JP2019185591A - Experiment assisting device and experiment assisting method - Google Patents

Abstract

To efficiently determine an appropriate experiment plan.SOLUTION: An experiment assisting device of the present invention is characterized by comprising: an input processing unit for accepting track record data including a product design value and a performance value, an optimization goal, and constraints; a feature quantity calculation unit for calculating feature quantity of the accepted track record data; an experiment plan/optimization technique selection unit for selecting an experiment plan or an optimization technique in accordance with the calculated feature quantity; an experiment plan/optimization execution unit for executing the selected experiment plan or the optimization technique; an experiment plan/optimization technique effect calculation unit for calculating an effect of the experiment plan or the optimization technique on the basis of an execution result of the experiment plan or the optimization technique; and an output processing unit for outputting the effect calculated in association with the experiment plan or the optimization technique.SELECTED DRAWING: Figure 10

Description

  The present invention relates to an experiment support apparatus and an experiment support method.

  The design conditions of the product are determined according to the design conditions so that the product satisfies desired performance and cost. For this purpose, many experiments with prototypes corresponding to various design conditions or simulations under analysis conditions corresponding to the design conditions are repeated. Some experiments and simulations are costly and time consuming, and it is important to minimize wasteful trial and error. Therefore, it is necessary to plan the experiment / simulation conditions so that the relationship between the design conditions and the performance / cost and the design conditions satisfying the desired performance / cost are clarified with the minimum number of experiments or simulations. This is called an experimental design. In the initial stage of design, an experiment plan is first designed to grasp the relationship between design conditions, performance, and cost. If a sufficient relationship can be grasped, an experiment plan for approaching the performance goal / cost goal can be performed, or a design plan closest to the goal can be obtained based on the grasped tendency. Obtaining the design conditions closest to the performance goal and cost goal, or performing an experimental design to obtain the design conditions closest to the goal is called optimization. These experimental design and optimization are collectively referred to as experimental design / optimization. Hereinafter, experiments and simulations are collectively referred to as experiments.

  In recent years, many experimental design and optimization techniques have been used. A heuristic method for global optimization represented by genetic algorithms, a statistical method represented by Bayesian optimization for determining optimal design conditions on the response surface indicating the relationship between design conditions and performance, orthogonal table An example is a classical method using the above. Each of these methods differs in the characteristics and number of experimental designs (optimal solutions) obtained thereby. In general, product design involves the steps of grasping the relationship between design conditions and performance, then trial and error through experiments based on the obtained relationships, and further optimizing the design conditions to output the desired performance. Trace in order. Each method described above cannot be preferably applied at any stage. At each stage, it is necessary to select a method.

  The experiment planning support method of Patent Document 1 accepts that the user inputs the number of experiments, the number of levels of design items (factors), and the like, and outputs an experiment plan suitable for it. The allocation method of Patent Document 2 accepts the user's input of the number of design item levels and the interaction between design items, and outputs an orthogonal table type suitable for the input. The test plan creation method of Patent Document 3 predicts the man-hours of a test case after setting priorities to the test cases, and allocates human resources necessary for the man-hours according to the priorities. Furthermore, the test plan creation method reviews the allocation of human resources when the test case does not end by a predetermined time limit.

US Pat. No. 6,448,983 JP 2001-265830 A JP 2011-257947 A

Product designers must select an appropriate experimental design / optimization method for the problem they are addressing from among many candidates, depending on the stage of the problem. Such a choice is difficult, especially for designers with little experience. In addition, even a skilled designer with abundant know-how may have a problem that the know-how is damaged and the objective experimental design / optimization method cannot be reached. The experimental design output (suggested) by the experimental design support method of Patent Literature 1 and the allocation method of Patent Literature 2 is premised on using an orthogonal table or a multi-way allocation table. These are effective when knowing the trend between design conditions and performance at the beginning of the experiment. However, if there is already some experience in experiments, another method (for example, a statistical method) is often required. The test plan creation method of Patent Document 3 focuses on a test case for software development, and a separate device is required to determine an experiment plan / optimization method in designing a general machine or the like.
Therefore, an object of the present invention is to efficiently determine an appropriate experiment plan.

The experiment support apparatus of the present invention includes an input processing unit that receives actual data including design values and performance values of products, an optimization target, and a constraint, and a characteristic amount calculation unit that calculates a characteristic amount of the received actual data, An experiment plan / optimization method selection unit that selects an experiment plan or optimization method according to the calculated feature value, an experiment plan / optimization execution unit that executes the selected experiment plan or optimization method, and an experiment plan or optimization An experiment plan / optimization effect calculation unit that calculates the effect of the experiment plan or the optimization method based on the execution result of the optimization method, and an output processing unit that outputs the effect calculated in association with the experiment plan or the optimization method It is characterized by that.
Other means will be described in the embodiment for carrying out the invention.

  According to the present invention, an appropriate experimental design can be determined efficiently.

It is a figure explaining the structure of an experiment assistance apparatus. It is a figure explaining a product. It is a figure explaining design value management information. It is a figure explaining feature-value information. It is a flowchart of the whole processing procedure. It is a flowchart of an experiment plan and optimization determination processing procedure. It is a flowchart of a deletion trial processing procedure. It is a flowchart of an additional trial process procedure. It is a figure which shows an example of an input screen. It is a figure which shows an example of an output screen.

  Hereinafter, a mode for carrying out the present invention (referred to as “the present embodiment”) will be described in detail with reference to the drawings. This embodiment is an example in which a designer (experiment support apparatus user of the present embodiment) having experiment result data accumulated to some extent creates an experiment plan for the next stage.

(Configuration of experiment support device)
The configuration of the experiment support apparatus 1 will be described with reference to FIG. The experiment support apparatus 1 is a general computer, and includes a central control device 11, an input device 12 such as a keyboard and a touch panel, an output device 13 such as a display, a main storage device 14, and an auxiliary storage device 15. These are connected to each other by a bus. The auxiliary storage device 15 stores design value management information 31, feature amount information 32, and an experiment plan / optimization method group 33 (all of which will be described later in detail).

  In the main storage device 14, an input processing unit 21, a feature amount calculation unit 22, an experiment plan / optimization method selection unit 23, an experiment plan / optimization execution unit 24, an experiment plan / optimization effect calculation unit 25, and an output processing unit 26. Is a program. In the following, when the operation subject is described as “XX section”, it means that the central control section 11 reads “XX section” from the auxiliary storage device 15 and loads it into the main storage device 14, and then the processing described later. Means to execute.

(Product)
The product will be described with reference to FIG. The product in the present embodiment is, for example, a rotary machine 41. The rotating machine 41 includes a rotating shaft 42, a bearing 43, and a bearing 44. The electric motor (not shown) gives a rotational force to the rotary shaft 42. Now, the user is trying to determine design conditions that minimize the vibration of the rotating shaft 42, the power consumption of the motor, the bearing 43, the temperature of the rotating shaft 42 in the bearing 43, and the like.

(Experiment, simulation, etc.)
An experiment is to place a prototype of a product in an operating environment, operate the prototype with design values (immediately described later), and obtain performance values (immediately later) of the prototype with various types of sensors. Simulation refers to inputting design values for a model on a computer and outputting performance values to the model. A model is a set of mathematical expressions including a plurality of variables. The experimental design / optimization method refers to an optimization method for searching for a design condition that outputs desired performance on a model or a response surface (detailed later), and a design condition for outputting desired performance. It is an experimental design method for planning an experiment. A plurality of different software applications for realizing the experiment design / optimization technique is the experiment design / optimization technique group 33 in FIG. The feature of this embodiment is that an appropriate experiment design / optimization method is selected.

(Design value)
Various design values can be assumed depending on the type and application of the rotating shaft 42. Here, it is assumed that the following items are design values.
・ Axial length “a” of the rotating shaft 42
The position of the bearing 43, that is, the distance “b” from the left end of the rotating shaft 42
The position of the bearing 44, that is, the distance “c” from the left end of the rotating shaft 42
The rotational speed of the rotating shaft 42 The force applied to the center of the rotating shaft 42 at an arbitrary position. This force may be an assumed load value (such as the gravity of a turbine blade) or an experimental excitation force.

(Performance value)
Various performance values can be assumed depending on the type and application of the rotating shaft 42. Here, it is assumed that the following items are performance values.
The primary mode vibration frequency of the rotating shaft 42, the secondary mode vibration frequency of the rotating shaft 42, the electric power consumption of the motor that applies a rotational force to the rotating shaft 42, and the temperature of the bearing 43 (equal to the temperature of the rotating shaft 42 in the bearing 43).
The temperature of the bearing 44 (equal to the temperature of the rotating shaft 42 in the bearing 44)

  As can be seen from the above, the design value of this embodiment is a five-dimensional vector, and the performance value is also a five-dimensional vector. It is a coincidence that the orders of both coincide. It is also possible to interpret that the design value is the “cause” of the causal relationship and the performance value is the “result” of the causal relationship. That is, the design value X can be interpreted as an input (explanatory variable) of a certain function F, and the performance value Y can be interpreted as an output (object variable) of the function F. The relationship among F, X, and Y is expressed by mathematical formulas below.

Y = F (X)
X = (x ₁ , x ₂ , x ₃ , x ₄ , x ₅ )
Y = (y ₁ , y ₂ , y ₃ , y ₄ , y ₅ )

Here, x ₁ , x ₂ ,..., X ₅ are the above-described five types of design values. y ₁ , y ₂ ,..., y ₅ are the five types of performance values described above. Even if the function F is unknown, if there are a sufficient number of combinations of the design value X and the performance value Y (experimental result data), the function F can be estimated using the actual data as learning data. .

(Response surface)
When discrete data of the design value X and the performance value Y obtained by experiments and simulations are approximated by a continuous curved surface, this curved surface is called a response curved surface. Data X and Y used to create this response surface are called learning data. As an example of the response surface, an example in which the function F is approximated by a linear function G as follows is given.
Y = F (X) = G (X) + error = a ₀ + a ₁ x ₁ + a ₂ x ₂ +... + A ₅ x ₅ + error where a ₁ , a ₂ , a ₃ ,. It is a parameter estimated by learning, and is determined so that the square error with the following function F is minimized.
(F (X) -G (X)) ²
This determination method of a ₁ , a ₂ , a ₃ ,... Is called a least square method.

(Response surface parameters)
In addition to the learning data, a value for determining the position and shape of the required response surface is called a “hyper parameter”. Hyper parameters are parameters that are not determined by learning, and are distinguished from parameters a ₁ , a ₂ , a ₃ ,... Estimated by learning.
As an example of a hyperparameter, Ridge regression that is an extension of the least square method is given. In Ridge regression, a ₁ , a ₂ , a ₃ ,... Are determined so that the following values are minimized so that a ₁ , a ₂ , a ₃ ,.
(F (X) −G (X)) ² + λ (a ₀ ² + a ₁ ² +... + A ₅ ² )
Here, newly introduced λ is a hyper parameter, and by changing the size of λ, the sizes of a ₁ , a ₂ , a ₃ ... Can be controlled, and overlearning can be prevented. . Since λ cannot be determined by learning, it is a hyperparameter. The learned parameters and hyperparameters are called response surface parameters.

  The experiment design / optimization method of the present embodiment has the following two types.

<Type 1: Experiment design / optimization method for finding optimal solution of design value>
The type 1 experiment design / optimization method optimizes the design value X so that one or a plurality of performance values y _i selected as the optimization target take a desired value. Therefore, the user performs an experiment with a prototype using the optimized design value X, and confirms that the prototype actually outputs a desired performance value. Specific examples of Type 1 are as follows.

・ "Gradient method" using derivatives
・ "Linear programming" such as simplex method
・ "Genetic algorithm" that multiplies two experimental points with good results

<Type 2: Experiment design / optimization method for finding multiple candidates for the optimal solution of the design value>
The type 2 experiment design / optimization method determines a plurality of design value X candidates that are likely to be optimal solutions, even if the optimal solution is not determined as one. Therefore, the user uses a plurality of design value X candidates in a trial and error manner, performs an experiment with a prototype, and determines the design value X such that the prototype outputs a desired performance value. Type 2 is slightly more detoured than Type 1 in that it imposes trial and error on the user. However, the existence of Type 2 is that the number of experiments can be reduced. A specific example of type 2 is as follows.

・ "Orthogonal table" that creates an orthogonal table based on the given number of levels, number of factors (number of types of design values), interaction between factors, etc.
-“Multiple layout table” that creates brute force combinations of design values based on the number of levels, factors, etc.
・ "Optimum planning" using existing experimental points such as D-optimal programming and I-optimal programming
・ "Design by density" that adds experimental points as a new sample to the part where the density of existing experimental points is low
“Bayesian optimization” that uses kriging, bootstrap sampling and machine learning to calculate the predicted value and variance of the response surface and uses an acquisition function such as EI (Expected Improvement)

For example, if there are five types of design values as described above and there are three levels for each of them, theoretically, 3 ⁵ = 243 experiments are required. However, the experimental design / optimization method for creating the orthogonal table reduces this 243 times to 27 times, for example.

(Effects of experimental design and optimization methods)
The experiment support apparatus 1 calculates the effect on the result of executing the experiment design / optimization technique. Examples of specific indices used as effects include the performance improvement degree A (formula (1)), the prediction accuracy improvement degree I (formula (2)), and the screening achievement degree D (formula (3)) defined below. ), Reliability C (formula (4)) and required number of experiments.

The performance improvement degree A is defined by equation (1). Here, “x _c ” is a design value proposed by the experiment design / optimization method. E in open bold is the expected value of the random variable. “˜” indicates that the random variable follows a probability distribution. The performance improvement degree A is an expected improvement amount (EI value) of y at the experimental point to be added.

The prediction accuracy improvement degree I is defined by equation (2). Here, “L1” is the L1 norm. “Μ” and “σ” are the mean and variance of the expected values of the performance value y obtained by linear regression, kriging or the like, respectively. “Μ ′” and “σ ′” are the mean and variance of the predicted value of y when (x _c , μ (x _c )) is added to the learning data when learning the response surface. The decrease in the prediction accuracy improvement degree I indicates that the prediction accuracy of the performance value y is improved in the entire search region (x definition region).

  Screening achievement degree D is defined by equation (3). Here, “L∞” is an L∞ norm. The decrease in the screening achievement degree D indicates that the parameters of the response surface can be well predicted when linear regression or the like is performed.

  The reliability C is defined by equation (4). Here, “p (x)” is the experimental point density. “Α” is a constant.

  The required number of experiments is, for example, the number of rows in the orthogonal table when an experiment is performed using the orthogonal table.

(Design value management information)
The design value management information 31 will be described with reference to FIG. In the design value management information 31, in association with the experiment ID stored in the experiment ID column 101, the performance value column 102 has a performance value, the design value column 103 has a design value, and the acquisition time column 104 has an acquisition time point. However, the actual / plan column 105 stores the actual / plan flag, the experiment plan / optimization method column 106 stores the experiment plan / optimization method, the effect column 107 stores the effect, and the proposal column 108 stores the proposal flag. Has been.

The experiment ID in the experiment ID column 101 is an identifier that uniquely identifies the result data or the experiment plan or optimization plan based on the experiment plan / optimization method.
The performance value in the performance value column 102 is the above-described performance value. Here, the five types of performance values described above are stored in the small columns 102a to 102e.
The design value in the design value column 103 is the above-described design value. Here, the above-described five types of design values are stored in the small columns 103a to 103e.
The acquisition time point in the acquisition time column 104 is the date when the performance value is actually acquired through an experiment.

The actual / plan flag in the actual / plan column 105 is either “actual” or “plan”. “Actual” indicates that an experiment using the design value is performed and a performance value as an experimental result has already been acquired. “Plan” indicates that an experiment using the design value has not yet been performed.
The experiment plan / optimization technique in the experiment plan / optimization technique column 106 is the name of the experiment plan / optimization technique.
The effect in the effect column 107 is a result of applying a predetermined threshold to an index value (A, I, D, C, or necessary number of experiments) indicating the effect of the above-described experiment design / optimization method. Then, it is either “large” or “small”. Among the above-mentioned five types of effects, there are those in which the effect here becomes “large” as the index value increases, and the effect here becomes “increase” as the index value decreases. There is also.

The proposal flag in the proposal column 108 is one of “adopted”, “rejected”, and “added”. “Rejected” indicates that the experimental design / optimization method was once determined to be output (proposed) to the user, but was not finally output as a result of review. “Adopted” indicates that the experimental design / optimization method is once determined to be output to the user and finally output as scheduled. “Addition” indicates that the experimental design / optimization method was once determined as not to be output to the user, but was finally output as a result of the review.
In the design value management information 31 of FIG. 3, “#” represents different design values in an abbreviated manner. “♭” represents different performance values in an abbreviated manner.

  In FIG. 3, the record above the double line 109 is a record related to the performance data. Accordingly, data is stored in the experiment ID column 101 to the actual / plan column 105, while the experiment plan / optimization method column 106 to the proposal column 108 are blank.

  A record below the double line 109 is a record related to the content proposed by the experiment support apparatus 1 to the user. Therefore, the acquisition time column 104 is blank. The performance value column 102 may store data or may be blank. When the experiment design / optimization method belongs to type 1 (for example, the gradient method), five “♭” are stored in the performance value columns 102a to 102e. One or more of them have a desired value as an optimization target. When the experiment design / optimization method belongs to type 2 (for example, orthogonal table L4), the performance value column 102 may be blank. This is because Type 2 only proposes candidate design values expected to output a desired performance value.

(Feature information)
The feature amount information 32 will be described with reference to FIG. In the feature amount information 32, an experiment plan / optimization method is stored in the experiment plan / optimization method column 112 in association with the actual data feature amount stored in the result data feature amount column 111. The experiment support apparatus 1 may use the feature amount information 32 shown in FIG. 4 when selecting an experiment plan / optimization method according to the feature amount of the actual data, or the experiment plan / optimization described later in FIG. A technique determination procedure may be used. According to the feature amount information 32, the experiment support apparatus 1 can determine one or a plurality of experiment plans / optimization methods suitable for the feature amount when the feature amount of the performance data is given.

  The actual data feature amount column 111 includes small columns 111a to 111g. In these small columns, a plurality of feature amounts of the performance data are stored. The feature amount is a value that can be calculated by the experiment support apparatus 1 by an arbitrary method based on the design value and the performance value that the user has already acquired (for the record above the double line 109 in FIG. 3). “...” Represents different feature amounts in an abbreviated manner.

  The experimental point distribution (column 111e) is information on the distribution of variables or the like that have been dimensionally compressed by a self-encoder or the like, for example. The information on the response surface (column 111f) is information on the response surface obtained by machine learning such as linear regression, kernel regression, grigging, ensemble learning, or deep learning. The information (column 111g) obtained from the response surface is information obtained from the response surface, such as a predicted value that can be estimated using grigging or bootstrap sampling. The information regarding the response surface and the information obtained from the response surface include the parameters estimated when learning the response surface.

(Overall procedure)
The overall processing procedure will be described with reference to FIG. In the description, FIGS. 9 and 10 are referred to as appropriate.
In step S201, the input processing unit 21 of the experiment support apparatus 1 displays the input screen 51 (FIG. 9). Specifically, the input processing unit 21 displays the input screen 51 on the output device 13.

  In step S202, the input processing unit 21 receives performance data and the like. Specifically, first, the input processing unit 21 operates the file selection buttons 52a and 52b in the result data column 52 of the input screen 51, and the design value result data and the performance value result data, respectively. Accept to select. Here, the design value and performance value of the record above the double line 109 of the design value management information 31 in FIG. 3 may be stored in advance in an arbitrary external server (not shown in FIG. 1) or the like. It is a premise. Thereafter, the input processing unit 21 acquires design values and performance values from an external server or the like, creates design value management information 31 using the acquired design values and performance values, and stores them in the auxiliary storage device 15.

  Secondly, the input processing unit 21 accepts that the user operates the drop-down menu 53a in the optimization target column 53 to select a performance value as an optimization target. The candidate performance values to be selected are the primary mode vibration frequency, the secondary mode vibration frequency,... Furthermore, the input processing unit 21 accepts that the user operates the drop-down menu 53b to select “maximize”, “minimize”, or “specific value”. When the user selects “specific value”, the user inputs the value (or the range that the value takes). Furthermore, the input processing unit 21 accepts that the user presses a button 53d for adding an optimization target or a button 53c for deleting an optimization target as necessary.

  Third, the input processing unit 21 accepts that the user inputs a design value constraint in the constraint column 54. For each design value, the user inputs the lower limit of the design value in the lower limit text box 54a and the upper limit of the design value in the upper limit text box 54b. The user presses the button 54d for adding or deleting the design value for inputting the lower limit and the upper limit as necessary. The user inputs a mathematical expression that defines a constraint between a plurality of design values in the inequality text box 54e. The user presses the button 54g for adding or deleting the restriction condition as necessary.

Fourth, the input processing unit 21 accepts that the user inputs the number of possible experiments in the number of possible experiments text box 55. The number of possible experiments here is, for example, the maximum number of rows of the orthogonal table 63b displayed on the output screen 61 (FIG. 10).
Fifth, the input processing unit 21 accepts that the user presses the execution button 56.

  In step S203, the feature amount calculation unit 22 of the experiment support apparatus 1 calculates a feature amount. Specifically, the feature amount calculation unit 22 calculates an arbitrary feature amount by an arbitrary method based on the data received in step S202. As described above with reference to FIG. 4, the feature amount example includes information on the response surface and information obtained from the response surface. Therefore, the feature amount calculation unit 22 learns the function F using the record data as learning data. For example, the feature amount calculation unit 22 performs multiple regression analysis in a multivariable space and learns a curved surface (response curved surface) that approximates the function F.

  In step S204, the experiment plan / optimization method selection unit 23 of the experiment support apparatus 1 selects an experiment plan / optimization method. Specifically, first, the experiment plan / optimization method selection unit 23 uses the feature amount information 32 (FIG. 4) or the experiment plan / optimization method decision processing procedure (details described later) in FIG. ) To select one or more experimental design / optimization methods. Note that “selecting an experiment design or optimization method” in the claims corresponds to selecting an experiment design / optimization method. Among these, “selecting an experimental design” corresponds to selecting an experimental design / optimization method that shows a plurality of combinations of design values, such as an orthogonal table. “Selecting an optimization method” corresponds to selecting an experimental design / optimization method that shows a combination of design values that are theoretically guaranteed to be optimal solutions, such as the gradient method. .

  Second, the experiment design / optimization technique selection unit 23 creates a new record of the design value management information 31 (FIG. 3). Then, the experiment plan / optimization method selection unit 23 puts the new record in the experiment / plan column 105, the experiment plan / optimization method column 106, and the proposal column 108 in the “plan” and “first” in step S204, respectively. The selected experiment design / optimization method and “adopted” are stored.

  In step S205, the experiment plan / optimization execution unit 24 of the experiment support apparatus 1 executes the experiment plan / optimization technique. Specifically, the experiment plan / optimization execution unit 24 executes the experiment plan / optimization method selected in “first” in step S <b> 204 with a program stored in the experiment plan / optimization method group 33. To do.

  The experiment plan / optimization execution unit 24 stores the experiment plan / design value plan that is the execution result of the experiment plan / optimization at step S205 in the design value column 103 of the corresponding record in the design value management information 31.

  The experiment plan / optimization execution unit 24 repeats the process of step S205 for all the experiment plans / optimization methods selected in “first” of step S204.

In step S206, the experiment plan / optimization effect calculation unit 25 of the experiment support apparatus 1 calculates an effect. Specifically, first, the experiment plan / optimization effect calculation unit 25 determines the performance improvement degree A, the prediction accuracy improvement degree I, the screening based on the results output by the respective experiment plans / optimization methods. At least one of achievement level D, reliability level C, and required number of experiments is calculated.
Second, the experiment design / optimization effect calculation unit 25 applies a predetermined threshold to the value acquired in “first” in step S206, and acquires “large” or “small” as the effect.
Third, the experiment design / optimization effect calculation unit 25 stores “large” or “small” in the effect column 107 of the record created in “second” in step S204.

  In step S207, the experiment plan / optimization effect calculation unit 25 attempts to delete the experiment plan / optimization technique. Details of step S207 will be described later as a deletion trial processing procedure in FIG. As a result, the experiment plan / optimization effect calculation unit 25 determines that the proposal flag “of” is selected for the experiment plan / optimization method selected in “first” in step S204 that has the effect “small”. Make a “rejection”. For example, the proposal flag “rejected” is set for an object having a short distance from the existing experimental point. Of course, there may be no experimental design / optimization method for setting the proposal flag “reject”.

  In step S208, the experiment design / optimization effect calculation unit 25 attempts to add an experiment design / optimization technique. Details of step S208 will be described later as an additional trial processing procedure in FIG. As a result, the experiment plan / optimization effect calculation unit 25 prevents the proposal flag “rejection” from being set for all the experiment plans / optimization methods selected in “first” in step S204.

  In step S209, the experiment plan / optimization effect calculation unit 25 determines whether there is an added experiment plan / optimization technique. Specifically, the experiment plan / optimization effect calculation unit 25 returns to step S205 when there is an experiment plan / optimization method added in step S208 (step S209 “Yes”). In other cases (step S209 “No”), the experiment plan / optimization effect calculation unit 25 proceeds to step S210.

In step S210, the output processing unit 26 of the experiment support apparatus 1 displays the output screen 61 (FIG. 10). Specifically, first, the output processing unit 26 displays the output screen 61 on the output device 13.
Secondly, the output processing unit 26 displays the experiment plan / optimization technique proposed to the user in the experiment plan / optimization technique column 62 b of the output screen 61. Of the records of the design value management information 31 (FIG. 3) (below the double line 109), the experiment plan / optimization method of the record whose proposal flag is “adopted” or “added” is displayed in the experiment plan method column 62b. Will be displayed.

Third, the output processing unit 26 includes four values (performance improvement degree A, prediction accuracy improvement degree I, screening achievement degree D, reliability degree) calculated in “first” in step S206 in the effect column 62c. C) is displayed. These four values are preferably normalized to facilitate comparison with each other.
Note that the output processing unit 26 may display a graph (for example, a scatter diagram of design values) that is dimensionally compressed by principal component analysis, t-SNE, or the like, in addition to those illustrated in FIG. 10.

  Fourth, the output processing unit 26 displays the design value in the experiment condition column 62a. In many cases, the displayed design value is a design value that is theoretically guaranteed to output a desired performance value that is an optimization target (reference numeral 63a). However, when the experimental design / optimization method is an orthogonal table or the like belonging to type 2, the design values displayed here are a plurality of candidates expected to output desired performance values (reference numeral 63b). . Looking at the code 63b, two levels (“10” and “30”) as the design value 1 and two levels (“0” and “0.1”) as the design value 2 are displayed twice in each column. Has been (orthogonal). Incidentally, the number of rows “4” of the code 63b is used for determining the effect of the “orthogonal table (L4)” as the required number of experiments. Thereafter, the entire processing procedure is terminated.

(Experimental design / optimization method decision process)
The experimental design / optimization method decision processing procedure will be described with reference to FIG. The experimental design / optimization technique determination processing procedure is the details of step S204 (FIG. 5).
In step S301, the experiment plan / optimization method selection unit 23 of the experiment support apparatus 1 acquires a feature amount. Specifically, the experiment design / optimization method selection unit 23 acquires the feature amount calculated in step S203.

  In step S302, the experiment plan / optimization technique selection unit 23 determines whether or not the number of experiments is equal to or greater than a threshold value. The number of experimental points is the number of actual data, that is, the number of records above the double line 109 in the design value management information 31 of FIG. The experiment plan / optimization method selection unit 23 proceeds to step S304 if the number of experiment points is equal to or greater than a predetermined threshold (step S302 “Yes”), and proceeds to step S303 otherwise (step S302 “No”). .

  In step S303, the experiment design / optimization technique selection unit 23 adds the orthogonal table, the optimum design method, and the multi-way arrangement table to the candidates. Specifically, the experiment design / optimization method selection unit 23 gives an appropriate design value (argument) and then adds the orthogonal table, the optimum design method, and the multi-way layout table to the experiment design / optimization method candidates. To do. Thereafter, the experimental design / optimization method decision processing procedure is terminated.

  In step S304, the experiment design / optimization technique selection unit 23 determines whether or not there is an area with a low experiment point density that is equal to or greater than a threshold. The experimental point density is the number of design values that are actual data in a predetermined region of design values. The experiment plan / optimization method selection unit 23 proceeds to step S305 if the region having a low experiment point density is equal to or higher than the predetermined threshold (step S304 “Yes”), otherwise (step S304 “No”). The process proceeds to S306.

  In step S305, the experiment design / optimization technique selection unit 23 adds the orthogonal table, the optimum design method, and the multi-way arrangement table to the candidates. Specifically, the experiment design / optimization technique selection unit 23 adds an orthogonal table, an optimum design method, and a multi-way arrangement table to the experiment design / optimization technique candidates for the region where the experiment point density is low. Thereafter, the experiment design / optimization technique selection unit 23 proceeds to step S306.

  In step S306, the experiment design / optimization method selection unit 23 determines whether or not the response surface parameter is normal. Specifically, when the parameter of the response surface calculated as the feature amount is within a predetermined normal range (step S306 “Yes”), the experiment design / optimization method selection unit 23 proceeds to step S308, and otherwise If so (step S306 “No”), the process proceeds to step S307.

In step S307, the experiment design / optimization technique selection unit 23 adds the optimum design method and the density design method to the experiment design / optimization technique candidates. Thereafter, the experimental design / optimization method decision processing procedure is terminated.
The “planning method by density” is a method of adding a sample of design values to the point where the logarithmic experimental point density (logp) calculated as the feature quantity is the smallest. The log experimental point density is defined by equation (5). “2” on the right side of Expression (5) represents the Euclidean norm, and “t” is a positive constant that is arbitrarily set by the user.

  In step S308, the experiment design / optimization technique selection unit 23 adds the density design method and Bayesian optimization to the experiment design / optimization technique candidates.

  In step S309, the experiment design / optimization method selection unit 23 determines whether or not the variance of the response surface is equal to or less than a threshold value. Specifically, the experiment design / optimization method selection unit 23 proceeds to step S310 if the variance of the performance value y of the response surface is equal to or smaller than a predetermined threshold (step S309 “Yes”), and otherwise ( In step S309 “No”), the experimental design / optimization method decision processing procedure ends.

  In step S310, the experiment design / optimization technique selection unit 23 determines whether there is one optimization target. Specifically, the experiment plan / optimization method selection unit 23 proceeds to step S311 when the design value of the optimization target is one type (step S310 “Yes”), and otherwise (step S310 “No”). "), The process proceeds to step S312.

  In step S311, the experiment design / optimization technique selection unit 23 adds the gradient method to the experiment design / optimization technique candidates. Thereafter, the experimental design / optimization method decision processing procedure is terminated.

  In step S312, the experiment design / optimization technique selection unit 23 adds the linear programming method and the genetic algorithm to the experiment design / optimization technique candidates. As an example of linear programming, there is a simplex method. Thereafter, the experimental design / optimization method decision processing procedure is terminated.

  In the experiment planning / optimization technique determination processing procedure, the criterion for selecting the experiment planning / optimization technique is the feature value of the actual data. However, in order to select a more appropriate experiment design / optimization technique, it is necessary to verify the effect of the selected experiment design / optimization technique. The experiment support apparatus 1 deletes an experiment plan / optimization method whose effect does not reach a predetermined standard in a deletion trial process procedure to be described later, and an experiment plan in which the effect may increase in an additional trial process procedure to be described later・ Add optimization methods.

(Deletion trial processing procedure)
The deletion trial processing procedure will be described with reference to FIG. The deletion trial processing procedure is the details of step S207 (FIG. 5).
In step S401, the experiment plan / optimization effect calculation unit 25 of the experiment support apparatus 1 acquires an experiment plan / optimization technique to be processed. Specifically, the experiment plan / optimization effect calculation unit 25 acquires an unprocessed one of the experiment plans / optimization methods selected in “first” in step S204.

In step S402, the experiment plan / optimization effect calculation unit 25 determines whether the added experiment plan / optimization method is used. Specifically, first, the experiment plan / optimization effect calculation unit 25 determines whether the experiment plan / optimization technique to be processed has been added in step S502 (FIG. 8) to be described later. Check for no.
Second, the experiment plan / optimization effect calculation unit 25 proceeds to step S403 if the experiment plan / optimization method is not added (step S402 “No”), and otherwise (step S402 “Yes”). ”), The deletion trial processing procedure is terminated.

  In step S403, the experiment plan / optimization effect calculation unit 25 determines whether the distance from the existing experiment point is equal to or less than a threshold value. Specifically, the experiment plan / optimization effect calculation unit 25 calculates a representative value (for example, an average value) of design values set in the experiment plan / optimization method to be processed and a design value of actual data. Calculate the distance. If the distance is equal to or smaller than the predetermined threshold (step S403 “Yes”), the experiment plan / optimization effect calculation unit 25 proceeds to step S405, and otherwise (step S403 “No”) proceeds to step S404. move on.

  In step S404, the experiment plan / optimization effect calculation unit 25 determines whether the effect is equal to or less than a threshold value. Specifically, the experiment plan / optimization effect calculation unit 25 proceeds to step S405 when the effect of the experiment plan / optimization method to be processed is equal to or less than a predetermined threshold (step S404 “Yes”). In other cases (step S404 “No”), the deletion trial processing procedure is terminated.

  In step S405, the experiment plan / optimization effect calculation unit 25 deletes the experiment plan / optimization technique. Specifically, the experiment plan / optimization effect calculation unit 25 sets the proposal flag “rejected” to the experiment plan / optimization method to be processed. Thereafter, the deletion trial processing procedure is terminated. The experiment plan / optimization effect calculation unit 25 repeats the deletion trial processing procedure for each unprocessed experiment plan / optimization technique.

(Additional trial processing procedure)
The additional trial processing procedure will be described with reference to FIG. The additional trial processing procedure is the details of step S208 (FIG. 5).
In step S501, the experiment plan / optimization effect calculation unit 25 of the experiment support apparatus 1 determines whether the candidate experiment plan / optimization method is empty. “Empty” means that no experimental design / optimization method has been selected in “first” of step S204, or the proposal flag “reject” is set for all selected experimental designs / optimization methods. Means that The experiment plan / optimization effect calculation unit 25 proceeds to step S502 if it is empty (step S501 “Yes”), and proceeds to step S503 otherwise (step S501 “No”).

  In step S502, the experiment plan / optimization effect calculation unit 25 adds the density design method and Bayesian optimization to the experiment plan / optimization method candidates. Here, the Bayesian optimization is added to the experimental design / optimization method candidates when the parameters of the response surface are within the normal range. Here, the experiment plan / optimization effect calculation unit 25 stores a record in which the proposal flag is “added” at the end of the design value management information 31 (FIG. 3).

  In step S503, the experiment plan / optimization effect calculation unit 25 acquires the experiment plan / optimization technique to be processed. Specifically, the experiment plan / optimization effect calculation unit 25 selects one of the unprocessed experiment plans / optimization methods selected in “first” in step S204 or added in step S502. get.

In step S504, the experiment design / optimization effect calculation unit 25 determines whether it is necessary to adjust the design value or the like. Specifically, first, the experiment design / optimization effect calculation unit 25 compares the relationship between the experiment design / optimization technique used in step S205 and the effect calculated in step S206 by comparing magnitude or linearity. Analyze by regression including regression and kriging.
Second, the experiment design / optimization effect calculation unit 25 determines whether there is a possibility that the effect is improved by changing the design value and / or the parameter based on the result of the analysis.
Third, the experiment plan / optimization effect calculation unit 25 proceeds to step S505 if there is a possibility (step S504 “Yes”), and otherwise performs the additional trial process procedure (step S504 “No”). finish.

In step S505, the experiment plan / optimization effect calculation unit 25 adjusts the design value and the like. Specifically, the experiment design / optimization effect calculation unit 25 sets the design value column 103 of the record of the design value management information 31 stored in step S502 to the design value after the change in “second” in step S504. Update. Furthermore, the experiment design / optimization effect calculation unit 25 updates the parameter of the function F (response surface) with the parameter after the change in “second” in step S504.
The experiment plan / optimization effect calculation unit 25 repeats the processes of steps S503 to S505 for each unprocessed experiment plan / optimization technique. Thereafter, the additional trial processing procedure is terminated.

(Effect of this embodiment)
The effects of the experiment support apparatus of this embodiment are as follows.
(1) The experiment support apparatus can efficiently output an appropriate experiment plan / optimization method based on the experiment result data.
(2) The experiment support apparatus can delete or add the experiment plan / optimization technique to be output according to the effect.
(3) The experiment support apparatus can determine in advance the relationship between the actual data and the experiment plan / optimization technique.
(4) The experiment support device can output an optimized design value if the conditions are met, or otherwise output an experiment plan that reduces the burden on the designer like an orthogonal table. Can do.

  In addition, this invention is not limited to an above-described Example, Various modifications are included. For example, the above-described embodiments have been described in detail for easy understanding of the present invention, and are not necessarily limited to those having all the configurations described. Further, a part of the configuration of one embodiment can be replaced with the configuration of another embodiment, and the configuration of another embodiment can be added to the configuration of one embodiment. Further, it is possible to add, delete, and replace other configurations for a part of the configuration of each embodiment.

Each of the above-described configurations, functions, processing units, processing means, and the like may be realized by hardware by designing a part or all of them with, for example, an integrated circuit. Each of the above-described configurations, functions, and the like may be realized by software by interpreting and executing a program that realizes each function by the processor. Information such as programs, tables, and files that realize each function can be stored in a recording device such as a memory, a hard disk, or an SSD (Solid State Drive), or a recording medium such as an IC card, an SD card, or a DVD.
In addition, the control lines and information lines are those that are considered necessary for the explanation, and not all the control lines and information lines on the product are necessarily shown. In practice, it may be considered that almost all the components are connected to each other.

DESCRIPTION OF SYMBOLS 1 Experiment support apparatus 11 Central control apparatus 12 Input apparatus 13 Output apparatus 14 Main memory 15 Auxiliary memory 21 Input processing part 22 Feature-value calculation part 23 Experiment plan and optimization method selection part 24 Experiment plan and optimization execution part 25 Experiment Planning / optimization effect calculation unit 26 Output processing unit 31 Design value management information 32 Feature quantity information 33 Experiment plan / optimization method group

Claims (5)

An input processing unit that accepts performance data including product design values and performance values, optimization targets, and constraints;
A feature amount calculation unit for calculating a feature amount of the received record data;
An experiment plan / optimization method selection unit for selecting an experiment plan or an optimization method according to the calculated feature amount;
An experiment plan / optimization execution unit for executing the selected experiment plan or the optimization method;
An experiment plan / optimization effect calculation unit for calculating an effect of the experiment plan or the optimization method based on an execution result of the experiment plan or the optimization method;
An output processing unit for outputting the calculated effect in association with the experimental design or the optimization method;
An experiment support apparatus comprising: The experiment design / optimization effect calculation unit
Delete experimental designs or optimization methods where the effect does not reach a predetermined standard,
Adding experimental designs or optimization techniques that may increase the effect,
The experiment support apparatus according to claim 1. The experiment design / optimization method selection unit
Selecting the experiment plan or the optimization method according to information that defines in advance the relationship between the actual data and the experiment plan or the optimization method;
The experiment support apparatus according to claim 2. The experiment design / optimization method selection unit
Selecting the experimental design or the optimization method from among optimal programming, density programming, Bayesian optimization, gradient method, linear programming, genetic algorithm, orthogonal table and multi-way table,
The experiment support apparatus according to claim 3. The input processor of the experiment support device
Accept actual data including product design and performance values, optimization targets, and constraints,
The feature amount calculation unit of the experiment support apparatus includes:
Calculating a feature amount of the received performance data;
The experiment planning / optimization method selection unit of the experiment support device is:
Select an experiment plan or an optimization method according to the calculated feature amount,
The experiment plan / optimization execution unit of the experiment support apparatus is:
Performing the selected experimental design or the optimization technique;
The experiment plan / optimization effect calculation unit of the experiment support device
Calculate the effect of the experiment plan or the optimization method based on the execution result of the experiment plan or the optimization method,
The output processing unit of the experiment support apparatus includes:
Outputting the calculated effect in association with the experimental design or the optimization method;
The experiment support method of the experiment support apparatus characterized by this.

Images