、前記要求仕様が分類される前記所属グループおよび前記関連グループのネットワーク構造を推定する
    　ことを特徴とする設計支援方法。     In the design support method according to claim 10,
    In the learning model generation step, the learning model is generated by inputting the required specifications and outputting the network structure of the belonging group and the related group.
    In the machine learning step, the network structure of the related group is further added as a teacher value to the learning data, and machine learning of the learning model is performed.
    In the classification step, the requirement specifications acquired by the requirement specification input step are given to the input of the learning model, and the requirement specifications are classified based on the processing result of the learning model of the affiliation group and the related group. A design support method characterized by estimating the network structure.
  12. 　コンピュータシステムを、請求項１～６のいずれか一項に記載の前記設計支援システムとして機能させる
    　ことを特徴とする設計支援プログラム。     A design support program characterized in that the computer system functions as the design support system according to any one of claims 1 to 6.
PCT/JP2020/006914 2019-05-07 2020-02-20 Design assistance system, design assistance method, and design assistance program WO2020225959A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2019087697A JP7270454B2 (en) 2019-05-07 2019-05-07 Design support system, design support method and design support program
JP2019-087697 2019-05-07

Publications (1)

Publication Number Publication Date
WO2020225959A1 true WO2020225959A1 (en) 2020-11-12

Family

ID=73045206

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2020/006914 WO2020225959A1 (en) 2019-05-07 2020-02-20 Design assistance system, design assistance method, and design assistance program

Country Status (2)

Country Link
JP (1) JP7270454B2 (en)
WO (1) WO2020225959A1 (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023176823A1 (en) * 2022-03-16 2023-09-21 三菱電機株式会社 Printed circuit board design assistance device and design assistance method, and program for causing computer to execute design assistance method

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2007128442A (en) * 2005-11-07 2007-05-24 Mazda Motor Corp Internal combustion engine design support system
JP2012003524A (en) * 2010-06-17 2012-01-05 Kobe Steel Ltd Steel product design support system, steel product design support method, and computer program

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP6611401B2 (en) * 2017-02-07 2019-11-27 株式会社日立製作所 Design support device
JP7140567B2 (en) * 2018-06-28 2022-09-21 株式会社日立製作所 Design proposal generator

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2007128442A (en) * 2005-11-07 2007-05-24 Mazda Motor Corp Internal combustion engine design support system
JP2012003524A (en) * 2010-06-17 2012-01-05 Kobe Steel Ltd Steel product design support system, steel product design support method, and computer program

Also Published As

Publication number Publication date
JP2020184159A (en) 2020-11-12
JP7270454B2 (en) 2023-05-10

Similar Documents

Publication Publication Date Title
Alinezhad et al. Sensitivity analysis of TOPSIS technique: the results of change in the weight of one attribute on the final ranking of alternatives
Ramirez et al. Measuring knowledge work: the knowledge work quantification framework
Aggarwal Compensative weighted averaging aggregation operators
CN108780313A (en) Increment related data digging technology for assembly line
CN110084627A (en) The method and apparatus for predicting target variable
Mousavi et al. Solving group decision-making problems in manufacturing systems by an uncertain compromise ranking method
Wang et al. Hybrid customer requirements rating method for customer-oriented product design using QFD
Vodopija et al. Characterization of constrained continuous multiobjective optimization problems: A feature space perspective
Habibi et al. Determination of hospital rank by using technique for order preference by similiarity to ideal solution (TOPSIS) and multi objective optimization on the basis of ratio analysis (MOORA)
WO2020225959A1 (en) Design assistance system, design assistance method, and design assistance program
Bobek et al. Enhancing cluster analysis with explainable AI and multidimensional cluster prototypes
Cui et al. A hybrid MCDM model with Monte Carlo simulation to improve decision-making stability and reliability
Meritxell et al. On the evaluation, management and improvement of data quality in streaming time series
Shi et al. A dynamic novel approach for bid/no-bid decision-making
Elahi et al. Evaluating software architectural styles based on quality features through hierarchical analysis and fuzzy integral (FAHP)
Nursal et al. The application of Fuzzy TOPSIS to the selection of building information modeling software
Szachniuk et al. MLP accompanied beam search for the resonance assignment problem
Zaabar et al. A two-phase part family formation model to optimize resource planning: a case study in the electronics industry
Gutiérrez-Salcedo et al. Identification and Visualization of the Conceptual Structure and Main Research Themes of Studies in Informatics and Control Journal from 2008 to 2019
Park et al. Development of automatic assembly sequence generating system based on the new type of parts liaison graph
Saini et al. Comparative analysis of classification algorithms using Weka
Basri Novelty ranking approach with z-score and fuzzy multi-attribute decision making combination
WO2019103773A1 (en) Automatically identifying alternative functional capabilities of designed artifacts
Helms et al. Classification of Methods for the Indication of Change Propagation-a Literature Review
Gonçalves et al. Decision Methodology for Maintenance KPI Selection: Based on ELECTRE I

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20802638

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20802638

Country of ref document: EP

Kind code of ref document: A1