Method for designing predictive model, method for generating predictive model, program, predicting device, and predictive model design device

Abstract

In this method for designing a predictive model to be executed by a computer, the following steps are repeated until a predetermined end condition is satisfied, namely: generating a predictive model using teacher data in which a parameter set prepared on the basis of a plurality of types of parameters indicating characteristics of spring steel and test results relating to a specific characteristic of the spring steel are associated with each other; using a plurality of algorithms for outputting a ranking of a degree of contribution of each type of parameter to the test results, to rank the types of parameters of the teacher data used in the predictive model generated immediately before; and, on the basis of the ranking of the types of parameters, regenerating the predictive model using teacher data in which a parameter set having a reduced number of parameter types and the test results relating to the specific characteristic of the spring steel are associated with each other.

Description

Prediction model design method, prediction model generation method, program, prediction device, and prediction model design device

[0001] One embodiment of the present invention is a method for designing a prediction model for predicting the properties of steel material, a method for generating a prediction model, a program, a prediction device, and a prediction model design device.

[0002] In the production and development of steel materials, in order to simply and quickly evaluate the performance of materials, it is required to accurately predict the properties of the produced steel material from its metal structure, manufacturing conditions, etc. Patent Document 1 discloses a material quality prediction model for predicting the mechanical properties of cold-rolled steel sheets for processing.

[0003] Japanese Patent Application Laid-Open No. 2022-136745

[0004] The mechanical properties of steel are difficult to predict because factors such as the steel's composition and the macrostructure of the steel, such as the amount of carbides and grain size, are intricately intertwined. To accurately predict the mechanical properties of manufactured steel, it is desirable to generate a prediction model using a dataset that associates measured values ​​of many types of parameters with test results of the mechanical properties. However, because the measured values ​​of many types of parameters and test results of the mechanical properties are required, preparing the dataset takes time.

[0005] On the other hand, if the number of data sets is small, the accuracy of the prediction model will be poor even if there are many types of parameters.In addition, since the appropriate parameters differ for each type of mechanical property, it is difficult to select the appropriate parameters from a large number of multiple parameters.

[0006] In view of the above problems, one object of one embodiment of the present invention is to provide a prediction model that has high prediction accuracy even with a small data set.

[0007] One embodiment of the present invention is a computer-implemented method for designing a predictive model, which generates a predictive model using training data in which a parameter set prepared based on multiple types of parameters that indicate the characteristics of spring steel is associated with test results related to predetermined properties of the spring steel, ranks the parameter types of the training data used in the predictive model previously generated using multiple algorithms that output a ranking of the contribution of each parameter type to the test results, and regenerates a predictive model using training data in which a parameter set with a reduced number of parameter types is associated with test results related to the predetermined properties of the spring steel based on the ranking of the parameter types, until a predetermined termination condition is met, and provides information for selecting at least one predictive model from the multiple predictive models generated based on the difference between the predicted values ​​of the test results output for each of the multiple predictive models and the actual measured values ​​of the test results.

[0008] In the method for designing a prediction model, selecting at least one prediction model from the plurality of prediction models may include selecting a prediction model with a smallest difference value.

[0009] In the above-mentioned method for designing a prediction model, ranking multiple types of parameters by multiple algorithms may include averaging the rankings of the multiple algorithms for each of the multiple types and ranking the averaged parameter types.

[0010] In the method for designing a prediction model, the training data may be such that the lowest ranking parameter type among the averaged parameter types is excluded.

[0011] In the above-mentioned method for designing a prediction model, providing information for selecting at least one prediction model from the plurality of prediction models may include providing the prediction model two models prior to the most recently generated prediction model when, when a first difference between the actual measured values ​​and predicted values ​​of the test results of the prediction model generated immediately before is compared with a second difference between the actual measured values ​​and predicted values ​​of the test results of the prediction model two models prior to the most recently generated prediction model, the first difference is greater than the second difference.

[0012] In the above-mentioned method for designing a predictive model, the multiple algorithms may be two or more selected from AIC, BIC, Lasso regression, ANN, and random forest, XGBoost, LightGBM, a combination of ANN and SHAP, a combination of random forest and SHAP, a combination of XGBoost and SHAP, and a combination of LightGBM and SHAP.

[0013] In the above method for designing a prediction model, the multiple algorithms may be AIC, BIC, Lasso regression, a combination of ANN and SHAP, ANN, and random forest.

[0014] In the method for designing the prediction model, the multiple types of parameters may include at least one of a parameter representing a grain size, a parameter representing a carbide morphology, a parameter representing a dislocation density, a parameter representing a solute carbon amount, a parameter representing residual γ, a parameter representing hardness, and a parameter related to the manufacturing conditions of the spring steel.

[0015] In the above-mentioned method for designing a prediction model, when the multiple types of parameters are parameters representing the crystal grain size, they may include at least one of the circle equivalent diameter of the crystal grain size, the circularity of the crystal grain size, and the aspect ratio of the crystal grain size.

[0016] In the above-mentioned method for designing a prediction model, when multiple types of parameters are parameters that represent the morphology of carbides, the parameters that represent the morphology of carbides may include at least one of a circle-equivalent diameter of carbides, a number density of carbides, an aspect ratio of carbides, an area ratio of carbides, and a volume ratio of carbides.

[0017] In the method for designing the prediction model, the plurality of types of parameters may include a circle-equivalent diameter of the crystal grain size, an area ratio of carbides, an aspect ratio of carbides, a number density of carbides, a circle-equivalent diameter of carbides, residual γ, hardness, and elements constituting the spring steel.

[0018] In the above-described method for designing a prediction model, the predetermined property of the spring steel may be at least one of tensile strength, toughness, sag, and fatigue.

[0019] A program according to one embodiment of the present invention causes a computer to execute the above-described method for designing a prediction model.

[0020] A method for generating a prediction model according to one embodiment of the present invention determines a prediction model by the above-described method for designing a prediction model.

[0021] A prediction device according to one embodiment of the present invention predicts test results relating to predetermined properties of spring steel using a prediction model generated using the above-described method for generating a prediction model.

[0022] A prediction model design device according to one embodiment of the present invention includes: a prediction model generation unit that generates a prediction model using training data that associates a parameter set prepared based on multiple types of parameters that indicate the characteristics of spring steel with test results related to specified properties of the spring steel; a ranking determination unit that ranks the types of parameters in the training data used in the most recently generated prediction model using multiple algorithms that output a ranking of the contribution of each parameter type to the test results; a training data generation unit that generates training data that associates a parameter set with a reduced number of parameter types with test results related to the specified properties of the spring steel based on the ranking of the parameter types; and a prediction model providing unit that provides information for selecting at least one prediction model from the multiple prediction models based on the difference between the predicted values ​​of the test results output for each of the multiple generated prediction models and the actual measured values ​​of the test results.

[0023] According to one embodiment of the present invention, it is possible to provide a prediction model with high prediction accuracy even with a small data set.

[0024] FIG. 1 is a block diagram illustrating the configuration of a prediction model design device according to one embodiment of the present invention. FIG. 2 is a block diagram illustrating various types of information stored in a storage unit. FIG. 3 is a block diagram illustrating a prediction model design function according to one embodiment of the present invention. FIG. 4 is a block diagram illustrating a teacher data storage unit and a prediction model storage unit stored in a storage unit. FIG. 5 is a table in which multiple types of parameters are ranked according to their contribution using multiple types of algorithms. FIG. 6 is a flowchart illustrating a method for designing a prediction model according to one embodiment of the present invention. FIG. 7 is a table showing 16 parameters and their rankings. FIG. 8 is a table showing the ranking of parameters when the number of parameter types is reduced 11 times. FIG. 9 is a block diagram illustrating a teacher data storage unit and a prediction model storage unit stored in a storage unit. FIG. 10 is a block diagram illustrating a prediction device that realizes a prediction function according to one embodiment of the present invention.

[0025] Each embodiment will be described below with reference to the drawings. However, the present invention can be embodied in various forms without departing from the spirit of the invention, and should not be construed as being limited to the description of the embodiments exemplified below.

[0026] Furthermore, in this specification and each drawing, elements having the same functions as those explained in the previous drawings may be given the same reference numerals, and duplicate explanations may be omitted.

[0027] 1. Configuration of a Prediction Model Design Apparatus Fig. 1 is a block diagram showing the configuration of a prediction model design apparatus 10 according to this embodiment. The prediction model design apparatus 10 includes a control unit 11, a storage unit 12, a communication unit 13, an input / output I / F 14, and a display unit 15. The prediction model design apparatus 10 is not limited to one that includes all of these components.

[0028] The control unit 11 is an example of a computer including a processor (arithmetic processing circuit) such as a CPU and a storage device such as a RAM. The control unit 11 executes various programs stored in the storage unit 12 using the processor to realize various functions in the prediction model design device. For example, by executing a program for designing a prediction model, the control unit 11 can design a prediction model that allows optimal parameters to be selected from multiple types of parameters required to predict the mechanical properties of spring steel.

[0029] The storage unit 12 is a storage device such as a non-volatile memory or a hard disk drive. The storage unit 12 stores a spring steel database in which multiple types of parameters indicating the characteristics of spring steel are associated with test results regarding predetermined properties of the spring steel. The storage unit 12 stores a prediction model design program and a prediction program as programs. The various types of data stored in the storage unit 12 will be described in detail later. Note that although the storage unit 12 is illustrated in FIG. 1 as a single storage device, it may be composed of multiple storage devices.

[0030] Spring steel generally refers to steels such as carbon-based, silicon-manganese-based, manganese-chromium-based, and chromium-vanadium-based steels. Furthermore, spring steels are not limited to the above-mentioned steels and may include steels that are suitable for springs. Examples of elements constituting spring steel include Ni, Mn, C, Si, Cu, Cr, Ti, V, Mo, and B.

[0031] The communication unit 13, under the control of the control unit 11, connects to the network NW and transmits and receives information to and from other computers, servers, or other devices connected to the network NW.

[0032] The input / output I / F 14 is an interface with a user and includes at least one of a keyboard, a mouse, a speaker, a printer, a scanner, etc. The input / output I / F 14 may also be connected to various measuring devices 141 for performing various tests related to the characteristics and mechanical properties of the spring steel. Examples of the measuring device 141 include an optical microscope, a scanning electron microscope (SEM), an electron backscatter diffraction (EBSD), a transmission electron microscope (TEM), an X-ray diffraction (XRD), a surface roughness meter (contact or non-contact), an atomic force microscope (AFM), small angle X-ray scattering, and an atom probe tomography (APT). The prediction model design device 10 may be connected to the measurement device 141 via the communication unit 13. Alternatively, the prediction model design device 10 may be connected to the measurement device 141 or a server connected to the measurement device 141 via the communication unit 13.

[0033] The display unit 15 is, for example, a display device such as a liquid crystal display panel or an EL display panel. The display unit 15 may be a display device that is integrally connected within a housing that constitutes the prediction model design device 10. Alternatively, the display unit 15 may be an external display device that is connected to the input / output I / F 14. The display unit 15 displays a display screen related to an acquisition unit, which is a functional block of the design function described later.

[0034] 2. Configuration of the Storage Unit Fig. 2 is a block diagram showing various types of information stored in the storage unit 12. The storage unit 12 stores a spring steel database 121, a teacher data storage unit 122, a prediction model storage unit 124, a prediction model design program 126, a prediction program 127, and an algorithm storage unit 128. The storage unit 12 is not limited to storing all of these components, and may store only a part of these components. The spring steel database 121, the teacher data storage unit 122, the prediction model storage unit 124, the prediction model design program 126, the prediction program 127, and the algorithm storage unit 128 shown in Fig. 2 are not limited to being stored in a single storage device, but may be stored in a distributed manner across multiple storage devices.

[0035] 2-1. Spring Steel Database The spring steel database 121 associates multiple types of parameters that indicate the characteristics of spring steels that have been measured in the past with test results for specific properties of the spring steels. Each spring steel is assigned a unique identifier, and the identifier, multiple types of parameters, and test results are stored in association with each other. The various parameters that indicate the characteristics of spring steels will be explained below.

[0036] The various parameters that indicate the characteristics of the spring steel include at least one of a parameter that indicates the grain size, a parameter that indicates the morphology of carbides, a parameter that indicates the dislocation density, a parameter that indicates the amount of solute carbon, a parameter that indicates the residual γ, a parameter that indicates the hardness, and a parameter related to the manufacturing conditions of the spring steel.

[0037] The grain size can be obtained by analyzing image data obtained by observing the surface of the spring steel using an optical microscope or SEM / EBSD to obtain surface information. Parameters that represent the grain size include the equivalent circle diameter (μm), circularity, and aspect ratio. The equivalent circle diameter (μm) is expressed using the arithmetic mean, weighted mean, maximum value, minimum value, median, and standard deviation. The circularity is the value obtained by dividing the equivalent circle perimeter by the actual perimeter. The aspect ratio is the ratio of the grain size in the longitudinal direction (rolling direction) of the spring steel to the grain size in the width direction (direction perpendicular to the longitudinal direction) of the spring steel.

[0038] The morphology of carbides can be obtained by analyzing image data obtained by observing the surface of spring steel using an SEM as surface information. Furthermore, the morphology of carbides can be obtained by analyzing the profile of X-ray scattering angle and scattering intensity obtained using small-angle X-ray scattering as internal information. Examples of parameters that represent the morphology of carbides include the circle-equivalent diameter (μm), number density, aspect ratio, area ratio, and volume ratio. The circle-equivalent diameter (μm) is expressed using the arithmetic mean, weighted mean, maximum value, minimum value, median, and standard deviation. The number density (pieces / μm ２ ) refers to the number of carbide particles in a unit area. ３ ) refers to the number of carbide particles in a unit volume. Aspect ratio refers to the ratio of the major axis to the minor axis in the size of the carbide (ellipticity). Area ratio refers to the area of ​​the carbide that occupies the evaluation range. Volume ratio refers to the volume of the carbide that occupies the evaluation range. Here, the parameters of circle equivalent diameter, number density (number of carbide particles in a unit area), aspect ratio, and area ratio are used as surface information. The parameters of number density (number of carbide particles in a unit volume) and volume ratio are used as internal information.

[0039] Surface roughness can be obtained by analyzing surface information, such as image data obtained by measuring the surface of the spring steel using an AFM or shape data or image data obtained by measuring the surface of the spring steel using a surface roughness meter (contact or non-contact). Parameters that represent surface roughness include the arithmetic mean roughness Ra and the maximum height roughness Rz (axial, circumferential, and 45° directions). Here, with the longitudinal direction of the wire (or rolling direction) as the axis, the directions perpendicular to and at 45° to the axis are called the circumferential and 45° directions, respectively.

[0040] The dislocation density can be obtained as internal information by analyzing the XRD spectrum of the interior of the spring steel using XRD. －２ ) is used. Alternatively, image data obtained by observing a cross section of the spring steel using a TEM may be analyzed as internal information. The dislocation density can be quantified, for example, by measuring dislocation lines in an image observed using a TEM and dividing the measured value by the observation range (field of view size × sample thickness).

[0041] The amount of solute carbon can be obtained as internal information by analyzing the XRD spectrum of spring steel using XRD. The relationship between the lattice constant of Fe in steel and the amount of solute carbon is generally known. When using XRD, the amount of solute carbon can be obtained by determining the lattice constant using XRD and then converting it to the amount of solute carbon. When using electrical resistance measurement, the amount of solute carbon can be indirectly determined by preparing a known reference material, utilizing the correlation between electrical resistance value and the amount of solute carbon. Furthermore, the amount of solute carbon within the evaluation range can be obtained by analyzing 3D elemental mapping at the atomic level using APT. The amount of solute carbon (wt%) is used as a parameter representing the amount of solute carbon.

[0042] The residual γ is obtained as internal information by analyzing the XRD spectrum of the spring steel using XRD. Also, the surface information is obtained by analyzing image data obtained by observing the surface of the spring steel using SEM / EBSD. The residual γ content (%) is used as a parameter representing the residual γ.

[0043] The characteristics of dislocations are obtained as internal information by analyzing the XRD spectrum of the spring steel using XRD. The ratio (%) of edge dislocations to screw dislocations is used as a parameter representing the characteristics of dislocations.

[0044] Residual stress is obtained as internal information by analyzing the XRD spectrum of the spring steel using XRD. Surface residual stress (MPa), crossing point depth (mm), maximum residual stress (MPa), and maximum residual stress depth (mm) are used as parameters representing residual stress.

[0045] The inclusion distribution can be quantified by observing the fracture surface after a fatigue test or a hydrogen-charged tensile test. For example, image data of the inclusion distribution can be obtained as internal information by observing the cross section using a SEM. Depth (mm) and size (mm) are used as parameters representing the inclusion distribution. Depth (mm) and size (mm) can be obtained by statistically processing the average, variance, maximum, and minimum.

[0046] Representative elements constituting spring steel include Ni, Mn, C, Si, Cu, Cr, Ti, V, Mo, and B. The parameters of the elements constituting spring steel may be, for example, the content (wt%) or ratio of the elements contained in the spring steel. Spark optical emission spectroscopy (OES; Optical Emission Spectrometer) may be used as a method for analyzing the elements constituting spring steel. Alternatively, the parameters of the elements constituting spring steel may be values ​​listed on a mill sheet, which is a document certifying the quality of the steel (steel inspection certificate). Alternatively, information in a database such as Total Material (manufactured by KEY to METALS AG) may be referenced.

[0047] The hardness of spring steel is obtained, for example, by a Vickers hardness test (JIS Z2244). For example, Vickers hardness (HV) is used as a parameter representing hardness. Methods for evaluating hardness include a Rockwell hardness test (JIS Z2245), a Brinell hardness test (JIS 2243), and a Shore hardness test (JIS Z2246). Although the hardness of spring steel relates to mechanical properties, it is easier to test than tensile properties, toughness, settling, and fatigue, and therefore may be used as a parameter indicating the characteristics of spring steel.

[0048] Parameters related to the manufacturing conditions of spring steel include the heat treatment method of the spring steel. As the main heat treatment methods for spring steel, conditions for quenching and tempering can be used as input parameters. For example, in the case of quenching spring steel, temperature, time, and heating means (furnace heating, induction heating, electrical heating, etc.) can be used as parameters, and in the case of tempering spring steel, temperature, time, and heating means (furnace heating, induction heating, electrical heating, etc.) can be used as parameters. Furthermore, conditions for annealing or normalizing can also be used as input parameters for the heat treatment method. Even when conditions for annealing or normalizing are used as input parameters, the temperature, time, and heating means can be used as parameters. In other words, the condition information for manufacturing spring steel only needs to include at least one of quenching, tempering, annealing, and normalizing as the heat treatment method, and at least one of time, temperature, and heating means as the heat treatment method.

[0049] The various parameters stored in the spring steel database 121 do not need to include all of the above-mentioned parameters, but may store a plurality of parameters appropriately selected from the above-mentioned parameters.

[0050] 2-2. Property Test Results The property test results 123b include mechanical properties such as hardness, tensile properties, toughness, sag, and fatigue. Parameters obtained by measuring the hardness, tensile properties, toughness, sag, and fatigue of the spring steel may be stored as the property test results 123b, or parameters predicted by a prediction model may be stored. As described above, the hardness of the spring steel may be used as one of multiple types of parameters indicating the characteristics of the spring steel in the training data 123, or may be used as the property test results 123b. The parameters of the tensile properties, toughness, sag, and fatigue are obtained by the methods described below.

[0051] The tensile properties are obtained, for example, by a tensile test (JIS Z2241) as tensile properties (MPa), 0.2% proof stress (MPa), elongation (%), and reduction of area (%). The tensile properties and proof stress are calculated from the SS diagram obtained in the test, and the elongation and reduction of area are calculated from the dimensions of the test piece before and after the test. The toughness is obtained, for example, by a Charpy impact test (JIS Z2242) as the Charpy impact value (J / cm ２ ) Fatigue can be measured by measuring the fatigue limit (MPa) using, for example, the 14S-N test method (described in, for example, JSME S 002). Creep strain (%) can be obtained by conducting a creep test at a predetermined temperature, stress, and time. Note that the methods for measuring hardness, tensile properties, toughness, fatigue, and sag are not limited to those described above, and any known method may be used.

[0052] 2-3. Teacher Data Storage Unit The teacher data storage unit 122 stores teacher data 123. As will be described in detail later, the teacher data storage unit 122 stores a plurality of teacher data 123. The teacher data 123 includes a parameter set 123a prepared based on a plurality of types of parameters of spring steel and characteristic test results 123b. The parameter set 123a does not need to include all of the above-mentioned plurality of types of parameters; appropriately selected parameters may be used. Furthermore, the characteristic test results include at least one test result related to a predetermined characteristic of the spring steel. In this embodiment, the predetermined characteristic of the spring steel refers to a mechanical characteristic.

[0053] The prediction model storage unit 124 stores a prediction model 125. As will be described in detail later, the prediction model storage unit 124 stores a plurality of prediction models 125. The prediction model 125 is a model that has learned the relationship between a parameter set that is prepared in advance based on a plurality of types of parameters that indicate the characteristics of spring steel and test results regarding predetermined properties of spring steel.

[0054] The prediction model design program 126 is a program for designing a prediction model 125 to be used in the prediction program 127. The prediction program 127 obtains predicted values ​​of test results relating to predetermined properties of the spring steel by providing the prediction model with multiple types of parameters that indicate the acquired characteristics of the spring steel. The prediction model 125, the prediction model design program 126, and the prediction program 127 will be described in detail later.

[0055] 2-5. Algorithm Storage Unit The algorithm storage unit 128 stores an algorithm for selecting parameters (feature selection) to be used in the prediction model. As will be described in detail later, when there are a large number of parameters for predicting predetermined properties of spring steel, it is preferable to gradually remove parameters using an algorithm for performing feature selection. Feature selection involves evaluating the importance of each feature with respect to the output and retaining features that have a greater impact on the machine learning model. The features selected typically differ for each algorithm. The algorithm storage unit 128 stores multiple different types of algorithms for calculating the ranking of the contribution of parameters that indicate the characteristics of spring steel to mechanical properties.

[0056] When an information criterion is used as the algorithm stored in the algorithm storage unit 128, AIC (Akaike Information Criterion) or BIC (Bayes Information Criterion) may be used. Both AIC and BIC are indices for model selection. AIC selects a model by taking into account both the complexity of the model and its fit to the data. AIC generates a statistical model for each combination of all multiple parameters, compares the generated models, and determines the model with the smallest value as the relatively best model. BIC is based on the same concept as AIC, but penalizes models that include too many terms.

[0057] Lasso regression is a type of linear regression, a regression analysis method that performs variable selection through regularization to improve the predictive accuracy and interpretability of the generated statistical model. In Lasso regression, a regularization term is added to prevent the model from becoming too complex. As the regularization term parameter λ (λ > 0) increases, the weight of unnecessary features becomes 0. In other words, the idea is that features whose weight does not become 0 even when λ is increased are important for prediction.

[0058] A machine learning algorithm may be used to search for the optimal combination of features. For example, an ANN (artificial neural network) is composed of an input layer, a hidden layer, and an output layer, and is capable of learning and predicting complex relationships through multi-layered nonlinear transformations. The ANN calculates the importance of each feature by focusing on the weights of the model after learning. The larger the weight of a feature, the greater its contribution to prediction.

[0059] Furthermore, as machine learning algorithms, Random Forest, XGBoost, and LightGBM, which apply a decision tree model, may be used. Random Forest uses a technique called bagging, while XGBoost and LightGBM use a technique called boosting. Random Forest, XGBoost, and LightGBM can output feature values ​​that are important to the results by using a technique called Feature Importance (hereinafter referred to as FI) that calculates the importance of each feature value to the prediction.

[0060] Furthermore, the machine learning model generated by the above machine learning algorithm may be interpreted using SHAP (Shapley Additive Explanations). SHAP uses game theory concepts to quantify the contribution of each feature to the prediction result. This makes it possible to visualize the impact of an increase or decrease in the value of a feature variable.

[0061] As described above, the algorithms stored in the algorithm storage unit 128 may be any algorithm capable of calculating the importance and contribution of multiple feature quantities to the prediction model 125. Examples of algorithms stored in the algorithm storage unit 128 include AIC (Akaike Information Criterion), BIC (Bayes Information Criterion), Lasso regression, ANN (artificial neural network), Random Forest, XGBoost (eXtreme Gradient Boosting), and LightGBM (Light Gradient Boosting Machine). Furthermore, when using ANN, Random Forest, XGBoost, or LightGBM as an algorithm, SHAP may be combined to quantify the contribution of the parameters. While FIG. 2 illustrates an example in which two algorithms 129a and 129b are stored in the algorithm storage unit 128, the number of algorithms is not limited. The number of types of algorithms stored in the algorithm storage unit 128 may be two or more.

[0062] 3. Design Function of Prediction Model Next, the design function of the prediction model will be described. The control unit 11 executes a design program for the prediction model stored in the storage unit 12, and realizes the design function of the prediction model in accordance with the instructions written in the program.

[0063] 3 is a diagram showing the configuration of the prediction model design function. The prediction model design function is realized by a prediction model generation unit 261, a prediction unit 262, an error calculation unit 263, a determination unit 264, a ranking determination unit 265, a parameter selection unit 266, a training data generation unit 267, and a prediction model provision unit 268.

[0064] To accurately predict the mechanical properties of manufactured spring steel, it is desirable to generate a prediction model using many data sets in which measured values ​​of many types of parameters are associated with test results of the mechanical properties. However, because many types of measured values ​​of parameters and test results of the mechanical properties are required, preparing the data sets takes time. On the other hand, if the number of data sets is small, the accuracy of the prediction model may actually decrease even if there are many types of parameters. Furthermore, because appropriate parameters differ for each type of mechanical property, it is difficult to select appropriate parameters from a large number of multiple parameters. Therefore, the prediction model design function according to one embodiment of the present invention enables selection of appropriate parameters (also known as feature selection) according to the type of mechanical property, even with a small data set.

[0065] The training data storage unit 122 stores training data in which a parameter set prepared based on a plurality of types of parameters indicating the characteristics of spring steel obtained in the past and test results relating to predetermined properties of the spring steel are associated with each other. In this embodiment, a case will be described in which 16 parameters, namely, grain size, hardness, residual γ, carbide area ratio, solute carbon content, carbide aspect ratio, carbide number, carbide size, dislocation density, and elements constituting the spring steel (Ni, Mn, C, Si, Cu, Cr, and Ti), are used as the training data, along with toughness test results as test results relating to the predetermined properties of the spring steel.

[0066] The prediction model generation unit 261 acquires the training data 123 from the training data storage unit 122. The prediction model generation unit 261 uses multiple types of parameters indicating the characteristics of the spring steel as explanatory variables and test results related to the predetermined properties of the spring steel as objective variables to generate a prediction model by machine learning for predicting measured values ​​related to the predetermined properties of the spring steel from the multiple types of parameters. The machine learning may use at least one machine learning algorithm from among neural networks, decision tree learning, and support vector regression. In this embodiment, the prediction model generation unit 261 uses the training data to input 16 parameters indicating the characteristics of the spring steel and generate a neural network that outputs predicted values ​​of toughness test results for the spring steel. The prediction model generation unit 261 stores the generated prediction model in the prediction model storage unit 124.

[0067] 4 is a block diagram showing the teacher data storage unit 122 and the prediction model storage unit 124 stored in the memory unit 12. The teacher data storage unit 122 stores, as teacher data 123-1, a parameter set 123a-1 prepared based on 16 parameters indicating the characteristics of spring steel, and toughness test results as characteristic test results 123b. The prediction model storage unit 124 stores a prediction model 125-1 corresponding to the teacher data 123-1.

[0068] The prediction unit 262 inputs 16 parameters of the spring steel using the prediction model 125-1 stored in the prediction model storage unit 124, and outputs a predicted value of the test result of the toughness of the spring steel.

[0069] The error calculation unit 263 acquires the predicted value of the test result of the toughness of the spring steel output from the prediction model 125-1 and the actual measured value of the test of the toughness of the spring steel. The error calculation unit 263 calculates the difference (also called error) between the predicted value of the test result of the toughness of the spring steel output from the prediction model acquired by the acquisition unit and the actual measured value of the test of the toughness of the spring steel. This makes it possible to calculate the prediction accuracy of the generated prediction model 125-1.

[0070] The ranking unit 265 calculates and ranks the contributions of each of the 16 parameters in the prediction model 125 using multiple algorithms stored in the algorithm storage unit 128. As described above, the algorithms stored in the algorithm storage unit 128 may be at least two selected from AIC, BIC, Lasso regression, ANN weight, Random Forest FI, XGBoost FI, LightGBM FI, a combination of ANN and SHAP, a combination of Random Forest and SHAP, a combination of XGBoost and SHAP, and a combination of LightGBM and SHAP. Furthermore, the statistical models and machine learning models generated by the ranking unit 265 using each algorithm may be stored in the algorithm storage unit 128.

[0071] FIG. 5 is a table in which 16 types of parameters are ranked according to their contributions to five types of algorithms. As shown in FIG. 5, the 16 parameters are ranked according to their contributions for each of the algorithms: AIC, BIC, Lasso regression, a combination of ANN and SHAP, ANN (weight), and random forest (FI). As shown in FIG. 5, the rankings of the 16 parameters vary depending on the characteristics of the algorithm. Therefore, to reduce the influence of the characteristics of the algorithm, the rankings are averaged for each parameter. FIG. 5 shows the rankings of the parameters obtained by averaging the rankings for each parameter.

[0072] As explained in the algorithm storage unit 128, the approach and calculation method for each feature value differ depending on the algorithm. Therefore, by ranking the parameters using multiple methods and averaging the rankings for each parameter, highly reliable results can be obtained regarding the contribution of each parameter.

[0073] The parameter selection unit 266 excludes the parameter with the lowest ranking in the parameter ranking. In Figure 5, the lowest-ranked parameters are Cr and Ti, so these parameters are excluded. The parameter selection unit 266 reads the teacher data 123-1 in the teacher data storage unit 122, generates new teacher data 123-2 by excluding the lowest-ranked parameter, and stores it in the teacher data storage unit 122.

[0074] The prediction model design function generates new training data by using multiple algorithms to exclude the lowest-order parameters from multiple types of parameters, and then generates a prediction model using the training data, repeating this process until a predetermined termination condition is met. Furthermore, multiple prediction models are generated depending on the number of repetitions. The determination unit 264 determines whether to reduce the number of parameters based on the prediction accuracy of the prediction model 125-1. If the determination unit 264 determines to reduce the number of parameters, it generates new training data by using multiple algorithms to exclude the lowest-order parameters from multiple types of parameters, and then generates a prediction model using the training data. If the determination unit 264 determines not to reduce the number of parameters, it terminates the repetitive process, assuming that a predetermined termination condition has been satisfied. The predetermined termination condition may, for example, be a comparison of the error of the most recently generated prediction model 125 with the error of the prediction model 125 generated two models prior to the current prediction model 125, and terminate the repetitive process when the error of the most recently generated prediction model 125 becomes larger than the error of the prediction model 125 generated two models prior to the current prediction model 125.

[0075] The prediction model providing unit 268 provides information for selecting at least one prediction model 125 based on a predetermined provision condition from among the multiple generated prediction models 125. For example, the predetermined provision condition may be to compare the error of the most recently generated prediction model 125 with the error of the prediction model 125 generated two models ago, and select the prediction model 125 generated two models ago if the error of the most recently generated prediction model 125 is larger than that of the prediction model 125 generated two models ago. For example, the prediction model providing unit 268 may provide (output) information for selecting at least one prediction model 125 to a display device.

[0076] 4. Prediction Model Design Method Next, the prediction model design method executed in the prediction model design function will be described with reference to Figures 6 to 10. Figure 6 explains how the prediction model design method searches for the optimal type and number of parameters for obtaining a predicted toughness value from the 16 parameters described above and toughness test results.

[0077] FIG. 6 is a flowchart illustrating a method for designing a prediction model. This flowchart begins when the prediction model generation unit 261 acquires the training data 123. The prediction model generation unit 261 uses multiple types of parameters that indicate the characteristics of spring steel as explanatory variables and test results related to the specified properties of spring steel as objective variables to generate a prediction model 125-1 for predicting measured values ​​related to the specified properties of spring steel from the multiple types of parameters (step S101). Next, the prediction unit 262 inputs actual measured values ​​of the multiple types of parameters into the generated prediction model 125-1, thereby outputting predicted values ​​of the test results output from the prediction model 125-1. The error calculation unit 263 acquires actual measured values ​​of the test results corresponding to the actual measured values ​​of the multiple types of parameters and predicted values ​​of the test results output from the prediction model 125-1, and calculates the difference (also referred to as error) between the actual measured values ​​and the predicted values. The smaller the difference between the actual measured values ​​and the predicted values, the higher the prediction accuracy of the prediction model. This allows the prediction accuracy of the prediction model 125-1 to be output (step S102). The difference between the actual measurement value and the predicted value calculated by the error calculation unit 263 may be stored in the prediction model storage unit 124 together with the prediction model 125-1. Next, the determination unit 264 determines whether to reduce the number of parameter types based on the prediction accuracy of the prediction model 125-1 (step S103). When generating the first prediction model 125-1, since there is no previously generated prediction model, the process proceeds to step S104.

[0078] The ranking unit 265 uses multiple algorithms to rank the contributions of multiple types of parameters (step S104). Next, the ranking unit 265 averages the contribution rankings of each algorithm for each type of parameter and performs ranking (step S105). In this embodiment, the ranking unit 265 uses AIC, BIC, Lasso regression, a combination of ANN and SHAP, ANN (weighted), and random forest (FI) as algorithms. The average may be an arithmetic mean, a geometric mean, a harmonic mean, or a weighted mean. FIG. 7 is a table showing 16 parameters and their rankings. FIG. 7 also shows the prediction accuracy of the prediction model 125-1. The ranking of the parameters shown in FIG. 7 is the ranking after averaging for each type of parameter. The ranking of the parameters shown in FIG. 7 corresponds to the ranking after averaging for each type of parameter shown in FIG. 5. Based on this result, the parameter selection unit excludes the lowest order parameters (Cr, Ti) (step S106).

[0079] Next, the teacher data generation unit 267 generates teacher data 123-2 (see FIG. 4 ) that associates a parameter set 123a-2 prepared based on 14 parameters excluding the lowest-order parameters (Cu, Ti) with the characteristic test results 123b (here, toughness test results) (step S107). The teacher data 123-2 generated by the teacher data generation unit 267 is stored in the teacher data storage unit 122. Then, the process returns to step S101. Teacher data 123 and a prediction model 125 are generated depending on the number of times the processes of steps S103 to S107, S101, and S102 are repeated.

[0080] FIG. 8 is a table showing the ranking of parameters when the number of parameter types is reduced 11 times. FIG. 8 shows the ranking of parameters after steps S101 and S102 are performed, followed by 10 repetitions of steps S103 to S107, S101, and S102. The parameter ranking shown in FIG. 8 is the ranking after ranking multiple algorithms and averaging the rankings for each parameter type. FIG. 9 is a block diagram showing the teacher data storage unit 122 and the prediction model storage unit 124 stored in the memory unit 12. FIG. 9 shows the teacher data storage unit 122 and the prediction model storage unit 124 after 10 repetitions of steps S103 to S107, S101, and S102. As shown in FIG. 9, the teacher data storage unit 122 stores 11 teacher data 123, and the prediction model storage unit 124 stores 11 prediction models 125 corresponding to the 11 teacher data 123.

[0081] As shown in FIG. 8 , the training data 123-1 is associated with a parameter set 123a-1 using 16 parameters and characteristic test results 123b. FIG. 8 also lists the number of parameters and the error of the prediction model 125 corresponding to the number of parameters. The training data storage unit 122 stores a prediction model 125-1 generated from the training data 123-1. Each time the generation of the prediction model 125 and the training data 123 is repeated, the training data storage unit 122 stores training data 123 generated according to the number of parameters. The prediction model storage unit 124 stores a prediction model 125 generated from the training data 123. In other words, each time the number of parameters is reduced in a stepwise manner, training data 123 and a prediction model 125 are generated.

[0082] As shown in FIG. 9 , teacher data 123-9 is associated with parameter set 123a-9 using five parameters and characteristic test results 123b. Prediction model 125-9 is generated from teacher data 123-9. Teacher data 123-11 is associated with parameter set 123a-11 using three parameters and characteristic test results 123b. Prediction model 125-11 is generated from teacher data 123-11. Note that in FIG. 9 , teacher data 123 and prediction models 125 corresponding to teacher data 123 for cases where the number of parameters is 14, 13, 12, 11, 10, 9, 7, or 4 are not shown.

[0083] In this way, by repeating the processes of steps S103 to S107, S101, and S102, the error of the prediction model 125 decreases. In step S103, the determination unit 264 determines whether to reduce the number of parameter types based on the magnitude of the error output by the prediction model 125. In this embodiment, the error of the most recently generated prediction model is compared with the error of the prediction model generated two models prior to the most recently generated prediction model. If the error of the most recently generated prediction model 125 is greater than the error of the prediction model generated two models prior to the most recently generated prediction model, the determination unit 264 determines not to reduce the number of parameter types. As shown in FIG. 9, the error of the most recently generated prediction model 125-11 is 11.0, as shown in FIG. 8, while the error of the prediction model 125-9 generated two models prior to the most recently generated prediction model is 9.7. In other words, the error of the most recently generated prediction model 125-11 is greater than the error of the prediction model 125-9 generated two models prior to the most recently generated prediction model. In such a case, the determination unit 264 determines not to reduce the number of parameter types (step S103; NO). The error of the prediction model 125-9 is the smallest among the errors of the prediction models 125-1 to 125-11.

[0084] Finally, in step S108, the prediction model providing unit 268 compares the error of the most recently generated prediction model 125-11 with the error of the prediction model 125-9 generated two models earlier. If the error of the most recently generated prediction model 125-11 is greater than that of the prediction model 125-9 generated two models earlier, the prediction model providing unit 268 selects the prediction model 125-9 generated two models earlier. The prediction model providing unit 268 provides information for selecting the prediction model 125-9, thereby completing the processing of the prediction model design method. Here, the information for selecting the prediction model 125-9 may be, for example, the errors of the prediction models 125-9 to 125-11. Furthermore, the prediction model providing unit 268 may display the prediction model 125-9, among the prediction models 125-9 to 125-11, in a different format from the prediction models 125-10 and 125-11. The prediction model 125-9 provided by the prediction model providing unit 268 is stored in the storage unit and is used in the prediction function described later.

[0085] According to a method for designing a prediction model according to one embodiment of the present invention, a prediction model is generated while gradually reducing the number of parameters used as training data based on multiple algorithms. This makes it possible to select appropriate parameters depending on the number of data sets for test results of specific properties of spring steel to be predicted. This reduces the time required to create many data sets that associate measured values ​​of many types of parameters with test results of mechanical properties. Furthermore, a highly accurate prediction model can be designed even with a small number of data sets.

[0086] As described above, according to the method for designing a prediction model according to one embodiment of the present invention, it is possible to select appropriate parameters depending on the type of mechanical property even with a small data set.

[0087] Furthermore, in the method for designing a prediction model according to one embodiment of the present invention, the prediction model 125 may be regenerated from scratch when a predetermined number of data sets, each of which associates multiple types of parameters that indicate the characteristics of spring steel with the test results thereof, has been accumulated in the spring steel database. As the number of accumulated data sets increases, the prediction accuracy of the prediction model 125 can be further improved.

[0088] Furthermore, a prediction model generation method for generating a prediction model can be provided by using the prediction model design method according to one embodiment of the present invention.

[0089] 5. Overview of Prediction Function The prediction function according to one embodiment of the present invention predicts test results for predetermined properties of spring steel from multiple parameters that indicate the characteristics of spring steel using a prediction model 125 that has previously learned the relationship between multiple parameters that indicate the characteristics of spring steel and test results for predetermined properties of spring steel. This makes it possible to select appropriate input parameters for test results for predetermined properties of spring steel even with a small data set, thereby making it possible to accurately predict test results for mechanical properties of spring steel that have previously been difficult to predict. Furthermore, by predicting the mechanical properties of newly manufactured spring steel using this prediction model, it is possible to shorten development time and reduce costs.

[0090] The control unit 11 executes a prediction program stored in the storage unit 12 and realizes a prediction function in accordance with instructions written in the program. In this embodiment, a case where the prediction function is executed by the control unit 11 of the prediction model design device 10 will be described, but one embodiment of the present invention is not limited to this. The control unit 11 may be a control unit of a communication terminal communicably connected to the prediction model design device, or a control unit of a cloud server. There are no particular limitations on the device as long as it is capable of realizing the prediction function according to one embodiment of the present invention.

[0091] 10 is a block diagram illustrating a prediction device that realizes a prediction function according to an embodiment of the present invention. The prediction function is realized by an input unit 271, a prediction unit 272, and an output unit 273. The prediction model 125-9 described in FIGS. 6 to 9 is applied to the prediction unit 272.

[0092] The input unit 271 receives input of multiple types of parameters, including Ni, grain size, hardness, residual γ, and carbide area ratio. The multiple types of parameters are parameters selected when generating the prediction model 125. When these parameters are provided to the prediction model 125-9 of the prediction unit 272, the prediction model 125-9 can output a predicted value of the toughness test result. The output unit 273 outputs the predicted value of the toughness test result. The output destination of the predicted value of the toughness test result by the output unit 273 is not particularly limited, and may be a display device, a control unit of a communication terminal connected to the prediction model design device so as to be able to communicate with the prediction model, or a control unit of a cloud server.

[0093] <Modifications> The present invention is not limited to the above-described embodiment, and includes various other modifications. For example, the above-described embodiment has been described in detail to clearly explain the present invention, and is not necessarily limited to an embodiment having all of the described configurations. Other configurations may be added to, deleted from, or substituted for part of the configuration of the embodiment and part of the modifications described below. Modifications are described below.

[0094] (1) In this embodiment, the case where the training data contains one test result (toughness) related to the predetermined property of spring steel has been described. However, one embodiment of the present invention is not limited to this. For example, there may be two or more test results related to the predetermined property of spring steel. For example, a prediction model may be generated using multiple types of parameters that indicate the characteristics of spring steel and test results related to the predetermined property of spring steel (fatigue and sag) as training data. The number and combination of test results related to the predetermined property of spring steel are arbitrary. Furthermore, when predicting test results for the predetermined property of multiple spring steels, a prediction model may be generated for each test result of the predetermined property of spring steel.

[0095] (2) The predetermined termination condition in the determination unit 264 may be determined as a result of comparing the error of the most recently generated prediction model 125 with the error of the prediction model 125 generated immediately before and determining that the error of the most recently generated prediction model 125 is larger than the error of the prediction model 125 generated immediately before. For example, the determination unit 264 may compare the error of the most recently generated prediction model 125-10 with the error of the prediction model 125-9 generated immediately before and determine that the number of parameter types should not be reduced if the error of the most recently generated prediction model 125-10 is larger than the error of the prediction model 125-9 generated immediately before. Alternatively, the predetermined termination condition may be determined as a result of the number of parameters remaining being reduced to one.

[0096] (3) As a predetermined provision condition for the prediction model providing unit 268, for example, the error of the most recently generated prediction model 125 may be compared with the error of the prediction model 125 generated immediately before, and if the error of the most recently generated prediction model 125 is larger than the error of the prediction model 125 generated immediately before, the prediction model providing unit 268 may select the prediction model 125. For example, the prediction model providing unit 268 may compare the error of the most recently generated prediction model 125-10 with the error of the prediction model 125-9 generated immediately before, and if the error of the most recently generated prediction model 125-10 is larger than the error of the prediction model 125-9 generated immediately before, the prediction model providing unit 268 may select the prediction model 125-9. Alternatively, the number of multiple types of parameters may be gradually reduced until the number of parameters is reduced to one, and the prediction model 125 with the smallest error may be selected from among the multiple generated prediction models 125. Furthermore, for example, a predetermined provision condition may be to select the prediction model 125 with the smallest error through the prediction model 125 with the fewest errors. The information for selecting the prediction model 125 may be, for example, the errors corresponding to all of the generated prediction models 125, or the errors of the prediction model 125 with the smallest error and the prediction model 125 with the second smallest error, or the errors of the prediction model 125 with the smallest error through the prediction model 125 with the fifth smallest error.

[0097] The above-described embodiments of the present invention may be combined as appropriate as long as they are not mutually inconsistent. Furthermore, even if a person skilled in the art adds or deletes components or modifies the design based on the embodiments, such addition or deletion of components is included in the scope of the present invention as long as the gist of the present invention is maintained.

[0098] Furthermore, even if there are other effects and advantages different from those brought about by the above-mentioned embodiments, those that are clear from the description in this specification or that can be easily predicted by a person skilled in the art are naturally understood to be brought about by the present invention.

[0099] 10: Design device, 11: Control unit, 12: Storage unit, 13: Communication unit, 15: Display unit, 121: Spring steel database, 122: Training data storage unit, 123, 123-1, 123-2, 123-9, 123-11: Training data, 123a, 123a-1, 123a-2, 123a-9, 123a-11: Parameter set, 123b: Characteristics test results, 124: Prediction model Storage unit, 125, 125-1, 125-9, 125-11: prediction model, 126: prediction model design program, 127: prediction program, 128: algorithm storage unit, 261: prediction model generation unit, 262: prediction unit, 263: error calculation unit, 264: judgment unit, 265: ranking determination unit, 266: parameter selection unit, 267: training data generation unit, 268: prediction model provision unit

Claims

1. A method for designing a predictive model executed by a computer, comprising: generating a predictive model using training data in which a parameter set prepared based on multiple types of parameters that indicate the characteristics of spring steel is associated with test results for predetermined properties of the spring steel; ranking the parameter types of the training data used in the predictive model previously generated using multiple algorithms that output a ranking of the contribution of each parameter type to the test results; and regenerating a predictive model using training data in which a parameter set with a reduced number of parameter types is associated with test results for the predetermined properties of the spring steel based on the ranking of the parameter types, until a predetermined termination condition is met; and providing information for selecting at least one of the multiple predictive models based on the difference between the predicted value of the test result output for each of the multiple generated predictive models and the actual value of the test result.

2. The method for designing a forecast model according to claim 1, wherein selecting at least one from the plurality of forecast models includes selecting the forecast model with the smallest difference value.

3. A method for designing a predictive model as described in claim 1, wherein ranking the plurality of types of parameters for each of a plurality of algorithms includes averaging the rankings of the plurality of algorithms for each of the plurality of types, and ranking the averaged parameter types.

4. A method for designing a predictive model as described in claim 1, wherein the parameter set with the reduced number of parameter types excludes the lowest-ranked parameter type in the parameter set before the number of parameter types was reduced from the ranking of the averaged parameter types.

5. A method for designing a predictive model as described in claim 1, wherein providing information for selecting at least one of the plurality of predictive models includes providing the second previous predictive model when, when comparing a first difference between the actual values ​​and predicted values ​​of the test results of the most recently generated predictive model and a second difference between the actual values ​​and predicted values ​​of the test results of the predictive model two models before the most recently generated predictive model, the first difference is greater than the second difference.

6. The method for designing a predictive model according to claim 1, wherein the plurality of algorithms are at least two selected from AIC, BIC, Lasso regression, ANN, and random forest, XGBoost, LightGBM, a combination of ANN and SHAP, a combination of random forest and SHAP, a combination of XGBoost and SHAP, and a combination of LightGBM and SHAP.

7. The method for designing a predictive model according to claim 1, wherein the plurality of algorithms include AIC, BIC, Lasso regression, a combination of ANN and SHAP, ANN, and random forest.

8. The method for designing a prediction model according to claim 1, wherein the plurality of types of parameters include at least one of a parameter representing crystal grain size, a parameter representing carbide morphology, a parameter representing dislocation density, a parameter representing the amount of solute carbon, a parameter representing residual γ, a parameter representing hardness, and a parameter relating to the manufacturing conditions of spring steel.

9. The method for designing a prediction model described in claim 8, wherein when the multiple types of parameters are parameters representing the crystal grain size, they include at least one of the circle equivalent diameter of the crystal grain size, the circularity of the crystal grain size, and the aspect ratio of the crystal grain size.

10. A method for designing a prediction model as described in claim 8, wherein when the multiple types of parameters are parameters representing the morphology of the carbides, the parameters representing the morphology of the carbides include at least one of a circle equivalent diameter of the carbides, a number density of the carbides, an aspect ratio of the carbides, an area ratio of the carbides, and a volume ratio of the carbides.

11. A method for designing a prediction model as described in claim 1, wherein the multiple types of parameters include the circle-equivalent diameter of the crystal grain size, the area ratio of carbides, the aspect ratio of carbides, the number density of carbides, the circle-equivalent diameter of carbides, residual γ, hardness, and the elements that make up the spring steel.

12. The method for designing a prediction model according to claim 1, wherein the predetermined property of the spring steel is at least one of tensile strength, toughness, sag, and fatigue.

13. A program for causing a computer to execute the method for designing a prediction model according to any one of claims 1 to 12.

14. A method for generating a predictive model, which generates a predictive model using the method for designing a predictive model according to any one of claims 1 to 12.

15. A prediction device that predicts test results regarding predetermined properties of the spring steel using a prediction model generated using the method for generating a prediction model according to claim 14.

16. A prediction model design device comprising: a prediction model generation unit that generates a prediction model using training data that associates a parameter set prepared based on multiple types of parameters that indicate the characteristics of spring steel with test results related to specified properties of the spring steel; a ranking determination unit that ranks the types of parameters in the training data used in the most recently generated prediction model using multiple algorithms that output a ranking of the contribution of each parameter type to the test results; a training data generation unit that generates training data that associates a parameter set with a reduced number of parameter types with test results related to the specified properties of the spring steel based on the ranking of the parameter types; and a prediction model provision unit that provides information for selecting at least one prediction model from the multiple prediction models generated based on the difference between the predicted values ​​of the test results output for each of the multiple prediction models and the actual measured values ​​of the test results.

Images