JP2022087429A - Physical property value estimate system and physical property value estimate method - Google Patents

Abstract

To estimate a physical property value of composite material with high accuracy.SOLUTION: A physical property value estimate device 100 comprises: a known data input unit 301 to which known data is inputted, the known data including a compounding ratio of materials associated with a physical property value corresponding to the compounding ratio; a level data generation unit 302 which adds a level to which the physical property value belongs from among a plurality of preset levels to the known data and generates level data including the compounding ratio, the physical property value and the level in association with each other; a level approximation function generation unit 303 which generates a level approximation function which outputs the level to which the physical property value corresponding to a first compounding ratio belongs when the first compounding ratio having unknown correspondence with the physical property value is inputted, based on the level data; a compounding ratio input unit 305 in which the first compounding ratio is inputted; and a level estimation unit 306 which estimates the level corresponding to the first compounding ratio, based on the first compounding ratio and the level approximation function.SELECTED DRAWING: Figure 2

Description

The present invention relates to a physical property value estimation system and a physical property value estimation method, and relates to, for example, a technique applicable to a technique for estimating a physical property value according to a blending ratio of a resin composite material.

Japanese Unexamined Patent Publication No. 2020-77257 (Patent Document 1) describes a technique for estimating the physical properties of a material to be designed by utilizing artificial intelligence.

Japanese Unexamined Patent Publication No. 2020-77257

In recent years, composite materials have been developed in which new performance is imparted to the characteristics of the resin itself by compounding a plurality of types of resins and compounding agents. In this regard, in order to develop a new composite material, it is necessary to develop the material while adjusting the composition ratio of each composition until the composite material has desired properties. For this reason, the development of composite materials requires enormous costs. Therefore, from the viewpoint of improving the efficiency of composite material development, it is desirable to be able to estimate the physical properties of the composite material to be tested at the experimental planning stage to some extent. However, for example, in composite materials for electric wire coating materials, there are many types of compounding agents, and the physical property values may change significantly depending on the compounding composition ratio. From this, it is difficult to estimate the physical property values of the composite material. From the above, a technique capable of estimating the physical property value of the composite material with high accuracy is desired.

The physical property value estimation system in one embodiment is configured to input known data in which a blending ratio of a material composed of a plurality of types of compositions and a physical property value corresponding to the blending ratio are related to each other. Level data generation configured to generate level data in which the blending ratio, the physical property value, and the level are related by adding the level to which the physical property value belongs from a plurality of preset levels to the known data. Based on the part and the level data, it is configured to generate a level approximation function that outputs the level to which the physical property value corresponding to the first compounding ratio belongs when the first compounding ratio whose correspondence with the physical property value is unknown is input. Based on the level approximation function generation unit, the blending ratio input section configured to input the first blending ratio, and the first blending ratio and the level approximation function, the level corresponding to the first blending ratio is determined. It includes a level estimation unit configured to estimate.

Further, the method for estimating the physical property value in one embodiment includes a known data input step of inputting known data in which known data relating the blending ratio of a material composed of a plurality of types of compositions and the physical property value corresponding to the blending ratio are input. Based on the level data generation process and the level data generation process, which adds the level to which the physical property value belongs from a plurality of preset levels to the known data and generates the level data in which the blending ratio, the physical property value, and the level are related. , A level approximation function generation step that generates a level approximation function that outputs the level to which the physical property value corresponding to the first compounding ratio belongs when the first compounding ratio whose correspondence with the physical property value is unknown is input, and the first compounding ratio. A blending ratio input step for inputting the above, and a level estimation step for estimating the level corresponding to the first blending ratio based on the first blending ratio and the approximate function for the level are provided.

According to one embodiment, the physical property values of the composite material can be estimated with high accuracy.

It is a figure which shows an example of the hardware composition of the physical characteristic value estimation apparatus. It is a functional block diagram which shows the function of a physical property value estimation device. It is a figure which shows the structural example of the approximate function generation part for a level. It is a figure which shows the structural example of the approximate function generation part for a physical characteristic value. It is a flowchart explaining the generation operation of the approximation function for a level. It is a flowchart explaining the generation operation of the approximation function for a physical characteristic value. It is a flowchart explaining the estimation operation of the physical characteristic value corresponding to the compounding ratio. It is a functional block diagram which shows the example which configures the physical characteristic value estimation system by the physical characteristic value estimation apparatus and the approximate function generation apparatus. It is a graph which shows the correspondence relation between the physical property value estimated by the approximation function for a single physical characteristic value, and the measured value. It is a graph which shows the correspondence relationship between the physical property value estimated by the approximation function for the physical property value for each level, and the measured value. It is a graph which shows the correspondence relationship between the physical property value estimated by the approximation function for the physical characteristic value of the 1st level, and the measured value. It is a graph which shows the correspondence relationship between the physical property value estimated by the approximation function for the 2nd level physical characteristic value, and the measured value. It is a figure which shows the functional block composition of a level estimation apparatus. It is a flowchart explaining the estimation operation of the level to which a physical characteristic value corresponding to a compounding ratio belongs.

In all the drawings for explaining the embodiments, the same members are designated by the same reference numerals in principle, and the repeated description thereof will be omitted. In addition, in order to make the drawing easier to understand, hatching may be added even if it is a plan view.

The technical idea in the present embodiment is an idea regarding a physical characteristic value estimation system that estimates a physical characteristic value corresponding to a compounding ratio in a composite material in which a plurality of types of resins and compounding agents are composited.

Here, as the composite material, for example, an electric wire coating material containing a resin or a compounding agent can be mentioned, and as a physical property value, for example, the elongation property of the material can be mentioned.

The resin is, for example, a polyolefin such as high-density polyethylene, low-density polyethylene, or an ethylene-acrylic acid copolymer, or an elastoma such as chlorinated polyethylene. On the other hand, examples of the compounding agent include fillers such as talc, calcium carbonate and silica, plasticizers, cross-linking agents and stabilizers. However, the types and numbers of compositions such as resins and compounding agents constituting the composite material are not limited.

<Explanation of related technologies>
First, a related technique related to a physical characteristic value estimation system for estimating a physical characteristic value corresponding to a blending ratio will be described. The "related technique" referred to in the present specification is not a known technique, but is a technique having a problem found by the present inventor and is a technique which is a premise of the present invention.

For example, as a physical property value estimation system, the following related techniques can be considered. That is, the relationship of generating an approximate function by learning a neural network in which the mixing ratio and the physical property value corresponding to the mixing ratio are known as teacher data, the input is the mixing ratio, and the output is the physical property value. Technology is conceivable.

In this related technique, for example, when a compounding ratio whose correspondence with the physical property value is unknown is input to the approximate function, the physical property value is output from the approximate function. Therefore, if an approximation function with high estimation accuracy can be obtained, it is possible to estimate the physical property value with high accuracy, which is highly likely to realize the correspondence with the physical property value at an unknown compounding ratio.

In this regard, the present inventor trains a neural network using these data as teacher data when the numerical range of the physical property values is wide in the data in which the compounding ratio and the physical property value corresponding to the compounding ratio are known. It was newly found that it is difficult to estimate the physical property value with high accuracy, which is likely to be realized by the unknown compounding ratio in the approximate function generated by. That is, in the related technique, a "single approximation function" is generated by using data having a wide numerical range of physical property values as training data, but the "single approximation function" generated in this way has accurate physical properties. The present inventor has newly found that it is difficult to estimate the value. That is, there is room for improvement in the related technology from the viewpoint of estimating the physical property value corresponding to the blending ratio with high accuracy.

Therefore, in the present embodiment, some measures are taken for the room for improvement existing in the related techniques. Hereinafter, the technical idea in the present embodiment to which this device has been devised will be described.

<Basic idea in the embodiment>
The basic idea in the present embodiment is to divide the numerical range of the physical property value into a plurality of ranges in the data in which the blending ratio and the physical property value corresponding to the blending ratio are known, and to divide the data belonging to each of the divided multiple ranges. The idea is to generate an approximate function peculiar to each of a plurality of ranges as teacher data. That is, the basic idea in the present embodiment is an idea to generate a different approximation function for each of a plurality of ranges. That is, according to the basic idea, instead of generating a "single approximation function" as in the related technology, a "multiple approximation functions" are generated.

According to this basic idea, since data in a narrow numerical range belonging to each of a plurality of ranges is used as teacher data, the approximate function generated by training the neural network using these data as teacher data is a physical property value. It is possible to estimate with high accuracy the physical property values that are likely to be realized with an unknown compounding ratio.

Based on this basic idea, there is a high possibility that the correspondence with the physical property value will be realized with an unknown compounding ratio. It is necessary to accurately estimate whether it belongs to the range. This is because if it is not possible to accurately estimate which range of multiple ranges the compounding ratio whose correspondence with the physical property value is unknown belongs to, an approximation function in a range different from the actual one will be used. As a result, it becomes impossible to estimate the physical property value with high accuracy.

Therefore, in embodying the basic idea, it is important to accurately estimate which range of multiple ranges the compounding ratio whose correspondence with the physical property value is unknown belongs to, and this point is also taken into consideration. We have devised ways to embody the idea.

In the following, an example in which a physical characteristic value estimation system embodying this basic idea is mainly composed of a single computer will be described. However, the physical characteristic value estimation system in the present embodiment is a distributed system composed of a plurality of computers. It is also possible to realize with.

<Structure of physical property value estimation device>
<< Hardware configuration >>
First, the hardware configuration of the physical property value estimation device in the present embodiment will be described.

FIG. 1 is a diagram showing an example of a hardware configuration of the physical property value estimation device 100 according to the present embodiment. The configuration shown in FIG. 1 is merely an example of the hardware configuration of the physical property value estimation device 100, and the hardware configuration of the physical property value estimation device 100 is not limited to the configuration shown in FIG. Other configurations may be used.

In FIG. 1, the physical property value estimation device 100 includes a CPU (Central Processing Unit) 101 that executes a program. The CPU 101 is electrically connected to, for example, a ROM (Read Only Memory) 102, a RAM (Random Access Memory) 103, and a hard disk device 112 via a bus 113, and controls these hardware devices. It is configured as follows.

The CPU 101 is also connected to an input device and an output device via the bus 113. Examples of the input device include a keyboard 105, a mouse 106, a communication board 107, a scanner 111, and the like. On the other hand, examples of the output device include a display 104, a communication board 107, a printer 110, and the like. Further, the CPU 101 may be connected to, for example, a removable disk device 108 or a CD / DVD-ROM device 109.

The physical property value estimation device 100 may be connected to, for example, a network. For example, when the physical property value estimation device 100 is connected to another external device via a network, the communication board 107 constituting a part of the physical property value estimation device 100 may be a LAN (local area network) or a WAN (wide area network). It is connected to the network) or the Internet.

The RAM 103 is an example of a volatile memory, and the recording medium of the ROM 102, the removable disk device 108, the CD / DVD-ROM device 109, and the hard disk device 112 is an example of the non-volatile memory. The storage device of the physical property value estimation device 100 is configured by these volatile memories and non-volatile memories.

The hard disk device 112 stores, for example, an operating system (OS) 201, a program group 202, and a file group 203. The program included in the program group 202 is executed by the CPU 101 while using the operating system 201. Further, in the RAM 103, at least a part of the program of the operating system 201 or the application program to be executed by the CPU 101 is temporarily stored, and various data necessary for processing by the CPU 101 are stored.

The BIOS 102 (Basic Input Output System) program is stored in the ROM 102, and the boot program is stored in the hard disk device 112. When the physical property value estimation device 100 is started, the BIOS program stored in the ROM 102 and the boot program stored in the hard disk device 112 are executed, and the operating system 201 is started by the BIOS program and the boot program.

The program group 202 stores a program that realizes the function of the physical property value estimation device 100, and this program is read and executed by the CPU 101. Further, in the file group 203, information, data, signal values, variable values and parameters indicating the result of processing by the CPU 101 are stored as each item of the file.

The file is recorded on a recording medium such as a hard disk device 112 or a memory. Information, data, signal values, variable values and parameters recorded in a recording medium such as a hard disk device 112 or a memory are read into a main memory or a cache memory by the CPU 101, and are extracted, searched, referenced, compared, calculated, processed, and processed. It is used for the operation of the CPU 101 represented by editing, output, printing, and display. For example, during the operation of the CPU 101 described above, information, data, signal values, variable values and parameters are temporarily stored in the main memory, registers, cache memory, buffer memory and the like.

The function of the physical property value estimation device 100 may be realized by the firmware stored in the ROM 102, or only the software, only the hardware represented by the element, the device, the board, and the wiring, and the software and the hardware. It may be realized in combination, and further in combination with firmware. The firmware and software are recorded as a program on a recording medium typified by a hard disk device 112, a removable disk, a CD-ROM, a DVD-ROM, or the like. The program is read and executed by the CPU 101. That is, the program causes the computer to function as the physical property value estimation device 100.

As described above, the physical property value estimation device 100 includes a CPU 101 as a processing device, a hard disk device 112 and a memory as a storage device, a keyboard 105 as an input device, a mouse 106, a communication board 107, a display 104 as an output device, and a printer 110. , A computer including a communication board 107. The function of the physical property value estimation device 100 is realized by using a processing device, a storage device, an input device, and an output device.

<< Functional block configuration >>
Next, the functional block configuration of the physical property value estimation device 100 will be described.

FIG. 2 is a functional block diagram showing the functions of the physical property value estimation device.

The physical property value estimation device 100 includes a known data input unit 301, a level data generation unit 302, a level approximation function generation unit 303, an accuracy determination unit 304, a blending ratio input unit 305, a level estimation unit 306, and a level. It has an output unit 307, an approximate function generation unit 308 for physical property values, a physical property value estimation unit 309, a physical property value output unit 310, and a data storage unit 311.

The known data input unit 301 is configured to input known data.

Here, the "known data" is data in which the blending ratio and the physical property value are related, and is defined as data in which the blending ratio and the physical property value are known.

The known data input to the known data input unit 301 is stored in the data storage unit 311.

The level data generation unit 302 is configured to generate level data from known data. Here, the "level data" refers to data in which a level indicating to which numerical range the physical property value of the known data belongs is added to the known data. Specifically, the "level data" is defined as data in which the blending ratio, the physical property value, and the level are related by adding the level to which the physical property value belongs from a plurality of preset levels to the known data. ..

For example, as known data, the first known data in which the blending ratio is "first blending ratio" and the physical property value is "100", and the second known data in which the blending ratio is "second blending ratio" and the physical property value is "500". And, it is assumed that there is a third known data in which the blending ratio is "third blending ratio" and the physical property value is "200". At this time, in the level data generation unit 302, for example, the threshold value of the physical property value is set to "300", the data having the physical property value "300" or less is associated with the first level, and the physical property value is higher than "300". It is assumed that the level is determined so that a large amount of data is associated with the second level.

In this case, since the level data generation unit 302 belongs to the first level in which the physical property value of the first known data is "100" and is "300" or less, the first known data and the first level are associated with each other and blended. Generates first level data in which the ratio is the "first blending ratio", the physical property value is "100", and the level is the "first level".

Further, since the level data generation unit 302 belongs to the second level in which the physical property value of the second known data is "500" and is larger than "300", the second known data and the second level are associated and blended. Generates second level data in which the ratio is the "second blending ratio", the physical property value is "500", and the level is the "second level".

Further, since the level data generation unit 302 belongs to the first level in which the physical property value of the third known data is "200" and is "300" or less, the third known data is related to the first level and the blending ratio is increased. Generates third level data in which "third compounding ratio", physical characteristic value is "200", and level is "first level".

As a result, the first level data and the third level data become the data belonging to the first level, and the second level data becomes the data belonging to the second level.

In this way, the level data generated by the level data generation unit 302 is classified by level and stored in the data storage unit 311, for example.

The level approximation function generation unit 303 has a function of generating a level approximation function based on the level data generated by the level data generation unit 302. That is, the level approximation function generation unit 303 has a function of generating a level approximation function that associates the blending ratio with the level to which the physical property value of the material at this blending ratio belongs. Specifically, as shown in FIG. 3, the level approximation function generation unit 303 uses level data as teacher data, and trains a neural network in which an input is a blending ratio and an output is a level, thereby performing level approximation. It is configured to generate a function.

Here, the "approximate function for level" is defined as a function that outputs the level to which the physical property value corresponding to the blending ratio belongs when the blending ratio is input. That is, the "approximate function for level" is a function that outputs the level to which the physical property value estimated to be realized by this compounding ratio belongs when the compounding ratio whose correspondence with the physical property value is unknown is input. Defined. In this way, the level approximation function can be said to be a function used to estimate the level corresponding to the compounding ratio whose correspondence with the physical property value is unknown.

The accuracy determination unit 304 has a function of determining whether the accuracy of the level approximation function generated by the level approximation function generation unit 303 is good.

The level approximation function is better, for example, as the probability that the level output when a compounding ratio whose correspondence with the physical property value is unknown is the level to which the actual physical property value at this compounding ratio belongs increases. It can be said that it has accuracy.

Therefore, the accuracy determination unit 304 is configured to determine whether or not the accuracy of the level approximation function is good by using the level data. For example, the accuracy determination unit 304 inputs the mixing ratio of the level data into the approximation function for the level, and when the output level has a probability of reproducing the level of the level data of 90% or more, the approximation for the level Judge that the accuracy of the function is good. The level approximation function determined by the accuracy determination unit 304 to have good accuracy is stored in the data storage unit 311.

The blending ratio input unit 305 inputs a blending ratio to be evaluated whose correspondence with the physical property value is unknown.

The level estimation unit 306 estimates the level corresponding to the input blending ratio based on the blending ratio input to the blending ratio input unit 305 and the level approximation function generated by the level approximation function generation unit 303. Has a function. That is, the level estimation unit 306 is configured to estimate the level corresponding to the blending ratio to be evaluated by using the level approximation function generated by the level approximation function generation unit 303.

The level output unit 307 outputs the level estimated by the level estimation unit 306.

Next, the physical property value approximation function generation unit 308 has a function of generating a physical property value approximation function that associates the blending ratio with the physical property value of the material for each of the plurality of levels. That is, the physical property value approximation function generation unit 308 is configured to generate different physical property value approximation functions for each different level. Therefore, there are as many approximation functions for physical properties as there are levels generated by the approximation function generation unit 308 for physical characteristics.

Specifically, as shown in FIG. 4, the approximation function generation unit 308 for the physical property value uses the level data whose level is the Nth level as the teacher data, the input as the blending ratio, and the output as the physical property. It has a function of generating an approximate function for physical property values at the Nth level by training a neural network as a value. For example, the approximation function generation unit 308 for the physical property value uses the level data whose level is each of the first level to the Nth level among the level data as the teacher data, the input as the blending ratio, and the output as the physical property value. By training the neural network, it is configured to generate an approximation function for physical property values at each of the first level to the Nth level. That is, when the level data is classified into N levels, the approximation function generation unit 308 for physical characteristic values generates N different approximation functions for physical characteristics values.

Here, the "approximate function for the physical property value" is defined as a function that outputs the physical property value according to the blending ratio when the blending ratio is input. That is, the "approximate function for physical property values" is defined as a function that outputs a physical property value that is presumed to be realized by this compounding ratio when a compounding ratio whose correspondence with the physical property value is unknown is input. To. As described above, the approximation function for the physical property value can be said to be a function used to estimate the physical property value corresponding to the blending ratio whose correspondence with the physical property value is unknown.

For each of the plurality of physical property value approximation functions generated by the physical property value approximation function generation unit 308, for example, the accuracy determination unit 304 described above determines whether the accuracy is good or bad. For example, the accuracy determination unit 304 inputs the blending ratio of the level data classified into the Nth level into the approximation function for the physical property value corresponding to the Nth level, and the physical property value output is the physical property of the level data. When the probability of reproducing the value is 90% or more, it is judged that the accuracy of the approximation function for the physical characteristic value is good. The approximation function for the physical characteristic value determined by the accuracy determination unit 304 to have good accuracy is stored in the data storage unit 311.

The physical characteristic value estimation unit 309 has a function of estimating the physical characteristic value corresponding to the blending ratio of the evaluation target based on the approximation function for the physical characteristic value at the level estimated by the level estimation unit 308.

For example, when the level estimated by the level estimation unit 308 is the first level, the physical property value estimation unit 309 is the first of a plurality of physical property value approximation functions generated by the physical property value approximation function generation unit 308. The physical property value corresponding to the compounding ratio to be evaluated will be estimated using the approximation function for the physical property value of the level.

The physical characteristic value output unit 310 outputs the physical characteristic value estimated by the physical characteristic value estimation unit 309.

In this way, the physical property value estimation device 100 is configured.

<Operation of physical property value estimation device>
The physical property value estimation device 100 in the present embodiment is configured as described above, and its operation will be described below. The operation of the physical property value estimation device 100 includes "the operation of generating the approximate function for the level", "the operation of generating the approximate function for the physical property value", and "the operation of estimating the physical property value corresponding to the blending ratio of the evaluation target". Therefore, these operations will be described below.

<< Generation operation of approximation function for level >>
FIG. 5 is a flowchart illustrating the generation operation of the level approximation function.

In FIG. 5, first, the known data input unit 301 inputs the known data (S101). Then, the known data input to the known data input unit 301 is stored in the data storage unit 311. Next, the level data generation unit 302 acquires the known data stored in the data storage unit 311 and generates level data from the known data based on the preset level classification standard (S102). For example, the preset level classification standard is set to the first level when the physical property value is equal to or less than a predetermined threshold value, while it is set to the second level when the physical property value is larger than the predetermined threshold value. You can think of the criteria of setting. The preset level classification standard may be not only a standard for classifying into two types of levels but also a standard for classifying into two or more types of levels.

After that, the level data generated by the level data generation unit 302 is stored in the data storage unit 311. Subsequently, the level approximation function generation unit 303 generates a level approximation function based on the level data generated by the level data generation unit 302 (S103). Specifically, the level approximation function generation unit 303 generates a level approximation function by learning a neural network in which the level data is used as teacher data, the input is the blending ratio, and the output is the level (FIG. 3). reference).

Next, the accuracy determination unit 304 determines whether the accuracy of the level approximation function generated by the level approximation function generation unit 303 is good (S104). Then, when the accuracy determination unit 304 determines that the accuracy of the level approximation function is good (S105), the physical property value estimation device 100 adopts the level approximation function generated by the level approximation function generation unit 303. (S106), and the determined level approximation function is stored in the data storage unit 311. On the other hand, when the accuracy determination unit 304 determines that the accuracy of the level approximation function is not good (S105), the physical property value estimation device 100 adds new known data (S107), and includes the added known data. Regenerate the level approximation function based on the known data. In this case, since the teacher data increases, it is considered that the accuracy of the regenerated level approximation function can be improved. This operation is repeated until the accuracy determination unit 304 determines that the accuracy of the level approximation function is good.

In this way, the operation of generating the approximation function for the level is performed.

<< Generation operation of approximation function for physical characteristics >>
Next, the operation of generating the approximation function for the physical property value will be described.

FIG. 6 is a flowchart illustrating the generation operation of the approximation function for the physical property value.

Here, it is assumed that the level data is already stored in the data storage unit 311. For example, it is assumed that the data storage unit 311 stores level data from the first level to the Nmax level.

First, the physical property value estimation device 100 sets “N” indicating the level to N = 1 (S201). Next, the physical property value estimation device 100 acquires the level data belonging to the Nth level among the level data stored in the data storage unit 311 (S202).

Then, the approximation function generation unit 308 for the physical property value generates an approximation function for the Nth level physical property value based on the level data belonging to the Nth level (S203). Specifically, the approximation function generation unit 308 for the physical property value is the Nth level by learning a neural network in which the level data belonging to the Nth level is used as the teacher data, the input is the blending ratio, and the output is the physical property value. Generate an approximation function for the physical property value of (see FIG. 4). Next, the accuracy determination unit 304 determines whether the accuracy of the Nth level approximation function for physical properties generated by the approximation function generation unit 308 for physical characteristics is good (S204).

Then, when the accuracy determination unit 304 determines that the accuracy of the N-level approximation function for physical properties is good (S205), the physical property value estimation device 100 is the Nth Nth generated by the approximation function generation unit 308 for physical properties values. It is decided to adopt the approximate function for the physical property value of the level (S206), and the determined Nth level approximate function for the physical property value is stored in the data storage unit 311. After that, the physical property value estimation device 100 determines whether or not "N" indicating the level is Nmax (S208). At this time, when "N" indicating the level is Nmax, the operation of generating the approximation function for the physical property value is terminated. On the other hand, when "N" indicating the level is not Nmax, N = N + 1 is substituted (S209), and the operation of generating the approximation function for the physical characteristic value of the N + 1th level is performed. When the accuracy determination unit 304 determines that the accuracy of the Nth level approximation function for the physical property value is not good (S205), the physical property value estimation device 100 adds new level data belonging to the Nth level. (S207), the approximation function for the Nth level physical property value is regenerated based on the Nth level level data including the added level data. In this case, since the teacher data increases, it is considered that the accuracy of the regenerated N-level approximation function for the physical property value can be improved. This operation is repeated until the accuracy determination unit 304 determines that the accuracy of the Nth level approximation function for the physical property value is good. In this way, it is possible to generate an approximate function for physical property values from the first level to the Nmax level.

<< Estimating operation of physical property values corresponding to the compounding ratio to be evaluated >>
Next, the operation of estimating the physical property value corresponding to the blending ratio of the evaluation target will be described.

FIG. 7 is a flowchart illustrating an operation of estimating a physical property value corresponding to a blending ratio of an evaluation target. Here, it is assumed that the level approximation function and the respective physical property value approximation functions from the first level to the Nmax level are already stored in the data storage unit 311.

In FIG. 7, first, the blending ratio input unit 305 inputs a blending ratio to be evaluated whose correspondence with the physical property value is unknown (S301). Next, the level estimation unit 306 estimates the level corresponding to the input blending ratio based on the blending ratio input to the blending ratio input unit 305 and the level approximation function stored in the data storage unit 311. (S302).

Subsequently, the physical characteristic value estimation unit 309 estimates the physical property value corresponding to the blending ratio of the evaluation target based on the approximation function for the physical characteristic value at the level estimated by the level estimation unit 308 (S303).

After that, the physical characteristic value output unit 310 outputs the physical characteristic value estimated by the physical characteristic value estimation unit 309 (S304).

In this way, according to the physical property value estimation device 100, it is possible to output a physical property value that is highly likely to be realized for a blending ratio whose correspondence with the physical property value is unknown.

<Characteristics in the embodiment>
Subsequently, the feature points in the present embodiment will be described.

The feature point in this embodiment is that the physical characteristic values corresponding to the blending ratios are divided into a plurality of levels, and approximation functions for the physical characteristic values different from each other are generated for each of the divided plurality of levels. As a result, it is possible to estimate the physical property value with high accuracy for the compounding ratio whose correspondence with the physical property value is unknown. This point will be described below.

For example, it is assumed that the numerical range of the physical property value corresponding to the blending ratio is "0" to "1000". In this case, by training a neural network in which the known data belonging to the numerical range of "0" to "1000" is used as the teacher data, the input is the compounding ratio, and the output is the physical property value, the approximation for a single physical property value is performed. It is conceivable to get a function.

However, since the numerical range of the physical property value is wide, it is difficult to estimate the physical property value with high accuracy in all the numerical range by a single approximation function for the physical property value even by machine learning. In other words, it is difficult to estimate the physical property value with high accuracy over a wide numerical range with a single approximation function for the physical property value, even if machine learning is used.

Therefore, in the present embodiment, the numerical range of the physical characteristic value is divided, and the optimum approximation function for the physical characteristic value is set for each divided narrow numerical range. For example, the above-mentioned numerical range of "0" to "1000" is divided into a first numerical range of "0" to "300" and a second numerical range of "301" to "1000", and the first numerical range is obtained. Is the first level, and the second numerical range is the second level. As a result, in the present embodiment, the level data of the first level in which the numerical range of the physical property value with respect to the blending ratio is the first numerical range and the second level in which the numerical range of the physical property value with respect to the blending ratio is the second numerical range. Level data and is generated.

Then, by learning a neural network in which the level data belonging to the first level is used as the teacher data, the input is the blending ratio, and the output is the physical property value, the approximation function for the physical property value of the first level is acquired. Similarly, by training a neural network in which the level data belonging to the second level is used as the teacher data, the input is the compounding ratio, and the output is the physical property value, the approximation function for the physical property value of the second level is acquired.

In this way, the numerical range (“0” to “300”) of the physical property values included in the level data used for the teacher data is narrowed, so that the approximation function for the physical property values of the first level can be estimated with high accuracy. Can be obtained. Similarly, since the numerical range (“301” to “1000”) of the physical property values included in the level data used for the teacher data becomes narrow, the second-level approximation function for the physical property values that can estimate the physical property values with high accuracy is obtained. It becomes possible to do.

As described above, the feature points in the present embodiment are based on the basic idea that a highly accurate approximation function for physical property values can be obtained by limiting the numerical range of the level data used for the teacher data. In order to embody this basic idea, in this embodiment, the known data is divided into a plurality of levels. Here, the important point is to accurately estimate the level corresponding to the compounding ratio whose correspondence with the physical property value is unknown. Because, if the correspondence with the physical property value is estimated to be different from the level to which the unknown compounding ratio should actually belong, it is not an approximate function for the physical property value used to estimate the physical property value with high accuracy. This is because the approximation function for the physical property value of is used, and as a result, the physical property value cannot be estimated with high accuracy. Therefore, in order to accurately estimate the level corresponding to the unknown compounding ratio with the physical property value, in the present embodiment, the level data relating the known data and the level is generated. Then, by using this level data as teacher data and learning a neural network in which the input is the compounding ratio and the output is the level, the approximation function for the level is acquired. In this way, by using the level approximation function, it is possible to accurately estimate the level corresponding to the compounding ratio whose correspondence with the physical property value is unknown.

That is, in the present embodiment, by using the level approximation function for estimating the level and the physical property value approximation function generated for each level, the physical properties corresponding to the compounding ratio whose correspondence with the physical property values is unknown. The value can be estimated with high accuracy.

<Modification example>
In the embodiment, as shown in FIG. 2, an example in which the physical characteristic value estimation system for estimating the physical property value according to the blending ratio is composed of a single physical property value estimation device 100 has been described, but the physical property value estimation system is described. Not limited to this configuration, for example, it can also be configured from a distributed system.

FIG. 8 is a functional block diagram showing an example in which the physical characteristic value estimation system is configured by the physical characteristic value estimation device and the approximate function generation device.

As shown in FIG. 8, the physical characteristic value estimation system is composed of a physical property value estimation device 400 and an approximate function generation device 500. For example, the physical property value estimation device 400 and the approximate function generation device 500 are connected by a network 600. ing.

The physical property value estimation device 400 includes a blending ratio input unit 305, a level estimation unit 306, a level output unit 307, a physical property value estimation unit 309, a physical property value output unit 310, a data storage unit 311 and a communication unit 320. Have.

The approximation function generation device 500 includes a known data input unit 301, a level data generation unit 302, a level approximation function generation unit 303, an accuracy determination unit 304, a physical property value approximation function generation unit 308, and a data storage unit 312. And a communication unit 330.

The physical characteristic value estimation device 400 and the approximate function generation device 500 configured in this way are configured so that data can be transmitted and received by the communication unit 320 and the communication unit 330 via the network 600. Then, in the approximation function generation device 500, the above-mentioned "generation operation of the approximation function for the level" and "generation operation of the approximation function for the physical property value" are performed, and the approximation function for the level and the approximation function for the physical property value for each level are performed. Is generated.

On the other hand, in the physical property value estimation device 400, the level approximation function generated by the approximation function generation device 500 and the physical property value approximation function for each level are input from the approximation function generation device 500 and stored in the data storage unit 311.

After that, in the physical property value estimation device 400, based on the level approximation function stored in the data storage unit 311 and the physical property value approximation function for each level, the above-mentioned "estimation of the physical property value corresponding to the blending ratio of the evaluation target" is performed. "Action" is performed.

In this way, the physical characteristic value estimation system according to the present embodiment can also be constructed by the distributed system including the physical characteristic value estimation device 400 and the approximate function generation device 500.

<Verification of effect>
Next, according to the physical property value estimation system in the present embodiment, the verification result capable of estimating the physical property value corresponding to the compounding ratio of the evaluation target with high accuracy will be described.

FIG. 9 shows a physical property value estimated by a single approximation function for a physical property value obtained by training a neural network using known data included in the numerical range “0” to “1000” of the physical property value as training data. It is a graph which shows the correspondence relationship with the measured value.

In FIG. 9, the horizontal axis shows the measured value, while the vertical axis shows the estimated value.

The closer the plot data is to the broken line shown in FIG. 9, the closer the estimated value is to the measured value. Therefore, the smaller the variation of the plot data from the straight line of the broken line, the higher the accuracy of the estimation. In this regard, it can be seen that the plot data shown in FIG. 9 has a large variation from the straight line of the broken line. This means that the estimation accuracy of the physical property value estimated by the approximation function for a single physical property value is low.

Subsequently, in FIG. 10, the numerical range “0” to “1000” of the physical property value is divided into the first level of the numerical range “0” to “300” and the second level of the numerical range “301” to “1000”. It is a graph showing the correspondence between the physical property value estimated by the approximate function for the physical property value for each level and the measured value.

In FIG. 10, the gray plot data show the physical property values estimated using the first level approximation function for the physical property values. On the other hand, the black plot data shows the physical property values estimated using the second-level approximation function for the physical property values.

For reference, FIG. 11 is an approximation for the first level physical property value obtained by training the neural network using the level data included in the numerical range “0” to “300” of the physical property value as the teacher data. It is a graph which shows by extracting the plot data (gray plot data) of the physical property value estimated by a function from the plot data shown in FIG.

On the other hand, FIG. 12 is estimated by a second-level approximation function for physical property values obtained by training a neural network using level data included in the numerical range “301” to “1000” of physical property values as teacher data. It is a graph which shows by extracting the plot data (black plot data) of the physical property value from the plot data shown in FIG.

It can be seen that the plot data shown in FIGS. 10 to 12 have a small variation from the straight line of the broken line. This means that the estimation accuracy of the physical characteristic value can be improved by dividing the numerical range of the physical characteristic value into a plurality of levels and using the approximation function for the physical characteristic value generated for each level. That is, it is supported that the physical property value estimation system in the present embodiment can estimate the physical property value with high accuracy.

<Application example>
In the embodiment, the physical characteristic value estimation system that estimates the physical characteristic value corresponding to the compounding ratio to be evaluated has been described. For example, the level estimation device that estimates the level to which the physical property value belongs instead of estimating the physical characteristic value. Is also useful.

Therefore, in this application example, a level estimation device that outputs the level to which the physical property value corresponding to the compounding ratio belongs when the compounding ratio to be evaluated is input will be described.

FIG. 13 is a diagram showing a functional block configuration of the level estimation device.

In FIG. 13, the level estimation device 700 includes a known data input unit 301, a level data generation unit 302, a level approximation function generation unit 303, an accuracy determination unit 304, a blending ratio input unit 305, and a level estimation unit 306. And has a level output unit 307.

The level estimation device 700 configured in this way generates a level approximation function by the same “level approximation function generation operation” as the physical property value estimation device 100 described in the embodiment (see FIG. 5). .. Then, as shown below, the level estimation device 700 estimates the level corresponding to the blending ratio of the evaluation target based on the level approximation function.

FIG. 14 is a flowchart illustrating an estimation operation of the level to which the physical characteristic value corresponding to the blending ratio of the evaluation target belongs. In FIG. 14, first, the blending ratio input unit 305 inputs a blending ratio to be evaluated whose correspondence with the physical property value is unknown (S401). Next, the level estimation unit 306 estimates the level corresponding to the input blending ratio based on the blending ratio input to the blending ratio input unit 305 and the level approximation function stored in the data storage unit 311. (S402). After that, the level output unit 307 outputs the level estimated by the level estimation unit 306. In this way, according to the level estimation device 700, it is possible to output a level corresponding to a blending ratio whose correspondence with the physical property value is unknown.

For example, the level estimation device 700 in this application example is useful in the following cases. That is, the physical property value corresponding to the blending ratio is divided into a first level and a second level. At this time, the physical property value belonging to the first level is set to a low value not worthy of experimental evaluation, while the physical property value belonging to the second level is set to a high value worthy of experimental evaluation.

Based on such a premise, in the level estimation device 700 in this application example, when the "first blending ratio" which is the evaluation target whose correspondence with the physical property value is unknown is input, the "first level" is output. Suppose. On the other hand, in the level estimation device 700 in this application example, when the "second blending ratio" whose correspondence with the physical property value is unknown is input, the "second level" is output. As a result, the user of the level estimation device 700 can determine that the "first compounding ratio" is not worthy of the experimental evaluation, while the "second compounding ratio" is worthy of the experimental evaluation. can. In this way, even in the level estimation device 700 configured to estimate the level to which the physical property value belongs instead of estimating the physical property value itself, the compounding ratio whose correspondence with the physical property value is unknown is suitable for experimental evaluation. It turns out to be useful in that it can be judged before the experiment whether or not it is.

Although the invention made by the present inventor has been specifically described above based on the embodiment thereof, the present invention is not limited to the embodiment and can be variously modified without departing from the gist thereof. Needless to say.

The embodiment includes the following embodiments.

(Appendix 1)
A known data input unit configured to input known data in which a blending ratio of a material composed of a plurality of types of compositions and a physical property value corresponding to the blending ratio are related to each other.
A level configured to generate level data in which the blending ratio, the physical property value, and the level are related by adding the level to which the physical property value belongs from a plurality of preset levels to the known data. Data generator and
Based on the level data, it is configured to generate a level approximation function that outputs the level to which the physical property value corresponding to the first compounding ratio belongs when the first compounding ratio whose correspondence with the physical property value is unknown is input. Approximate function generator for level and
A blending ratio input unit configured to input the first blending ratio, and
A level estimation unit configured to estimate the level corresponding to the first compounding ratio based on the first compounding ratio and the approximate function for the level.
In the physical characteristic value estimation device which is a component of the physical characteristic value estimation system including
The physical property value estimation device is at least
A storage unit configured to store the level approximation function,
The compounding ratio input unit and
With the level estimation unit
A physical property value estimation device.

(Appendix 2)
A known data input unit configured to input known data in which a blending ratio of a material composed of a plurality of types of compositions and a physical property value corresponding to the blending ratio are related to each other.
A level configured to generate level data in which the blending ratio, the physical property value, and the level are related by adding the level to which the physical property value belongs from a plurality of preset levels to the known data. Data generator and
Based on the level data, it is configured to generate a level approximation function that outputs the level to which the physical property value corresponding to the first compounding ratio belongs when the first compounding ratio whose correspondence with the physical property value is unknown is input. Approximate function generator for level and
A blending ratio input unit configured to input the first blending ratio, and
A level estimation unit configured to estimate the level corresponding to the first compounding ratio based on the first compounding ratio and the approximate function for the level.
In the approximate function generator, which is a component of the physical characteristic value estimation system,
The approximate function generator has at least an approximate function generator for the level.
The level approximation function generation unit is configured to generate the level approximation function by training a neural network using the level data as teacher data, input as a blending ratio, and output as a level. Approximate function generator.

100 Physical property value estimation device 101 CPU
102 ROM
103 RAM
104 Display 105 Keyboard 106 Mouse 107 Communication board 108 Removable disk device 109 CD / DVD-ROM device 110 Printer 111 Scanner 112 Hard disk device 113 Bus 201 Operating system 202 Program group 203 File group 301 Known data input unit 302 Level Data generation unit 303 level Approximate function generation unit 304 Accuracy judgment unit 305 Mixing ratio input unit 306 Level estimation unit 307 Level output unit 308 Physical property value approximation function generation unit 309 Physical property value estimation unit 310 Physical property value output unit 311 Data storage unit 312 Data storage unit 320 Communication Unit 330 Communication unit 400 Physical property value estimation device 500 Approximate function generator 600 Network 700 Level estimation device

Claims (7)

A known data input unit configured to input known data in which a blending ratio of a material composed of a plurality of types of compositions and a physical property value corresponding to the blending ratio are related to each other.
A level configured to generate level data in which the blending ratio, the physical property value, and the level are related by adding the level to which the physical property value belongs from a plurality of preset levels to the known data. Data generator and
Based on the level data, it is configured to generate a level approximation function that outputs the level to which the physical property value corresponding to the first compounding ratio belongs when the first compounding ratio whose correspondence with the physical property value is unknown is input. Approximate function generator for level and
A blending ratio input unit configured to input the first blending ratio, and
A level estimation unit configured to estimate the level corresponding to the first compounding ratio based on the first compounding ratio and the approximate function for the level.
A physical property value estimation system. In the physical property value estimation system according to claim 1,
The level approximation function generation unit is configured to generate the level approximation function by training a neural network using the level data as teacher data, input as a blending ratio, and output as a level. There is a physical property value estimation system. In the physical property value estimation system according to claim 1,
The physical property value estimation system in which the plurality of levels are set based on a threshold value. In the physical property value estimation system according to claim 1,
The physical property value estimation system is
Based on the level data belonging to each of the plurality of levels, when the first compounding ratio is input, an approximation function for physical characteristics that outputs the physical property value corresponding to the first compounding ratio is generated for each of the plurality of levels. Approximate function generator for physical characteristic values configured to
A physical property value estimation unit configured to estimate a physical characteristic value corresponding to the first compounding ratio based on an approximation function for a physical characteristic value at a level estimated by the level estimation unit.
A physical characteristic value output unit configured to output the physical characteristic value estimated by the physical characteristic value estimation unit, and a physical characteristic value output unit.
A physical property value estimation system. In the physical property value estimation system according to claim 4,
The approximation function generation unit for physical property values uses level data belonging to each of the plurality of levels as teacher data, and trains a neural network in which an input is a blending ratio and an output is a physical property value, thereby causing the plurality of levels. A physical property value estimation system configured to generate an approximation function for the physical property value for each of the above. A known data input step of inputting known data relating a blending ratio of a material composed of a plurality of types of compositions and a physical property value corresponding to the blending ratio, and
A level data generation step of adding a level to which the physical property value belongs from a plurality of preset levels to the known data and generating level data in which the compounding ratio, the physical property value, and the level are related to each other.
Based on the level data, when the first compounding ratio whose correspondence with the physical property value is unknown is input, the level approximation function for generating the level to which the physical property value corresponding to the first compounding ratio belongs is generated. Process and
In the compounding ratio input step of inputting the first compounding ratio,
A level estimation step for estimating the level corresponding to the first blending ratio based on the first blending ratio and the approximate function for the level, and a level estimation step.
A method for estimating physical characteristics. In the method for estimating the physical property value according to claim 6,
The method for estimating the physical property value is
Based on the level data belonging to each of the plurality of levels, when the first compounding ratio is input, an approximation function for physical characteristics that outputs the physical property value corresponding to the first compounding ratio is generated for each of the plurality of levels. Approximate function generation process for physical characteristics and
A physical property value estimation step for estimating a physical property value corresponding to the first compounding ratio based on an approximation function for a physical characteristic value at a level estimated in the level estimation step, and a physical property value estimation step.
A physical property value output process that outputs the physical property value estimated in the physical property value estimation process, and a physical property value output process.
A method for estimating physical characteristics.

Images