Learning device, operating method of learning device, operating procedure of learning device and application device

Abstract

This invention provides a learning apparatus, a method for operating the learning apparatus, an operating procedure for the learning apparatus, and an application apparatus for further improving the accuracy of product quality predictions made by a machine learning model when multidimensional physical property data derived from multidimensional physical property data of a product is provided as input data for learning. The learning unit of the learning apparatus provides input data, including related data with multiple items, derived from the physical property data of the product, to a first machine learning model for learning and outputs a temporary machine learning model. The extraction unit uses the temporary machine learning model to extract high-contribution items from the multiple items of the related data, ensuring that the degree of contribution to improving the accuracy of quality prediction meets predetermined conditions. The learning unit selectively provides the related data of the high-contribution items to a second machine learning model for learning and outputs the second machine learning model as a learned model for practical application.

Description

Technical Field

[0001] The technology disclosed herein relates to a learning device, a method of operating the learning device, a procedure for operating the learning device, and an application device. Background Technology

[0002] Machine learning models have been used to predict product quality. Patent document 1 proposes a machine learning model that uses physical property-related data derived from physical property data representing the product's physical properties as input data to improve prediction accuracy.

[0003] Patent Document 1 exemplifies coffee beans and other food and beverage products as products. Furthermore, Patent Document 1 exemplifies spectral data such as near-infrared (NIR) spectral analysis data, Fourier transform infrared (FT-IR) spectral analysis data, and nuclear magnetic resonance (NMR) spectral analysis data, or image data obtained by photographing the product with a camera, etc., as physical property data. Moreover, as physical property-related data, numerical values ​​obtained from spectral data can be cited, such as the waveform slope, periodicity, amplitude, peak height, and peak width of the spectrum. Furthermore, as physical property-related data, image data obtained by photographing the product with a camera, etc., can be cited as examples.

[0004] Previous technical documents

[0005] Patent documents

[0006] Patent Document 1: Japanese Patent Application Publication No. 2018-018354 Summary of the Invention

[0007] The technical problem to be solved by the invention

[0008] Physical property data can be categorized into two types: data where a single parameter represents a product's physical property, such as its weight, and data where multiple parameters represent a product's physical property, such as spectral data and image data. More specifically, in spectral data, the spectrum represents the product's physical property; for example, wavenumber and intensity are equivalent to multiple parameters. In image data, color represents the product's physical property; the pixel values ​​for red, green, and blue are equivalent to multiple parameters. When parameters are used as dimensions for data capture, such physical property data can be considered multidimensional. Conversely, physical property data where a single parameter represents a product's physical property can be considered one-dimensional. Hereinafter, physical property data where multiple parameters represent a product's physical property will be referred to as multidimensional physical property data. Furthermore, the physical property-related data derived from multidimensional physical property data will be called multidimensional physical property-related data.

[0009] Multidimensional physical property-related data consists of numerous numerical and graphical data points, resulting in a large number of items. Among these items, some data significantly contribute to improving the prediction accuracy of machine learning models, while others contribute almost nothing. Therefore, directly feeding multidimensional physical property-related data to machine learning models for learning yields limited results, and the prediction accuracy of the machine learning model may peak at a relatively low level.

[0010] The purpose of this disclosure is to provide a learning device, a method for operating the learning device, an operating procedure for the learning device, and an application device that can further improve the accuracy of product quality prediction by the machine learning model when multidimensional physical property data derived from multidimensional physical property data of a product is provided as input data for learning to a machine learning model.

[0011] means for solving technical problems

[0012] To achieve the above objectives, the learning apparatus of this disclosure comprises: a first acquisition unit that acquires learning input data, which is provided to a machine learning model for predicting the quality of a product, and includes multidimensional property-related data having multiple items derived from multidimensional property data representing the physical properties of the product; a temporary learning unit that provides the learning input data to the machine learning model for learning and outputs a temporary machine learning model; an extraction unit that uses the temporary machine learning model to extract high-contribution items from the multiple items of the multidimensional property-related data, such that the degree of contribution to improving the accuracy of quality prediction meets predetermined conditions; and a formal learning unit that selectively provides the multidimensional property-related data of high-contribution items to the machine learning model for learning and outputs the machine learning model as a learned model for practical application.

[0013] The preferred learning input data also includes production condition data set in the product's production process.

[0014] The preferred multidimensional physical property data includes spectral data obtained from spectral analysis of the product.

[0015] The preferred multidimensional physical property correlation data are representative values ​​of intensity derived for each of the multiple intervals of the spectral data.

[0016] The preferred multidimensional physical property data includes image data obtained from photographing the product.

[0017] The preferred products are produced using flow synthesis.

[0018] Preferably, it has an export section that applies at least a portion of an autoencoder to the multidimensional physical property data to export multidimensional physical property-related data.

[0019] The preferred automatic encoder learns from multidimensional property data of products of higher quality than a preset grade. The output unit provides multidimensional property data to the automatic encoder and outputs output data. Multidimensional property related data is derived based on the difference between the multidimensional property data provided to the automatic encoder and the output data.

[0020] The preferred output unit provides multidimensional physical property data to the autoencoder, outputs feature data from the encoder network of the autoencoder, and derives multidimensional physical property related data based on the feature data.

[0021] Preferred multidimensional physical property data includes image data of the spectrum represented by spectral data detected by spectral analysis of the product.

[0022] The optimal export section exports multidimensional physical property correlation data for each of the multiple intervals of the divided spectral data.

[0023] The application device disclosed herein includes: a second acquisition unit for acquiring a learned model output from the formal learning unit of a learning device; a third acquisition unit for acquiring multidimensional physical property correlation data for predicting products with unknown quality; a processing unit for providing the learned model acquired by the second acquisition unit with the multidimensional physical property correlation data for predicting products with unknown quality acquired by the third acquisition unit to predict quality; and an output control unit for controlling the output of the quality prediction result obtained by the learned model.

[0024] The operating method of the learning device disclosed herein includes: a first acquisition step, acquiring learning input data, which is provided to a machine learning model for predicting the quality of a product, and includes multidimensional physical property related data with multiple items derived from multidimensional physical property data representing the physical properties of the product; a temporary learning step, providing the learning input data to the machine learning model for learning, and outputting a temporary machine learning model; an extraction step, using the temporary machine learning model to extract high-contribution items from the multiple items of the multidimensional physical property related data, the degree of which helps to improve the accuracy of quality prediction meeting preset conditions; and a formal learning step, selectively providing the multidimensional physical property related data of high-contribution items to the machine learning model for learning, and outputting the machine learning model as a learned model for practical application.

[0025] The operating procedure of the learning device disclosed herein enables a computer to function as a first acquisition unit, a temporary learning unit, an extraction unit, and a formal learning unit. The first acquisition unit acquires learning input data, which is provided to a machine learning model for predicting product quality and includes multidimensional property-related data with multiple items derived from multidimensional property data representing the product's physical properties. The temporary learning unit provides the learning input data to the machine learning model for learning and outputs a temporary machine learning model. The extraction unit uses the temporary machine learning model to extract high-contribution items from the multiple items of the multidimensional property-related data, ensuring that the degree of contribution to improving the accuracy of quality prediction meets pre-set conditions. The formal learning unit selectively provides the multidimensional property-related data of high-contribution items to the machine learning model for learning and outputs the machine learning model as a learned model for practical application.

[0026] Invention Effects

[0027] According to the technology disclosed herein, a learning device, a method for operating the learning device, an operating procedure for the learning device, and an application device are provided that can further improve the prediction accuracy of product quality by the machine learning model when multidimensional physical property related data derived from multidimensional physical property data of a product are provided as input data for learning to a machine learning model. Attached Figure Description

[0028] Figure 1 It is a diagram representing a machine learning system, a flow reaction device, a physical property analysis device, and a quality evaluation device.

[0029] Figure 2 It is a diagram showing the outline of the processes in the flow reaction device, the physical property analysis device, and the quality evaluation device.

[0030] Figure 3 It is a diagram that shows an overview of the processing in the learning device and the application device.

[0031] Figure 4 It is a graph that represents production condition data, physical property data, related data, and quality data linked together by a common ID.

[0032] Figure 5 This is a diagram showing a flow reaction apparatus having a reaction section containing a T-shaped confluence.

[0033] Figure 6 This is a diagram showing a flow reaction apparatus having a reaction section containing a cross-shaped confluence.

[0034] Figure 7 It is a graph representing production condition data.

[0035] Figure 8 It is a graph representing spectral data and spectra.

[0036] Figure 9 It is a graph representing quality data.

[0037] Figure 10 This is a block diagram representing a computer that constitutes a learning device and an application device.

[0038] Figure 11 This is a block diagram representing the processing unit of the CPU in a learning device.

[0039] Figure 12 It is a graph that represents the division of spectral data into multiple intervals.

[0040] Figure 13 This diagram illustrates the scenario where the first derivation unit derives relevant data for each of the multiple intervals of the divided spectral data.

[0041] Figure 14 This is a diagram showing the details of the Study Department.

[0042] Figure 15 This is a diagram representing the processing of the first processing unit in temporary learning.

[0043] Figure 16 This is a diagram showing the details of the extraction section.

[0044] Figure 17 This is a diagram showing the processing of the second processing unit.

[0045] Figure 18 This is a diagram showing the processing of the third processing unit.

[0046] Figure 19 This is a diagram representing the processing of the computing unit.

[0047] Figure 20 This is a diagram illustrating a specific example of the processing by the computing unit.

[0048] Figure 21 It is a graph representing contribution information.

[0049] Figure 22 This is a diagram showing the processing of the decision-making unit.

[0050] Figure 23 This is a diagram representing the processing of the first processing unit in formal learning.

[0051] Figure 24 This is a block diagram representing the processing unit of the CPU in the device.

[0052] Figure 25 This is a diagram showing the processing of the 4th processing unit.

[0053] Figure 26 This is a diagram representing the quality prediction display screen.

[0054] Figure 27 This is a flowchart showing the processing sequence of the learning device.

[0055] Figure 28 This is a flowchart representing the processing order of temporary learning.

[0056] Figure 29 This is a flowchart showing the processing order for extracting high-contribution items.

[0057] Figure 30 This is a flowchart representing the processing sequence of formal learning.

[0058] Figure 31 It is a flowchart showing the processing sequence of the device.

[0059] Figure 32 This is a table showing the coefficients of determination for the molecular weight dispersion and the molecular weight of the comparative examples and embodiments.

[0060] Figure 33 It is a graph representing image data.

[0061] Figure 34 It is a graph that represents multiple regions of image data.

[0062] Figure 35 This diagram illustrates the situation where the first export unit exports relevant data for each of the multiple regions that are divided into image data regions.

[0063] Figure 36 This is a diagram illustrating the application of an autoencoder to physical property data to derive related data.

[0064] Figure 37 This is a diagram representing an automatic encoder.

[0065] Figure 38 This is an illustration of convolution processing.

[0066] Figure 39 It is a graph representing the feature map of an image.

[0067] Figure 40 This is a diagram showing the preparation of an autoencoder for each of the multiple intervals of spectral data.

[0068] Figure 41 This diagram illustrates a scenario where spectral image data of a product of higher quality than a pre-defined grade is provided to an automatic encoder as input image data for learning.

[0069] Figure 42This is a diagram illustrating the difference between the input and output image data of an autoencoder.

[0070] Figure 43 This diagram illustrates the case where the average and sum of the difference data are derived as relevant data for each of the multiple intervals of the divided spectral data through the first derivation unit.

[0071] Figure 44 This is a diagram illustrating the situation where relevant data is derived from image feature maps. Detailed Implementation

[0072] [First Implementation]

[0073] Figure 1 In this system, the machine learning system 2 includes a learning device 10 and an application device 11. The learning device 10 and the application device 11 are, for example, desktop personal computers. The learning device 10 and the application device 11 are communicatively connected to each other via a network 12. The network 12 is, for example, a LAN (Local Area Network) or a WAN (Wide Area Network), such as the Internet or a public communication network. The network 12 also connects to a flow reaction device 13, a physical property analysis device 14, and a quality evaluation device 15.

[0074] Figure 2 In this process, the flow reaction apparatus 13 produces product PR from raw material RM using a flow synthesis method, according to the production condition data PCD in the production process. The property analysis apparatus 14 analyzes the properties of product PR and outputs the analysis results, i.e., property data PD. The quality evaluation apparatus 15 evaluates the quality of product PR and outputs the evaluation results, i.e., quality data QD. The flow reaction apparatus 13 sends the production condition data PCD to the learning apparatus 10, the property analysis apparatus 14 sends the property data PD to the learning apparatus 10, and the quality evaluation apparatus 15 sends the quality data QD to the learning apparatus 10.

[0075] Figure 3 In this process, the learning device 10 acquires production condition data (PCD) from the flow reaction unit 13, physical property data (PD) from the physical property analysis unit 14, and quality data (QD) from the quality evaluation unit 15. The learning device 10 derives physical property related data (hereinafter referred to as related data) (PRD) from the physical property data PD. This related data PRD and the production condition data PCD constitute the learning input data IDL. For example... Figure 4 As shown, production condition data (PCD), physical property data (PD), related data (PRD) derived from physical property data (PD), and quality data (QD) are associated with a common ID (identification data) assigned to a product PR.

[0076] The learning device 10 has a machine learning model M. The machine learning model M is used to predict the quality of product PR. To improve the prediction accuracy of this machine learning model M, the learning device 10 provides the machine learning model M with learning input data IDL, consisting of production condition data PCD and related data PRD, both assigned the same ID. Figure 15 (etc.). The output of the machine learning model M is compared with the corresponding learning output data ODL to the input data IDL.

[0077] Quality data QD is used to respond to the learning output data ODL. The higher the prediction accuracy of the machine learning model M, the smaller the difference between the quality data QD and the learning output data OD. Therefore, the learning device 10 compares the learning output data ODL with the quality data QD, which has the same ID as the learning input data IDL, to evaluate the prediction accuracy of the machine learning model M. Then, based on the evaluation result, the machine learning model M is updated. The learning device 10 performs evaluation of the prediction accuracy of the machine learning model M, including the input of the learning input data IDL to the machine learning model M and the output of the learning output data ODL from the machine learning model M, while changing the learning input data IDL and the quality data QD. This series of processes is repeated until the prediction accuracy of the machine learning model M reaches a preset level. The learning device 10 sends the machine learning model M with the prediction accuracy reaching the preset level as a learned model TM for practical application to the application device 11.

[0078] The application device 11 receives the learned model TM from the learning device 10. The application device 11 provides the learned model TM with the production condition data (PCDF) for the product PR with unknown quality, i.e., the predictive production condition data, and the related data (PRDF) for the product PR with unknown quality, i.e., the predictive physical property related data (hereinafter referred to as the predictive related data). Similar to the related data PRD, the predictive related data PRDF is derived from the physical property data (PD) for the product PR with unknown quality, i.e., the predictive physical property data PDF (see reference). Figure 24 The data is exported. After learning, the Model™ outputs Quality Forecast Data (QFD) corresponding to these Predictive Production Condition Data (PCDF) and Predictive Relevance Data (PRDF).

[0079] The flow reaction carried out in the flow reaction apparatus 13 is a synthesis reaction of monomers or a polymerization reaction in which monomers react with each other to form a polymer. Therefore, the product PR can also be, for example, a compound in the growth stage that is the target of the polymerization reaction. In this example, the flow reaction apparatus 13 carries out anionic polymerization of polystyrene, which is the product PR, through a flow reaction.

[0080] Figure 5 In the process, the flow reaction device 13 includes a first raw material supply unit 20, a second raw material supply unit 21, a reaction zone 22, a temperature control unit 23, a waste recovery section 24, a setting unit 25, and a system controller 26.

[0081] The first raw material supply unit 20 is connected to the upstream end of the reaction section 22 via a pipe (not shown). The first raw material supply unit 20 supplies the first raw material RM1 to the reaction section 22. The first raw material supply unit 20 has a pump for conveying the first raw material RM1 to the reaction section 22. The flow rate of the first raw material RM1 conveyed from the first raw material supply unit 20 to the reaction section 22 is adjusted by controlling the speed of the pump.

[0082] In this example, the first raw material RM1 is a solution of lithium polystyrene dissolved in a solvent. Lithium polystyrene acts as an initiator for the anionic polymerization of product PR, i.e., polystyrene. Tetrahydrofuran is used as the solvent. Furthermore, a small amount of toluene and hexane are mixed in the solution. Alternatively, the raw materials for the flow reaction can be a mixture of reactants such as lithium polystyrene (RM1) with other substances, or they can consist solely of reactants.

[0083] The second raw material supply unit 21 is identical to the first raw material supply unit 20, and is connected to the upstream end of the reaction section 22 via a pipe (not shown). The second raw material supply unit 21 supplies the second raw material RM2 to the reaction section 22. The second raw material supply unit 21, also identical to the first raw material supply unit 20, has a pump for conveying the second raw material RM2 to the reaction section 22. The flow rate of the second raw material RM2 conveyed from the second raw material supply unit 21 to the reaction section 22 is adjusted by controlling the rotational speed of the pump.

[0084] In this example, the second raw material RM2 is an aqueous methanol solution. Methanol is used as a stop agent for the anionic polymerization reaction.

[0085] Reaction section 22 is the section used for the flow reaction (anionic polymerization in this example). Reaction section 22 has a confluence section 30 and a reaction section 31. The confluence section 30 is composed of a first tube section 32, a second tube section 33, and a third tube section 34. The first tube section 32 and the second tube section 33 are connected in a straight line, and the third tube section 34 intersects the first tube section 32 and the second tube section 33 perpendicularly. That is, the confluence section 30 is T-shaped.

[0086] The first pipe section 32 is connected to the first raw material supply section 20, and the second pipe section 33 is connected to the second raw material supply section 21. Furthermore, the third pipe section 34 is connected to the reaction section 31. The first raw material RM1 is supplied from the first raw material supply section 20 to the first pipe section 32, and the second raw material RM2 is supplied from the second raw material supply section 21 to the second pipe section 33. The first raw material RM1 and the second raw material RM2 are combined in the third pipe section 34 and transported to the reaction section 31 in a mixed state.

[0087] A first flow rate sensor 35 is provided in the first pipe section 32 to detect the flow rate of the first raw material RM1 passing through the first pipe section 32. A second flow rate sensor 36 is provided in the second pipe section 33 to detect the flow rate of the second raw material RM2 passing through the second pipe section 33. Furthermore, a third flow rate sensor 37 is provided in the third pipe section 34 to detect the flow rate of the mixture of the first raw material RM1 and the second raw material RM2 passing through the third pipe section 34.

[0088] The reaction section 31 is a slender tube formed by connecting multiple straight tubes of the same inner diameter in a straight line. The length L of the reaction section 31 can be changed by changing the number of connected straight tubes and / or the length of the straight tubes. Furthermore, the inner diameter Φ of the reaction section 31 can be changed by changing the inner diameter of the connected straight tubes.

[0089] The interior of the reaction section 31 is the flow path for the mixture of raw material RM1 and raw material RM2, and is the site of the flow reaction. The mixture passes through the reaction section 31, where it undergoes a flow reaction to become a polystyrene solution. Some flow reaction also occurs in the third pipe section 34 of the confluence section 30, but the length of the third pipe section 34 is very short relative to the length L of the reaction section 31. Therefore, the third pipe section 34 is ignored, and the length L of the reaction section 31 is considered as the length during the flow reaction, i.e., the reaction path length. Similarly, the inner diameter Φ of the reaction section 31 is considered as the diameter during the flow reaction, i.e., the reaction path.

[0090] The temperature regulating unit 23 includes a heater and / or a cooler to regulate the internal temperature of the reaction unit 31 (hereinafter referred to as the reaction temperature). A temperature sensor 38 for detecting the reaction temperature is provided at the downstream end of the reaction unit 31.

[0091] Waste recycling section 24 is the section where the product PR (polystyrene) is recycled and the waste from failed reactions is disposed of. Waste recycling section 24 has a recycling section 40 and a waste section 41. The recycling section 40 and the waste section 41 are connected to the downstream end of the reaction section 31 via a three-way valve 42. Through this three-way valve 42, the recycling line connecting the reaction section 31 and the recycling section 40, and the waste line connecting the reaction section 31 and the waste section 41, can be switched.

[0092] The recovery unit 40 precipitates polystyrene from the polystyrene solution. The recovery unit 40 extracts the precipitated polystyrene by filtering it from the solution. Then, the extracted polystyrene is dried. More specifically, the recovery unit 40 has a container equipped with a stirrer, in which methanol is contained, and the polystyrene solution is mixed into the stirred methanol, thereby precipitating polystyrene. Furthermore, the recovery unit 40 has a thermostatic bath with a reduced pressure function, and the interior of the thermostatic bath is heated under reduced pressure to dry the methanol.

[0093] Waste section 41 is a tank for storing waste. Here, waste refers to waste transported from reaction section 31 when the flow rate of raw material RM1, the flow rate of raw material RM2, the flow rate of the mixture, the reaction temperature, etc., are disrupted for some reason, making it impossible to produce according to the originally planned production conditions.

[0094] The setting unit 25 receives the production conditions settings for the product PR production process from the operator of the flow reaction device 13. The production conditions received by the setting unit 25 are registered as Production Condition Data (PCD) set in the product PR production process to the system controller 26.

[0095] The system controller 26 centrally controls the overall operation of the flow reaction device 13. The system controller 26 is connected to the first raw material supply unit 20, the second raw material supply unit 21, the temperature regulation unit 23, the first flow rate sensor 35, the second flow rate sensor 36, the third flow rate sensor 37, the temperature sensor 38, and the three-way valve 42.

[0096] The system controller 26 controls the rotational speed of the pump in the first raw material supply unit 20 and adjusts the flow rate of the first raw material RM1 based on the flow rate of the first raw material RM1 detected by the first flow rate sensor 35. Similarly, the system controller 26 controls the rotational speed of the pump in the second raw material supply unit 21 and adjusts the flow rate of the second raw material RM2 based on the flow rate of the second raw material RM2 detected by the second flow rate sensor 36. Furthermore, the system controller 26 drives the temperature regulating unit 23 based on the reaction temperature detected by the temperature sensor 38. In addition, the system controller 26 controls the three-way valve 42 to switch between the aforementioned recovery line and waste line.

[0097] It can also be used Figure 6 The reaction section 45 shown replaces the reaction section 22. Additionally, Figure 6 In the middle, to and Figure 5 The same components are marked with the same symbols, and the explanation is omitted.

[0098] Figure 6The confluence section 46 of the reaction section 45 shown is composed of a first tube section 47, a second tube section 48, a third tube section 49, and a fourth tube section 50. The first tube section 47 and the second tube section 48 are connected in a straight line. Similarly, the third tube section 49 and the fourth tube section 50 are connected in a straight line. Furthermore, the first tube section 47 and the third tube section 49, and the second tube section 48 and the fourth tube section 50, intersect perpendicularly. That is, the confluence section 46 is cross-shaped.

[0099] The first pipe section 47 and the second pipe section 48 are connected to the first raw material supply section 20, and the third pipe section 49 is connected to the second raw material supply section 21. Furthermore, the fourth pipe section 50 is connected to the reaction section 31. The first raw material RM1 is supplied from the first raw material supply section 20 to the first pipe section 47 and the second pipe section 48, respectively, and the second raw material RM2 is supplied from the second raw material supply section 21 to the third pipe section 49. The first raw material RM1 and the second raw material RM2 are combined in the fourth pipe section 50 and transported to the reaction section 31 in a mixed state.

[0100] A first flow rate sensor 51 is provided in the first pipe section 47 to detect the flow rate of the first raw material RM1 passing through the first pipe section 47, and a second flow rate sensor 52 is provided in the second pipe section 48 to detect the flow rate of the first raw material RM1 passing through the second pipe section 48. A third flow rate sensor 53 is provided in the third pipe section 49 to detect the flow rate of the second raw material RM2 passing through the third pipe section 49. A fourth flow rate sensor 54 is provided in the fourth pipe section 50 to detect the flow rate of the mixture of the first raw material RM1 and the second raw material RM2 passing through the fourth pipe section 50.

[0101] In this configuration, the system controller 26 controls the rotational speed of the pump in the first raw material supply unit 20 and adjusts the flow rate of the first raw material RM1 based on the average value of the flow rate of the first raw material RM1 detected by the first flow rate sensor 51 and the flow rate of the first raw material RM1 detected by the second flow rate sensor 52. Furthermore, the system controller 26 controls the rotational speed of the pump in the second raw material supply unit 21 and adjusts the flow rate of the second raw material RM2 based on the flow rate of the second raw material RM2 detected by the third flow rate sensor 53.

[0102] Figure 7 The production condition data (PCD) includes the concentration (mol / L) and flow rate (ml / min) of the first feedstock RM1, the concentration (mol / L) and flow rate (ml / min) of the second feedstock RM2, the shape of the confluence, the reaction path Φ (mm), the reaction path length L (m), and the reaction temperature (°C). Regarding the confluence shape, [the data is provided in the original text]. Figure 5 The "T-shaped" registration diagram shown in the reaction section 22 is used when the reaction section 22 is registered. Figure 6In the case of reaction section 45 shown in the figure, a "cross" shape is registered (see reference). Figure 17 (etc.). Alternatively, the flow rate of the mixture can be added as an item in the Production Condition Data (PCD).

[0103] When reaction section 22 is used, system controller 26 controls the speed of the pump in the first raw material supply unit 20 and adjusts the flow rate of the first raw material RM1 so that the flow rate of the first raw material RM1 detected by the first flow rate sensor 35 matches the flow rate of the first raw material RM1 registered in the production condition data PCD. Similarly, system controller 26 controls the speed of the pump in the second raw material supply unit 21 and adjusts the flow rate of the second raw material RM2 so that the flow rate of the second raw material RM2 detected by the second flow rate sensor 36 matches the flow rate of the second raw material RM2 registered in the production condition data PCD.

[0104] When reaction section 45 is used, system controller 26 controls the speed of the pump in the first raw material supply unit 20 and adjusts the flow rate of the first raw material RM1 so that the average value of the flow rate of the first raw material RM1 detected by the first flow rate sensor 51 and the flow rate of the first raw material RM1 detected by the second flow rate sensor 52 matches the flow rate of the first raw material RM1 registered in the production condition data PCD. Similarly, system controller 26 controls the speed of the pump in the second raw material supply unit 21 and adjusts the flow rate of the second raw material RM2 so that the flow rate of the second raw material RM2 detected by the third flow rate sensor 53 matches the flow rate of the second raw material RM2 registered in the production condition data PCD.

[0105] Furthermore, the system controller 26 drives the temperature regulation unit 23 to make the reaction temperature detected by the temperature sensor 38 consistent with the reaction temperature registered in the production condition data PCD.

[0106] If the deviation between the values ​​detected by sensors 35, 36, 38, 51, 52, and 53 and the values ​​registered in the Production Condition Data (PCD) exceeds a preset range, the system controller 26 controls the three-way valve 42 to switch to the waste line and guide the waste to the waste section 41. Furthermore, when a reaction fails and waste is generated, the physical property data (PD) and quality data (QD) are not output. Therefore, the Production Condition Data (PCD) generated when waste is generated is discarded instead of being sent to the learning device 10.

[0107] Figure 8 In this process, the physical property analysis device 14 performs spectral analysis on the product PR, and detects the spectral data SPD as physical property data PD. As for spectral analysis, the physical property analysis device 14 performs, for example, near-infrared spectroscopy, Fourier transform infrared spectroscopy, Raman spectroscopy, and nuclear magnetic resonance spectroscopy. Figure 8The spectral data SPD illustrated herein is obtained through Raman spectroscopy analysis and consists of groups of wavenumber and intensity. That is, the spectral data SPD is an example of the "multidimensional physical property data" involved in this disclosure. Plotting the intensity of the spectral data SPD by wavenumber and connecting the points with lines yields the spectrum SP. Alternatively, the spectral data SPD can be constructed using groups of wavelength and intensity instead of groups of wavenumber and intensity.

[0108] Figure 9 In this process, the quality evaluation device 15 outputs the molecular weight dispersion and molecular weight of product PR as quality data QD. Here, molecular weight is the number-average molecular weight. Furthermore, the molecular weight dispersion is the value obtained by dividing the weight-average molecular weight by the number-average molecular weight. The quality evaluation device 15 uses, for example, a polystyrene solution obtained by dissolving product PR, i.e., polystyrene, in tetrahydrofuran, and determines the molecular weight dispersion and molecular weight by gel permeation chromatography (hereinafter referred to as GPC).

[0109] GPC is conducted under the following conditions.

[0110] Device: HLC-8220GPC (manufactured by Tosoh Corporation)

[0111] Detector: Differential refractometer (RI (Refractive Index) detector)

[0112] Front-column: TSKGUARDCOLUMN HXL-L 6mm×40mm (manufactured by Tosoh Corporation)

[0113] Sample side column: Connect directly to the following 3 columns (1) to (3) in sequence (all manufactured by Tosoh Corporation)

[0114] (1) TSK-GEL GMHXL 7.8mm×300mm

[0115] (2)TSK-GEL G4000HXL 7.8mm×300mm

[0116] (3)TSK-GEL G2000HXL 7.8mm×300mm

[0117] Reference side pillar: TSK-GEL G1000HXL 7.8mm×300mm

[0118] Thermostatic bath temperature: 40℃

[0119] Mobile layer: Tetrahydrofuran

[0120] Sample-side moving layer flow rate: 1.0 ml / min

[0121] Reference side moving layer flow rate: 1.0 ml / min

[0122] Sample concentration: 0.1% by mass

[0123] Sample injection volume: 100 μl

[0124] Data extraction time: 5 to 45 minutes after sample injection

[0125] Sampling interval: 300msec

[0126] Alternatively, various methods such as infrared spectroscopy, nuclear magnetic resonance spectroscopy, high-performance liquid chromatography (HPLC), or gas chromatography (GC) can be used instead of GPC. Furthermore, the quality data (QD) is not limited to the molecular weight dispersion and molecular weight of the product PR. When the product PR is obtained in solution, the concentration of the product PR in the solution, i.e., the molar concentration, can also be used as the quality data QD. Alternatively, the yield of the product PR, obtained by dividing the amount of product PR by the amount of raw material RM, can also be used as the quality data QD. Moreover, in cases where by-products are produced, the yield of the by-products can also be used as the quality data QD.

[0127] Figure 10 In this system, the basic structure of the computer constituting the learning device 10 and the application device 11 is the same, including a storage device 60, a memory 61, a CPU (Central Processing Unit) 62, a communication unit 63, a display 64, and an input device 65. They are interconnected via a bus 66.

[0128] Storage device 60 is a hard disk drive built into the computer constituting learning device 10, or connected via cable or network. Alternatively, storage device 60 is a disk array with multiple hard disk drives connected together. Storage device 60 stores control programs such as operating systems, various application programs, and various data associated with these programs. Alternatively, it can replace a hard disk drive or use a solid-state drive.

[0129] Memory 61 is the working memory used by CPU 62 for performing processing. CPU 62 loads the program stored in storage device 60 into memory 61, performs processing according to the program, and thus centrally controls various parts of the computer.

[0130] The communication unit 63 is a network interface that controls the transmission of various information through network 12. The display 64 shows various screens. The computer constituting the learning device 10, etc., receives input from the input device 65 via various screens. The input device 65 may be a keyboard, mouse, touch panel, etc.

[0131] In addition, in the following description, the subscript "A" will be attached to each part of the learning device 10, and the subscript "B" will be attached to each part of the operating device 11 for distinction.

[0132] Figure 11 In the learning device 10, a first operating program 70 is stored in the storage device 60A. The first operating program 70 is an application program for enabling the computer to function as the learning device 10. That is, the first operating program 70 is an example of the "operating program of the learning device" involved in the technology of this disclosure.

[0133] The storage device 60A also stores production condition data (PCD) from the flow reaction unit 13, physical property data (PD) from the physical property analysis unit 14, and quality data (QD) from the quality evaluation unit 15. Furthermore, the storage device 60A also stores related data (PRD) derived from the physical property data (PD) and a machine learning model (M). Multiple sets of production condition data (PCD), physical property data (PD), related data (PRD), and quality data (QD) are stored together.

[0134] If the first action program 70 is started, the CPU 62A of the computer constituting the learning device 10, together with the memory 61, etc., will function as the first read / write (hereinafter referred to as RW) control unit 75, the first export unit 76, the learning unit 77, the retrieval unit 78, and the transmission control unit 79.

[0135] The first RW control unit 75 controls the reading and storage of various data in the storage device 60A. The first RW control unit 75 reads physical property data PD from the storage device 60A and outputs the physical property data PD to the first export unit 76. Furthermore, the first RW control unit 75 stores the relevant data PRD from the first export unit 76 into the storage device 60A.

[0136] The first RW control unit 75 reads relevant data PRD, production condition data PCD, and quality data QD from the storage device 60A and outputs them to the learning unit 77. Since the relevant data PRD and production condition data PCD constitute learning input data IDL, the first RW control unit 75 obtains the learning input data IDL by reading the relevant data PRD and production condition data PCD from the storage device 60A. That is, the first RW control unit 75 is an example of the "first acquisition unit" disclosed herein.

[0137] The first RW control unit 75 reads the machine learning model M from the storage device 60A and outputs the machine learning model M to any one of the learning unit 77, the extraction unit 78, and the transmission control unit 79. Furthermore, the first RW control unit 75 stores the machine learning model M from the learning unit 77 into the storage device 60A.

[0138] The first export unit 76 receives physical property data PD from the first RW control unit 75. The first export unit 76 exports related data PRD from the physical property data PD. The first export unit 76 assigns the exported related data PRD the same ID as the physical property data PD and outputs the related data PRD to the first RW control unit 75. Each time a new physical property data PD is sent from the physical property analysis device 14, the first export unit 76 exports the related data PRD.

[0139] Learning unit 77 receives learning input data IDL, quality data QD, and machine learning model M from the first RW control unit 75. Learning unit 77 provides the learning input data IDL to the machine learning model M for learning and outputs a temporary machine learning model PM (see reference). Figure 16 That is, learning unit 77 is an example of a "temporary learning unit" involved in the technology disclosed herein. Hereinafter, the machine learning model M used during the period when learning unit 77 functions as a temporary learning unit will be denoted as the first machine learning model M1 (see reference). Figure 15 ).

[0140] Extraction unit 78 receives machine learning model M from first RW control unit 75. The machine learning model M received by extraction unit 78 from first RW control unit 75 is a temporary machine learning model PM. Extraction unit 78 uses the temporary machine learning model PM to extract high-contribution items from multiple items in the relevant data PRD. High-contribution items are those that contribute to improving the accuracy of product PR quality prediction and meet pre-set conditions. Extraction unit 78 outputs the extraction result of high-contribution items, i.e., high-contribution item information HCII, to learning unit 77. Furthermore, although not illustrated, extraction unit 78 outputs high-contribution item information HCII to first RW control unit 75, and first RW control unit 75 stores the high-contribution item information HCII in storage device 60A.

[0141] Learning unit 77 receives high-contribution project information HCII from extraction unit 78. Based on the high-contribution project information HCII, learning unit 77 selectively provides relevant data PRD of high-contribution projects to machine learning model M for learning, and outputs machine learning model M as the learned model TM. That is, learning unit 77 is an example of a "formal learning unit" involved in the technology of this disclosure. Hereinafter, the machine learning model M used during the period when learning unit 77 functions as a formal learning unit will be referred to as the second machine learning model M2 (see reference). Figure 23 ).

[0142] The transmission control unit 79 receives the machine learning model M from the first RW control unit 75. The machine learning model M received by the transmission control unit 79 from the first RW control unit 75 is a learned model TM. The transmission control unit 79 controls the transmission of the learned model TM to the application device 11. Furthermore, the transmission control unit 79 controls the transmission to receive high contribution item information HCII from the first RW control unit 75 along with the learned model TM, and also transmits the high contribution item information HCII to the application device 11.

[0143] like Figure 12 and Figure 13 As shown, the first derivation unit 76 derives correlation data PRD for each of the multiple intervals that divide the spectral data SPD. The first derivation unit 76 derives representative values ​​of intensity as correlation data PRD. Furthermore, this correlation data PRD is an example of "multidimensional physical property correlation data" involved in the technology of this disclosure.

[0144] Figure 12 In the example, the spectral data SPD is divided into multiple intervals, exemplified by dividing the SPD into 20 equal parts: INT1, INT2, INT3, ..., INT20. Furthermore, Figure 13 The example shown illustrates how representative values ​​(mean, maximum, minimum, median, variance, skewness, and kurtosis) of the intensity of each spectral data point SPD_INT1 to SPD_INT20 in each interval INT1 to INT20 are used as related data PRD_INT1 to PRD_INT20. In this example, since there are 7 representative values ​​and 20 intervals, the total number of related data PRDs combining PRD_INT1 to PRD_INT20 is 7 × 20 = 140.

[0145] like Figure 14As shown, the learning unit 77 includes a first processing unit 85, an evaluation unit 86, and an update unit 87. The first processing unit 85 provides the learning input data IDL to the machine learning model M and outputs the learning output data ODL from the machine learning model M. Similar to the quality data QD, the learning output data ODL consists of molecular weight dispersion and molecular weight (see reference). Figure 15 (etc.). The first processing unit 85 outputs the learning output data ODL to the evaluation unit 86.

[0146] Evaluation unit 86 receives the learning output data ODL from the first processing unit 85. Evaluation unit 86 compares the learning output data ODL with the quality data QD to evaluate the prediction accuracy of the machine learning model M. Evaluation unit 86 outputs the evaluation result to update unit 87.

[0147] Evaluation Unit 86, for example, uses a loss function to evaluate the prediction accuracy of the machine learning model M. The loss function represents the degree of difference between the learning output data ODL and the quality data QD. The closer the calculated value of the loss function is to 0, the higher the prediction accuracy of the machine learning model M.

[0148] The update unit 87 updates the machine learning model M based on the evaluation results from the evaluation unit 86. For example, the update unit 87 changes the values ​​of various parameters of the machine learning model M using methods such as stochastic gradient descent with accompanying learning coefficients. The learning coefficients represent the magnitude of change in the values ​​of various parameters of the machine learning model M. That is, the larger the learning coefficients, the greater the magnitude of change in the values ​​of various parameters, and the larger the update process of the machine learning model M.

[0149] The input of learning data IDL for the machine learning model M by the first processing unit 85 and the output of learning data ODL for the evaluation unit 86 are continuously repeated, the evaluation of prediction accuracy by the evaluation unit 86 and the updating of the machine learning model M by the update unit 87 are performed until the prediction accuracy reaches a preset level.

[0150] like Figure 15As shown, during the period when the learning unit 77 functions as a temporary learning unit, the first processing unit 85 provides multiple items of the production condition data (PCD) and related data (PRD) as learning input data (IDL) to the first machine learning model M1. Then, the evaluation unit 86 evaluates the prediction accuracy of the first machine learning model M1, and the update unit 87 updates the first machine learning model M1. By continuously repeating this process while changing the learning input data (IDL), the prediction accuracy of the first machine learning model M1 reaches a preset level. The first machine learning model M1, whose prediction accuracy has reached the preset level, is stored as a temporary machine learning model (PM) in the storage device 60A via the first RW control unit 75.

[0151] like Figure 16 As shown, the extraction unit 78 includes a second processing unit 90, a third processing unit 91, a calculation unit 92, and a determination unit 93. The second processing unit 90 provides the learning input data IDL to the temporary machine learning model PM, and outputs temporary output data POD from the temporary machine learning model PM. Similar to the learning output data ODL, the temporary output data POD is composed of molecular weight dispersion and molecular weight (see reference). Figure 17 (etc.). The second processing unit 90 outputs the temporary output data POD to the computing unit 92.

[0152] The third processing unit 91 provides the temporary machine learning model PM with the same production condition data (PCD) and extraction property related data (hereinafter referred to as extraction related data) (PRDE) as the data provided to the temporary machine learning model PM in the second processing unit 90, and outputs extraction temporary output data (PODE) from the temporary machine learning model PM. Similar to the learning output data (ODL), the extraction temporary output data (PODE) consists of molecular weight dispersion and molecular weight (see reference). Figure 18 (etc.). The third processing unit 91 outputs the temporary output data PODE for extraction to the calculation unit 92.

[0153] The calculation unit 92 receives temporary output data POD from the second processing unit 90 and extraction temporary output data PODE from the third processing unit 91. Based on the temporary output data POD and extraction temporary output data PODE, the calculation unit 92 calculates the contribution of each item in the relevant data PRD to improving the quality prediction accuracy of the first machine learning model M1. The calculation unit 92 outputs the contribution calculation result, i.e., contribution information CI, to the determination unit 93.

[0154] The determination unit 93 receives contribution information CI from the calculation unit 92. Based on the contribution information CI and the set conditions SC, the determination unit 93 determines whether each item among the multiple items of the relevant data PRD is a high contribution item. The determination unit 93 outputs high contribution item information HCII to the learning unit 77 as the determination result.

[0155] like Figure 17 As shown, in the second processing unit 90, with Figure 15 Similarly, in the case of the first processing unit 85 shown, multiple items of the relevant data PRD are completely provided to the temporary machine learning model PM. In contrast, as... Figure 18 As shown, in the third processing unit 91, the relevant data PRDE for extracting one item from among multiple items of the relevant data PRD is provided to the temporary machine learning model PM. Furthermore, Figure 18 The example shown is an item that excludes the average value of the relevant data PRD_INT1 in the first interval INT1.

[0156] like Figure 19 and Figure 20 As shown, the calculation unit 92 first calculates the rate of change based on the temporary output data POD and the extraction temporary output data PODE. The rate of change is obtained by the following (Equation 1).

[0157] Rate of change = |Difference between temporary output data and temporary output data used for extraction| / temporary output data ... (Equation 1)

[0158] The rate of change is a value that indicates the extent to which the output data of the temporary machine learning model PM changes after the influence of items excluded from the relevant data PRDE is extracted.

[0159] Next, the calculation unit 92 uses a conversion table 100 to convert the rate of change into a contribution level. In the conversion table 100, the contribution level is recorded as 0 when the rate of change is 0 or higher and less than 0.05, as 1 when the rate of change is 0.05 or higher and less than 0.1, ..., as 9 when the rate of change is 0.45 or higher and less than 0.5, and as 10 when the rate of change is 0.5 or higher.

[0160] Figure 20 Examples Figure 17 The provisional output data shown in the figure has a molecular weight dispersion of 1.5865 and a molecular weight of 22000 for POD. Figure 18The example shown is for the temporary output data PODE used for extraction, where the molecular weight dispersion is 1.6043 and the molecular weight is 26000. In this case, the rate of change of molecular weight dispersion is |1.5865-1.6043| / 1.5865≈0.01. According to conversion table 100, the contribution is 0 when the rate of change is 0.01, therefore the contribution of the rate of change of molecular weight dispersion is calculated as 0. Similarly, the rate of change of molecular weight is |22000-26000| / 22000≈0.18, therefore the contribution of the rate of change of molecular weight is calculated as 3. The contribution calculated in this way is the item excluded from the relevant data PRDE used for extraction (in... Figure 18 The contribution of the items (the average value of the relevant data PRD_INT1) to the project.

[0161] The third processing unit 91 sequentially provides the extraction-related data PRDEs to the temporary machine learning model PM while changing the excluded items one by one. In this example, since the number of items in the related data PRDEs is 140, the third processing unit 91 provides 140 extraction-related data PRDEs to the temporary machine learning model PM. Furthermore, the calculation unit 92 calculates the contribution for each extraction-related data PRDE. In addition, to improve the reliability of the rate of change value, the rate of change can be calculated not only for a single set of production condition data PCD and related data PRDs, but also for multiple different sets of production condition data PCDs and related data PRDs, and their average value is converted into a contribution.

[0162] like Figure 21 As shown, the contribution information (CI) registers the contribution of each item in the relevant data (PRD). The two values ​​next to each item represent the contribution: the left value represents the contribution of the rate of change of molecular weight dispersion after conversion, and the right value represents the contribution of the rate of change of molecular weight after conversion. For example, the contribution of the rate of change of molecular weight dispersion after converting the maximum value of INT1 in the first interval is 8, and the contribution of the rate of change of molecular weight after conversion is 6.

[0163] Figure 22 This indicates that the setting condition SC is "the contribution of the rate of change of molecular weight dispersion and the contribution of the rate of change of molecular weight are both 6 or higher", and the contribution information CI is as follows: Figure 21 The example shown. In this case, the determination unit 93 determines the items represented by the shaded lines indicating the maximum value, variance, skewness, and kurtosis of the first interval INT1, the median value, variance, skewness, and kurtosis of the second interval INT2, ..., and the average value, maximum value, minimum value, and skewness of the 20th interval INT20 as high contribution items.

[0164] Furthermore, this method of extracting essential components (high-contribution items in this case) from multiple data sets (multiple items in the related data PRD) is called sparse modeling. Such sparse modeling can be performed using, for example, the glmnet package, which works on the R language. Detailed algorithms for sparse modeling are described, for example, in "Regularization Paths for Generalized Linear Models via Coordinate Descent" in the Journal of Statistical Software, vol. 33-1 (2010).

[0165] like Figure 23 As shown, during the period when the learning unit 77 is functioning as a formal learning unit, in the first processing unit 85, the relevant data PRD of high-contribution items is selectively provided to the second machine learning model M2. In other words, items other than high-contribution items are selectively excluded and provided to the second machine learning model M2. Then, the prediction accuracy of the second machine learning model M2 is evaluated by the evaluation unit 86, and the second machine learning model M2 is updated by the update unit 87. By continuously repeating this process while changing the learning input data IDL, the prediction accuracy of the second machine learning model M2 reaches a preset level. The second machine learning model M2, whose prediction accuracy has reached the preset level, is stored in the storage device 60A as a learned model TM by the first RW control unit 75. In addition, the second machine learning model M2 used during the period when the learning unit 77 is functioning as a formal learning unit has the same characteristics such as type and performance as the first machine learning model M1 used during the period when the learning unit 77 is functioning as a temporary learning unit.

[0166] Figure 24 In the application device 11, the storage device 60B stores the second operating program 110. The second operating program 110 is an application program used to enable the computer to function as the application device 11.

[0167] The storage device 60B also stores the learned model TM and high contribution item information HCII from the learning device 10, and the predictive property data PDF from the property analysis device 14. Furthermore, the storage device 60B also stores the predictive production condition data PCDF. The predictive production condition data PCDF is input by the operator via the input device 65B. More specifically, an input screen displaying the input boxes for each item of the predictive production condition data PCDF is shown on the display 64B, and the predictive production condition data PCDF is input via the input screen. The predictive production condition data PCDF and the predictive property data PDF are the production condition data PCD and property data PD for the product PR whose quality is unknown and whose quality will be predicted using the learned model TM in the future.

[0168] Furthermore, the storage device 60B also stores prediction property related data (hereinafter referred to as prediction related data) PRDF derived from the prediction property data PDF.

[0169] If the second action program 110 is started, the CPU 62B and memory 61 of the computer constituting the application device 11 cooperate to function as the second RW control unit 115, the second output unit 116, the fourth processing unit 117, and the display control unit 118.

[0170] Similar to the first RW control unit 75 of the learning device 10, the second RW control unit 115 controls the reading and storage of various data in the storage device 60B. The second RW control unit 115 reads the prediction property data PDF and high contribution item information HCII from the storage device 60B, and outputs the prediction property data PDF and high contribution item information HCII to the second export unit 116. Furthermore, the second RW control unit 115 stores the prediction correlation data PRDF from the second export unit 116 into the storage device 60B.

[0171] The second RW control unit 115 reads the learned model TM from the storage device 60B and outputs the learned model TM to the fourth processing unit 117. The second RW control unit 115 acquires the learned model TM by reading the learned model TM from the storage device 60B. That is, the second RW control unit 115 is an example of the "second acquisition unit" involved in the technology of this disclosure.

[0172] The second RW control unit 115 reads the prediction-related data PRDF and the prediction-related production condition data PCDF from the storage device 60B and outputs them to the fourth processing unit 117. The second RW control unit 115 acquires the prediction-related data PRDF by reading it from the storage device 60B. That is, the second RW control unit 115 is an example of the "third acquisition unit" involved in the technology of this disclosure.

[0173] The second export unit 116 receives the prediction property data PDF and high contribution item information HCII from the second RW control unit 115. The second export unit 116 exports the prediction correlation data PRDF from the prediction property data PDF. More specifically, similar to the first export unit 76 of the learning device 10, the second export unit 116 exports the average, maximum, minimum, median, variance, skewness, and kurtosis of the intensity for each of the multiple intervals INT1 to INT20 that divide the spectral data SPD. However, the second export unit 116 selectively exports high contribution items based on the high contribution item information HCII and does not export items other than high contribution items. Therefore, compared with... Figure 23 Similarly, the relevant data PRD for prediction becomes data that selectively excludes projects other than those with high contributions (see reference). Figure 25 ).

[0174] The fourth processing unit 117 receives the predictive production condition data (PCDF), the predictive correlation data (PRDF), and the learned model TM from the second RW control unit 115. The fourth processing unit 117 provides the predictive production condition data (PCDF) and the predictive correlation data (PRDF) to the learned model TM to predict quality. That is, the fourth processing unit 117 is an example of a "processing unit" according to the technology of this disclosure. The fourth processing unit 117 outputs the quality prediction result, i.e., the quality prediction data (QFD), obtained through the learned model TM to the display control unit 118. Similar to the quality data QD, the quality prediction data QFD is composed of molecular weight dispersion and molecular weight (see reference). Figure 25 wait).

[0175] The display control unit 118 controls the display of various screens on the display 64B. Among these various screens, in addition to the aforementioned input screen for predictive production condition data (PCDF), there is also a quality prediction display screen 120 for displaying quality prediction data (QFD) (see reference). Figure 26 That is, the display control unit 118 is an example of the "output control unit" involved in the technology of this disclosure.

[0176] Figure 25 This represents the scenario where the learned model TM is used to predict the quality of product PR (ID PR0500) with unknown quality. (And...) Figure 23Similarly, in the fourth processing unit 117, the prediction correlation data PRDF for high-contribution items among multiple items is selectively provided to the fully learned model TM. In other words, items other than high-contribution items are selectively excluded and provided to the fully learned model TM.

[0177] Figure 26 In the process, under the control of the display control unit 118, the quality prediction display screen 120 on the display 64B displays the quality prediction data QFD output from the learned model TM. By selecting the OK button 121, the display of the quality prediction display screen 120 disappears.

[0178] Next, refer to Figures 27-31 The flowchart below explains the effects of the aforementioned structure. First, if the first action program 70 is initiated in the learning device 10, then... Figure 11 As shown, the CPU 62A of the learning device 10 functions as the first RW control unit 75, the first export unit 76, the learning unit 77, the extraction unit 78, and the transmission control unit 79. Figure 14 As shown, the learning unit 77 functions as the first processing unit 85, the evaluation unit 86, and the updating unit 87. Furthermore, as... Figure 16 As shown, the extraction unit 78 functions as the second processing unit 90, the third processing unit 91, the calculation unit 92, and the determination unit 93.

[0179] In the learning device 10, the first RW control unit 75 reads the physical property data PD, i.e., the spectral data SPD, from the storage device 60A and outputs it to the first export unit 76. Then, as... Figure 12 and Figure 13 As shown, the first export unit 76 exports representative values ​​(mean, maximum, minimum, median, variance, skewness, and kurtosis) of intensity as correlation data PRD for each of the multiple intervals INT1 to INT20 of the spectral data SPD. The correlation data PRD is output from the first export unit 76 to the first RW control unit 75 and stored in the storage device 60A by the first RW control unit 75. The correlation data PRD is exported each time a new physical property data PD is sent from the physical property analysis device 14.

[0180] Figure 27 In the learning device 10, the main processes are temporary learning (step ST100), extraction of high-contribution items (step ST200), and formal learning (step ST300). In temporary learning, a first machine learning model M1 is learned as a temporary machine learning model PM. The temporary machine learning model PM is used in the extraction of high-contribution items. In formal learning, a second machine learning model M2 is learned as the completed model TM.

[0181] like Figure 28 As shown, in the temporary learning process, the first step is to read the learning input data IDL, which consists of production condition data PCD and related data PRD set in the production process, from the storage device 60A via the first RW control unit 75 (step ST100) 1). The learning input data IDL is then output from the first RW control unit 75 to the learning unit 77. Step ST1001 is an example of the "first acquisition step" covered by the technology of this disclosure.

[0182] In Study Department 77, such as Figure 15 As shown, in the first processing unit 85, multiple items of the production condition data (PCD) and related data (PRD) are provided as learning input data (IDL) to the first machine learning model M1, and learning output data (ODL) is output from the first machine learning model M1 (step ST1002). Then, in the evaluation unit 86, the learning output data (ODL) is compared with the quality data (QD) to evaluate the prediction accuracy of the product PR quality made by the first machine learning model M1 (step ST1003).

[0183] If the evaluation result of the prediction accuracy of the first machine learning model M1 performed by the evaluation unit 86 is that the prediction accuracy of the first machine learning model M1 is less than a preset level ("No" in step ST1004), the first machine learning model M1 is updated by the update unit 87 (step ST1005). Then, using the updated first machine learning model M1, the processes of steps ST1001, ST1002, and ST1003 are repeated. If the evaluation result of the prediction accuracy of the first machine learning model M1 performed by the evaluation unit 86 is that the prediction accuracy of the first machine learning model M1 reaches a preset level ("Yes" in step ST1004), each of these processes of steps ST1001 to ST1003 ends. The first machine learning model M1 with the prediction accuracy reaching the preset level is output as a temporary machine learning model PM from the learning unit 77 to the first RW control unit 75 (step ST1006). The temporary machine learning model PM is stored in the storage device 60A by the first RW control unit 75. Furthermore, this series of steps ST1001 to ST1006 is an example of a "temporary learning step" involved in the technology of this disclosure.

[0184] like Figure 29 As shown, in the extraction of high-contribution projects, such as Figure 16 and Figure 17As shown, in the second processing unit 90, multiple items of production condition data PCD and related data PRD are completely provided to the temporary machine learning model PM, and temporary output data POD is output from the temporary machine learning model PM (step ST2001). The temporary output data POD is output from the second processing unit 90 to the calculation unit 92.

[0185] And, as Figure 16 and Figure 18 As shown, in the third processing unit 91, production condition data PCD and extraction related data PRDE (excluding related data PRD) are provided to the temporary machine learning model PM, and extraction temporary output data PODE is output from the temporary machine learning model PM (step ST2002). Extraction temporary output data PODE is output from the third processing unit 91 to the calculation unit 92.

[0186] like Figure 19 and Figure 20 As shown, in the calculation unit 92, the contribution of one item excluded from the extraction related data PRDE is calculated based on the temporary output data POD and the extraction temporary output data PODE (step ST2003). Steps ST2002 and ST2003 are repeated while changing the excluded items one by one until the contribution of all items in the related data PRD is calculated ("Yes" in step ST2004). The contribution of all items in the related data PRD calculated in this way is used as... Figure 21 The contribution information CI shown in the figure is output from the calculation unit 92 to the determination unit 93.

[0187] like Figure 22 As shown, in the determination unit 93, based on the contribution information CI and the set conditions SC, it is determined whether each item among the multiple items of the relevant data PRD is a high contribution item (step ST2005). Then, the determination result is output from the determination unit 93 to the learning unit 77 as high contribution item information HCII (step ST2006). Furthermore, this series of steps ST2001 to ST2006 is an example of the "extraction steps" involved in the technology of this disclosure.

[0188] like Figure 30 As shown, during formal learning, the first RW control unit 75 reads the learning input data IDL, which consists of production condition data PCD and related data PRD, from the storage device 60A (step ST3001). The learning input data IDL is then output from the first RW control unit 75 to the learning unit 77.

[0189] In Study Department 77, such as Figure 23As shown, in the first processing unit 85, production condition data (PCD) and related data (PRD) of high-contribution projects are selectively provided as learning input data (IDL) to the second machine learning model M2, and learning output data (ODL) is output from the second machine learning model M2 (step ST3002). Then, in the evaluation unit 86, the learning output data (ODL) is compared with the quality data (QD) to evaluate the prediction accuracy of product PR quality by the second machine learning model M2 (step ST3003).

[0190] If the evaluation result of the prediction accuracy of the second machine learning model M2 performed by the evaluation unit 86 is that the prediction accuracy of the second machine learning model M2 is less than the preset level ("No" in step ST3004), the second machine learning model M2 is updated by the update unit 87 (step ST3005). Then, using the updated second machine learning model M2, the processes of steps ST3001, ST3002, and ST3003 are repeated. If the evaluation result of the prediction accuracy of the second machine learning model M2 performed by the evaluation unit 86 is that the prediction accuracy of the second machine learning model M2 reaches the preset level ("Yes" in step ST3004), each of these processes of steps ST3001 to ST3003 ends. The second machine learning model M2 with the prediction accuracy reaching the preset level is output from the learning unit 77 to the first RW control unit 75 as the learned model TM (step ST3006). The learned model TM is stored in the storage device 60A by the first RW control unit 75. Then, the learned model TM is sent to the application device 11 via the transmission control unit 79. Furthermore, this series of steps ST3001 to ST3006 is an example of a "formal learning step" covered by the technology of this disclosure.

[0191] If the second action procedure 110 is initiated in the operating device 11, then as follows Figure 24 As shown, the CPU 62B of the application device 11 functions as the second RW control unit 115, the second output unit 116, the fourth processing unit 117, and the display control unit 118.

[0192] In the application device 11, the prediction property data PDF and high contribution item information HCII are read from the storage device 60B by the second RW control unit 115 and output to the second export unit 116. Then, the second export unit 116 derives representative intensity values ​​for each of the multiple intervals INT1 to INT20 of the spectral data SPD, which is divided into prediction property data PDF, limited to high contribution items. In this way, prediction correlation data PRDF is exported in the second export unit 116. The prediction correlation data PRDF is stored in the storage device 60B by the second RW control unit 115.

[0193] like Figure 31 As shown, in the application device 11, the learned model TM is read from the storage device 60B via the second RW control unit 115 (step ST500). Furthermore, the prediction production condition data PCDF and prediction correlation data PRDF set in the production process are read from the storage device 60B via the second RW control unit 115 (step ST600). The learned model TM, the prediction production condition data PCDF, and the prediction correlation data PRDF are output from the second RW control unit 115 to the fourth processing unit 117.

[0194] like Figure 25 As shown, in the fourth processing unit 117, the production condition data (PCDF) for prediction and the related data (PRDF) for prediction are provided to the learned model TM, and the quality prediction data (QFD) is output from the learned model TM (step ST700). The quality prediction data (QFD) is output from the fourth processing unit 117 to the display control unit 118.

[0195] Displayed on the monitor 64B via the display control unit 118 Figure 26 The quality prediction display screen 120 shown is shown in step ST800. Through this quality prediction display screen 120, the operator can view the quality prediction data QFD.

[0196] As described above, the learning device 10 acquires learning input data IDL containing relevant data PRD via the first RW control unit 75. In the learning unit 77, the learning input data IDL is provided to the first machine learning model M1 for learning, and a temporary machine learning model PM is output. Next, the extraction unit 78 uses the temporary machine learning model PM to extract high-contribution items from multiple items in the relevant data PRD. Then, in the learning unit 77, the relevant data PRD of the high-contribution items is selectively provided to the second machine learning model M2 for learning, and the learned model TM is output. Therefore, compared with the case where relevant data PRD other than high-contribution items is provided for learning, the prediction accuracy of product PR quality by the learned model TM can be improved.

[0197] More specifically, such as Figure 32As shown in Table 125, the following results were obtained: Compared with Comparative Examples 1 and 2, the example provided with Production Condition Data (PCD) and Related Data (PRD) of High Contribution Items as Learning Input Data (IDL) showed higher coefficients of determination for molecular weight dispersion and molecular weight. Furthermore, Comparative Example 1 involved providing only Production Condition Data (PCD) as Learning Input Data (IDL), while Comparative Example 2 involved providing and learning multiple items from both Production Condition Data (PCD) and Related Data (PRD) as Learning Input Data (IDL). Thus, it can be confirmed that by providing and learning the Related Data (PRD) of High Contribution Items as Learning Input Data (IDL), the prediction accuracy of product PR quality using the learned model TM can be further improved.

[0198] Furthermore, it can improve learning efficiency by preventing learning labor from being allocated to projects that contribute little to improving prediction accuracy.

[0199] The learning model uses input data IDL that includes not only relevant data PRD but also production condition data PCD. Therefore, it can output quality prediction data QFD from the machine learning model M that also takes into account the impact of production condition data PCD. Furthermore, it can extract high-contribution items by considering the impact of production condition data PCD, and further improve the appropriateness of high-contribution items.

[0200] The physical property data (PD) includes the spectral data (SPD) detected by spectral analysis of the product's physical property (PR). Furthermore, the correlation data (PRD) are representative intensity values ​​derived for each of the multiple intervals (INT1 to INT20) of the spectral data (SPD). Compared to using the intensity of all wavenumbers from the spectral data (SPD) as the correlation data (PRD), this reduces the amount of data in the correlation data (PRD).

[0201] The device 11 acquires the learned model TM and prediction correlation data PRDF via the second RW control unit 115. Next, the fourth processing unit 117 provides the prediction correlation data PRDF to the learned model TM to predict quality. Then, under the control of the display control unit 118, the display 64B displays a quality prediction display screen 120 containing the quality prediction result, i.e., quality prediction data QFD, obtained through the learned model TM. Therefore, the operator can easily grasp the quality level of the product PR without having to specifically apply it to the quality evaluation device 15 to actually evaluate the quality.

[0202] Here, when actually evaluating the quality of product PR in the quality evaluation device 15, the preprocessing of product PR and the quality evaluation process require a relatively long time, for example, about one to two weeks. In contrast, if a fully learned model TM is used, the quality of product PR can be predicted in a very short time. Moreover, as... Figure 32 As shown in the example, since the prediction accuracy of the learned model TM is relatively high, the operator can make full use of the quality prediction data (QFD) to formulate the production plan for future product PR.

[0203] [Second Implementation]

[0204] exist Figures 33-35 In the second embodiment shown, the image data IMD obtained by photographing the product PR is used as the physical property data PD.

[0205] Figure 33 In the second embodiment, the property analysis apparatus 130 is, for example, a digital optical microscope, which captures a product image (PR) and outputs the image data (IMD) as property data (PD). Furthermore, the image data (IMD) is an example of "multidimensional property data" as described in this disclosure.

[0206] like Figure 34 and Figure 35 As shown, in the second embodiment, the first export unit 135 exports relevant data PRD_AR1-1, PRD_AR1-2, ..., PRD_AR10-10 for each of the multiple regions AR1-1, AR1-2, ..., AR10-10 after the image data IMD is equally divided. Specifically, the first export unit 135 exports the average values ​​of the red, green, and blue pixel values ​​of the image data IMD_AR1-1 to IMD_AR10-10 for each region AR1-1 to AR10-10 as relevant data PRD_AR1-1 to PRD_AR10-10. Alternatively, instead of the average value of each color pixel value, or based on this, other representative values ​​such as the maximum value, minimum value, median value, and variance of each color pixel value may be exported as relevant data PRD. Furthermore, although the illustrations and detailed descriptions are omitted, in the second export section of the application device, similarly to the first export section 135, the average values ​​of the red, green, and blue pixels of each region AR1-1 to AR10-10 of the image data IMD of the product PR with unknown quality are exported as the prediction correlation data PRDF.

[0207] Thus, in the second embodiment, the image data IMD obtained from photographing the product PR is used as the physical property data PD. Therefore, it is easier to capture the physical properties of the product PR.

[0208] Furthermore, image data IMD can replace spectral data SPD in the first embodiment described above as property data PD, or it can be based on spectral data SPD as property data PD. Moreover, the property analysis device 130 is not limited to the exemplified digital optical microscope, but can also be a scanning electron microscope (SEM) or the like.

[0209] Multidimensional physical property data is not limited to the spectral data SPD of the first embodiment and the image data IMD of the second embodiment described above. It can be multidimensional physical property data related to any one of the five human senses: sight, hearing, smell, touch, and taste. Examples include tactile data of a product PR, odor data during product PR production, and audio data during product PR production. In the case of tactile data, the outputs of multiple tactile sensors on the product PR are used as tactile data, and the average value of the outputs of each tactile sensor is derived as related data PRD. In the case of audio data, audio data is recorded from the beginning to the end of production using a microphone, and this audio data is divided into multiple intervals. The average value of the frequency and the average value of the amplitude in each interval are derived as related data PRD.

[0210] [Third Implementation]

[0211] exist Figures 36-43 In the third embodiment shown, input image data IIMD, which serves as physical property data PD, is provided to the auto encoder AF, and output image data OIMD is output. Correlation data PRD is derived based on the difference data DD between the input image data IIMD and the output image data OIMD provided to the auto encoder AE.

[0212] Figure 36 In the third embodiment, the first export unit 140 uses an automatic encoder AE to export related data PRD from the input image data IIMD, which is physical property data PD. That is, the first export unit 140 is an example of an "export unit" according to the technology of this disclosure. In addition, the following describes the case where the image data SPIMD of the spectrum SP, which is an example of "multidimensional physical property data" according to the technology of this disclosure, is used as the input image data IIMD.

[0213] An autoencoder (AE) has multiple layers for analyzing input image data IIMD (Intense Image Depth). Each layer consists of a hierarchical machine learning model composed of convolutional neural networks that extract features of different spatial frequency bands contained in the IIMD of the input image data. Examples of convolutional neural networks include U-Net (U-Shaped Neural Network) and SegNet.

[0214] like Figure 37 As shown, the auto encoder AE consists of an encoder network 145 and a decoder network 146. The encoder network 145 uses a filter F (reference) for each layer. Figure 38 Convolution operations are performed, and the extracted image feature map (CMP) is then convolved. The decoder network 146 gradually enlarges the image size of the smallest image feature map (CMP) output from the lowest layer of the encoder network 145. Then, the gradually enlarged image feature map (CMP) is combined with the image feature maps (CMP) output from each layer of the encoder network 145 to generate output image data (OIMD) with the same image size as the input image data (IIMD).

[0215] Input data DI (reference) with multiple pixel values ​​arranged in a two-dimensional pattern Figure 38 The input image data IIMD is input into each layer of the encoder network 145. The encoder network 145 performs convolution processing on the input image data IIMD in each layer to extract an image feature map CMP. The input image data IIMD is input as input data DI into the topmost first layer of the encoder network 145. The first layer performs convolution processing on the input image data IIMD, for example, outputting an image feature map CMP with the same image size as the input image data IIMD. Below the second layer, the image feature maps CMP output from the upper layers are used as input data DI. Below the second layer, the image feature maps CMP are convolutionally processed, for example, outputting an image feature map CMP with the same image size as the input image feature map CMP.

[0216] Figure 38 In this process, convolution is performed as follows: A 3×3 filter F is applied to the input data DI, and the pixel value e of the pixel of interest Ip within the input data DI is convolved with the pixel values ​​a, b, c, d, f, g, h, i of the eight adjacent pixels Is of the pixel of interest Ip. This yields output data DIc, with the pixel values ​​arranged in two dimensions, similar to the input data DI. With the coefficients of filter F set to r, s, t, u, v, w, x, y, z, the pixel value k of pixel Icp in the output data DIc, resulting from the convolution operation on the pixel of interest In, is obtained, for example, by calculating the following (Equation 2).

[0217] k=az+by+cx+dw+ev+fu+gt+hs+ir……(Formula 2)

[0218] In convolution processing, each pixel of the input data DI undergoes a convolution operation as described above to output pixel values ​​k. This results in output data DIc with pixel values ​​k arranged in a two-dimensional array. One output data DIc is output to one filter F. When using multiple filters F of different types, output data DIc is output to each filter F.

[0219] like Figure 39 As shown, the output data DIC is the pixel value k arranged in a two-dimensional array, having width and height. Furthermore, when multiple filters F of different types are applied to output multiple output data DICs, the image feature map CMP becomes a collection of multiple output data DICs. In the image feature map CMP, the number of filters F is called the number of channels. Figure 39 The image feature map CMP shown is an example of a 4-channel image feature map CMP with 4 output data DICs applied to the outputs of 4 filters F.

[0220] like Figure 40 As shown, an autoencoder AE (AE_INT1 to AE_INT20) is prepared for each of the multiple intervals INT1 to INT20 of the spectral data SPD. Each autoencoder AE_INT1 to AE_INT20 is specifically used for the image data SPIMD_INT1 to SPIMD_INT20 of the spectral SP for each interval INT1 to INT20. Additionally, in the following description, the autoencoders AE_INT1 to AE_INT20 are sometimes collectively referred to as autoencoder AE.

[0221] like Figure 41 As shown, image data SPIMD of the spectrum SP of a product PR with a quality better than a preset grade is provided to the autoencoder AF as input image data IIMDL for learning. The autoencoder AE then learns to make the input image data IIMDL for learning consistent with the output image data OIMDL for learning. That is, when image data SPIMD of the spectrum SP of a product PR with a quality better than a preset grade is input as input image data IIMD, the autoencoder AE ideally outputs output image data OIMD with a spectrum SP having the same shape as the spectrum SP of the input image data IIMD. Furthermore, the preset grade is, for example, a molecular weight dispersion of 1.5 or higher, a molecular weight of 25,000 or higher, etc.

[0222] Figure 41 The example shows the autoencoder AE_INT1 from each of the autoencoders AE_INT1 to AE_INT20. Therefore, the image data SPIMD_INT1 is learned using the input image data IIMDL, which is the spectral SP of the interval INT1.

[0223] like Figure 42As shown, the first export unit 140 provides the image data SPIMD of the spectral SP as input image data IIMD to the autoencoder AE, and outputs output image data OIMD. Next, the first export unit 140 exports the difference data DD between the input image data IIMD and the output image data OIMD. Specifically, for each interval from INT1 to INT20, the first export unit 140 exports the difference in intensity of each wavenumber between the input image data IIMD and the output image data OIMD as difference data DD.

[0224] and Figure 41 Similarly, Figure 42 Only the autoencoder AE_INT1 is illustrated. The image data SPIMD_INT1 of the spectral SP of the INT1 interval is provided to the autoencoder AE_INT1 as the input image data IIMD. Then, based on the input image data IIMD and the output image data OIMD of the autoencoder AE_INT1, the differential data DD_INT1 of the INT1 interval is derived.

[0225] Here, as Figure 41 As shown, the autoencoder AE provides the spectral image data SPIMD of a product PR whose quality is better than a preset level as the learning input image data IIMDL and performs training. Therefore, if the quality of the product PR, the source of the input image data IIMD, is better than the preset level, the autoencoder AE outputs output image data OIMD that is approximately the same as the input image data IIMD. Thus, the intensity difference becomes relatively small. Conversely, if the quality of the product PR is below the preset level, the autoencoder AE outputs output image data OIMD that is different from the input image data IIMD. Thus, the intensity difference becomes relatively large. That is, based on the difference data DD, it can be determined whether the quality of the product PR is better than the preset level.

[0226] like Figure 43As shown, the first export unit 140 exports relevant data PRD_INT1, PRD_INT2, ..., PRD_INT20 for each data point in the difference data DD_INT1, DD_INT2, ..., DD_INT20 of each interval INT1 to INT20. Specifically, the first export unit 140 exports the average and sum of the intensity differences of each interval INT1 to INT20 as relevant data PRD_INT1 to PRD_INT20. Alternatively, instead of the average and sum of the intensity differences, or based on this, other representative values ​​such as the maximum, minimum, median, and variance of the intensity differences can be exported as relevant data PRD. Furthermore, although the illustrations and detailed descriptions are omitted, in the second export unit of the application device, similarly to the first export unit 140, the average and sum of the intensity differences of each interval INT1 to INT20 of the image data SPIMD of the spectral SP of the product PR with unknown quality are exported as prediction relevant data PRDF.

[0227] Thus, in the third embodiment, the first output unit 140 provides input image data IIMD, which serves as property data PD, to the autoencoder AE and outputs output image data OIMD. Correlation data PRD is derived based on the difference data DD between the input image data IIMD and the output image data OIMD provided to the autoencoder AE. Therefore, compared to representative values ​​of intensity such as average, maximum, minimum, median, variance, skewness, and kurtosis exemplified in the first embodiment, numerical values ​​accurately representing the property of the product PR (here, the average and sum of intensity differences) can be derived as correlation data PRD. Therefore, the prediction accuracy of the learned model TM can be further improved.

[0228] In addition, the relevant data PRD based on the difference data DD can be used instead of the representative values ​​of intensity such as average, maximum, minimum, median, variance, skewness and kurtosis in the first embodiment described above, or it can be used on this basis.

[0229] The input image data IIMD is not limited to the spectral SP image data SPIMD illustrated above. Alternatively, the spectral SP image data SPIMD can be used instead, or the image data IMD obtained from the photographing product PR of the second embodiment described above can be used as the input image data IIMD. In this case, an automatic encoder AE is prepared for each region in the region AR of the image data IMD.

[0230] [Fourth Implementation]

[0231] exist Figure 44In the fourth embodiment shown, related data PRD is derived from the image feature map CMP output from a portion of the autoencoder AE, namely the encoder network 155. That is, the image feature map CMP is an example of "feature data" involved in the technology of this disclosure.

[0232] Figure 44 Similar to the third embodiment described above, the first output unit 150 of the fourth embodiment provides the image data SPIMD of the physical property data PD, i.e., the spectrum SP, to the autoencoder AF as input image data IIMD. Furthermore, an image feature map CMP is output from the encoder network 155 of the autoencoder AE. The image feature map CMP is, for example, the smallest image feature map CMP output from the lowest layer of the encoder network 155. Additionally, similar to the autoencoder AE of the third embodiment described above, the autoencoder AF of the fourth embodiment is prepared for each of the multiple intervals INT1 to INT20 of the spectrum data SPD. However, unlike the autoencoder AF of the third embodiment described above, the autoencoder AE of the fourth embodiment provides and learns the image data SPIMD of the spectrum SP of various products PR as learning input image data IIMDL, regardless of the quality level.

[0233] The first export section 150 exports relevant data PRD based on the image feature map CMP. Figure 44 The example illustrates an image feature map CMP composed of output data DIc from channel 1 (Ch1), output data DIc from channel 2 (Ch2), output data DIc from channel 3 (Ch3), and output data DIc from channel 4 (Ch4). Furthermore, an example is shown where the first extraction unit 150 derives the average pixel value as correlation data PRD for each data point in the output data DIc from channels 1 to 4. Alternatively, the maximum value, minimum value, median value, dispersion, etc., can be derived as correlation data PRD instead of the average value. Although illustrations and detailed descriptions are omitted, in the second extraction unit of the application device, similarly to the first extraction unit 150, the average pixel value of the output data DIc from each interval INT1 to INT20 of the image feature map CMP of the spectral SP image data of a product PR with unknown quality is derived as prediction correlation data PRDF.

[0234] Figure 44Only the autoencoder AE_INT1 is shown in the diagram. The image data SPIMD_INT1 of the spectral SP of the INT1 interval is provided to the autoencoder AE_INT1 as input image data IIMD. The first export unit 150 exports the related data PRD_INT1 of the INT1 interval from the image feature map CMP of the encoder network 155 of the autoencoder AF_INT1. In addition, although the illustration and explanation are omitted, the related data PRD_INT2 to PRD_INT20 are also exported from the image feature map CMP of the encoder network 155 of the autoencoder AE_INT2 to AE_INT20 for the other intervals INT2 to INT20.

[0235] Thus, in the fourth embodiment, relevant data PRD is derived from the image feature map CMP output from the encoder network 155 of the autoencoder AE. Therefore, similar to the third embodiment described above, the numerical value accurately representing the physical properties of the product PR (here, the average pixel value of the output data DIC) can be derived as relevant data PRD, further improving the prediction accuracy of the learned model TM.

[0236] Image feature maps (CMPs) can be output from layers other than the lowest layer of encoder network 155.

[0237] Furthermore, the relevant data PRD based on the image feature map CMP can be used in place of the representative values ​​of intensity such as mean, maximum, minimum, median, variance, skewness, and kurtosis in the first embodiment described above, or it can be used in a similar manner. Also, the relevant data PRD based on the image feature map CMP can be used in place of the relevant data PRD based on the difference data DD in the third embodiment described above, or it can be used in a similar manner.

[0238] In the fourth embodiment, similar to the third embodiment described above, the input image data IIMD is not limited to the spectral SP image data SPIMD exemplified. Alternatively, the spectral SP image data SPIMD can be used, or the image data IMD obtained from the photographing product PR of the second embodiment described above can be used as the input image data IIMD. In this case, an automatic encoder AE is prepared for each region in the region AR of the image data IMD.

[0239] The intervals INT that divide spectral data SPD can overlap. Similarly, the regions AR that divide image data IMD can overlap.

[0240] Product PR is not limited to products produced using flow synthesis. For example, it can also include products produced using batch synthesis.

[0241] Furthermore, in the above embodiments, the production conditions received by the setting unit 25 of the flow reaction device 13 are used as production condition data PCD, but this is not a limitation. The measured values ​​detected by the first flow rate sensor 35, the second flow rate sensor 36, the third flow rate sensor 37, the temperature sensor 38, the first flow rate sensor 51, the second flow rate sensor 52, the third flow rate sensor 53, and the fourth flow rate sensor 54 may also be used as production condition data PCD.

[0242] In the above embodiments, a quality prediction display screen 120 is exemplified as the output method for quality prediction data (QFD), but it is not limited to this. Alternatively, the quality prediction display screen 120 can be used, or the quality prediction data QFD can be printed onto a paper medium or output as a data file.

[0243] Furthermore, as a machine learning model M, there exist machine learning models that use linear regression, Gaussian process regression, support vector regression, decision trees, ensemble methods, bagging, boosting, gradient boosting, and so on. Moreover, there are machine learning models that use simple perceptrons, multilayer perceptrons, deep neural networks, convolutional neural networks, deep belief networks, recurrent neural networks, stochastic neural networks, and so on. There are no particular restrictions on using any of the machine learning models M exemplified above; any machine learning model M using any method can be chosen.

[0244] As an ensemble method, Random Forest is an example. It is well known that Random Forest creates multiple groups of decision trees with low correlation by using randomly sampled learning data and randomly selected explanatory variables, and improves prediction accuracy by ensembling and averaging these predictions. The control parameters for the machine learning model M in this case include the number of selected explanatory variables and the number of branches in the decision trees.

[0245] Because deep neural networks have a large number of control parameters that can be flexibly combined, they can achieve high predictive performance on various data structures. These control parameters include the number of network layers and nodes, the type of activation function, the Dropout ratio, the mini-batch size, the number of epochs, and the learning rate.

[0246] Such a machine learning model M can be implemented using multiple frameworks, from which appropriate options can be selected. For example, it can be selected from Tensorflow, CNTK (Cognitive Toolkit), Theano, Caffe, mxnet, Keras, PyTorch, Chainer, Scikit-learn, Caret, MATLAB (registered trademark), etc.

[0247] The hardware structure of the computer constituting the machine learning system 2 can be varied. For example, the learning device 10 and the application device 11 can be integrated into a single computer. Furthermore, to improve processing power and reliability, at least one of the learning device 10 and the application device 11 can be constituted by multiple computers that are separate hardware components. For example, the functions of the first output unit 76, the learning unit 77, and the extraction unit 78 of the learning device 10 can be distributed among three computers. In this case, the learning device 10 is constituted by three computers.

[0248] In this way, the hardware structure of the computer in the machine learning system 2 can be appropriately modified according to the required performance such as processing power, security, and reliability. Moreover, it is not limited to hardware; for applications such as the first action program 70 and the second action program 110, in order to ensure security and reliability, they can also be dual-stored or distributed across multiple storage devices.

[0249] In the above embodiments, for example, as the hardware structure of the processing unit that performs various processes such as the first RW control unit 75, the first export unit 76, 135, 140, 150, the learning unit 77 (first processing unit 85, evaluation unit 86, update unit 87), the extraction unit 78 (second processing unit 90, third processing unit 91, calculation unit 92, determination unit 93), the transmission control unit 79, the second RW control unit 115, the second export unit 116, the fourth processing unit 117, and the display control unit 118, various processors shown below can be used. Among the various processors, in addition to general-purpose processors such as CPU62A and CPU62B that contain execution software (first action program 70 and second action program 110) and function as various processing units as described above, there are also programmable logic devices (PLDs) and dedicated circuits. The programmable logic device is a processor whose circuit structure can be changed after manufacturing FPGAs (Field Programmable Gate Arrays) and the dedicated circuit is a processor with a circuit structure specifically designed to perform specific processes such as ASICs (Application Specific Integrated Circuits).

[0250] A processing unit can consist of one of these various processors, or it can consist of a combination of two or more processors of the same or different types (e.g., a combination of multiple FPGAs, and / or a combination of a CPU and an FPGA). Furthermore, multiple processing units can also be composed of a single processor.

[0251] As examples of a single processor comprising multiple processing units, firstly, as exemplified by client and server computers, there are processors that consist of a combination of one or more CPUs and software, and that processor functions as multiple processing units. Secondly, as exemplified by System-on-Chip (SOC), there are processors that implement the overall system functionality containing multiple processing units using a single IC (Integrated Circuit) chip. In these cases, various processing units are constructed using one or more of the aforementioned processors as the hardware structure.

[0252] Moreover, the hardware structure of these various processors, more specifically, is a circuit composed of circuit elements such as semiconductor elements.

[0253] Based on the above records, the inventions described in Appendix 1 and Appendix 2 can be understood.

[0254] [Appendix Item 1]

[0255] A learning device comprising:

[0256] The first acquisition processor acquires learning input data, which is provided to a machine learning model for predicting product quality and includes multidimensional property-related data with multiple items derived from multidimensional property data representing the physical properties of the product.

[0257] A temporary learning processor provides the learning input data to the machine learning model for learning and outputs a temporary machine learning model.

[0258] The extraction processor, using the temporary machine learning model, extracts high-contribution items from the multiple items of the multidimensional physical property correlation data, which contribute to improving the accuracy of the quality prediction to a degree that meets pre-defined conditions; and

[0259] The formal learning processor selectively provides the multidimensional physical property-related data of the high-contribution projects to the machine learning model for learning, and outputs the machine learning model as a learned model for practical application.

[0260] [Appendix Item 2]

[0261] An application device comprising:

[0262] The second acquisition processor acquires the learned model output from the formal learning processor of the learning device described in Appendix 1;

[0263] The third acquisition processor acquires multidimensional physical property related data for predicting products with unknown quality.

[0264] The processor provides the learned model acquired by the second acquisition processor with the prediction multidimensional physical property correlation data acquired by the third acquisition processor to predict the quality; and

[0265] The output control processor controls the output of the quality prediction results obtained through the learned model.

[0266] The technology disclosed herein can also be appropriately combined with the various embodiments and modifications described above. Furthermore, it is not limited to the embodiments described above; various structures can be adopted as long as the main idea remains unchanged. Moreover, the technology disclosed herein relates not only to programs but also to storage media that do not temporarily store programs.

[0267] The foregoing descriptions and illustrations are detailed explanations of the parts covered by this disclosure and are merely one example of the technology disclosed herein. For example, the descriptions of the structure, function, role, and effect described above are one example of the structure, function, role, and effect of the parts covered by this disclosure. Therefore, it is natural that unnecessary parts can be deleted and new elements can be added or replaced in the foregoing descriptions and illustrations without departing from the spirit of this disclosure. Furthermore, to avoid complexity and facilitate understanding of the parts covered by this disclosure, explanations related to technical common sense that do not require special explanation have been omitted from the foregoing descriptions and illustrations, provided that the technology disclosed can be implemented.

[0268] In this specification, "A and / or B" is synonymous with "at least one of A and B". That is, "A and / or B" means that it can be only A, only B, or a combination of A and B. Furthermore, in this specification, when more than three cases are expressed in connection with "and / or", the same viewpoint applies as with "A and / or B".

[0269] All documents, patent applications and technical standards described in this specification are incorporated herein by reference to the same extent as those specifically and individually described and incorporated by reference into individual documents, patent applications and technical standards.

[0270] Symbol Explanation

[0271] 2-Machine Learning System, 10-Learning Device, 11-Application Device, 12-Network, 13-Flow Reaction Device, 14, 130-Property Analysis Device, 15-Quality Evaluation Device, 20-First Raw Material Supply Section, 21-Second Raw Material Supply Section, 22, 45-Reaction Zone, 23-Temperature Control Section, 24-Waste Recycling Section, 25-Setting Section, 26-System Controller, 30, 46-Merging Section, 31-Reaction Section, 32-34-First Pipe Section-Third Pipe Section, 35-37-First Flow Rate Sensor-Third Flow Rate Sensor, 38-Temperature Sensor, 40-Recycling Section, 41-Waste Section, 42-Three-Way Valve, 47-50-First Pipe Section-Fourth Pipe Section, 51-54-First Flow Rate Sensor-Fourth Flow Rate Sensor Speed ​​sensor; 60, 60A, 60B - Storage device; 61 - Memory; 62, 62A, 62B - CPU; 63 - Communication unit; 64, 64B - Display; 65, 65B - Input device; 66 - Bus; 70 - First operation program (operation program of learning device); 75 - First read / write control unit (first RW control unit, first acquisition unit); 76, 135, 140, 150 - First export unit (export unit); 77 - Learning unit (temporary learning unit, formal learning unit); 78 - Extraction unit; 79 - Transmission control unit; 85 - First processing unit; 86 - Evaluation unit; 87 - Update unit; 90 - Second processing unit; 91 - Third processing unit; 92 - Calculation unit; 93 - Judgment unit; 100 - Conversion table 110 - Second Action Procedure, 115 - Second Read / Write Control Unit (Second RW Control Unit, Second Acquisition Unit, Third Acquisition Unit), 116 - Second Output Unit, 117 - Fourth Processing Unit, 118 - Display Control Unit (Output Control Unit), 120 - Quality Prediction Display Screen, 121 - OK Button, 125 - Table, 145, 155 - Encoder Network, 146 - Decoder Network, AE - Autoencoder, AR - Region, CI - Contribution Information, CMP - Image Feature Map, DD - Differential Data, DI - Autoencoder Input Data, DIC - Autoencoder Output Data, F - Filter, HCII - High Contribution Item Information, Icp - Pixels in Output Data, IDL - Input for Learning Data, IIMD - Input image data, IIMDL - Input image data for learning, IMD - Image data, INT - Interval, Ip - Pixel of interest, Is - Neighboring pixels, L - Response path length, M - Machine learning model, M1, M2 - First machine learning model, Second machine learning model, ODL - Output data for learning, OIMD - Output image data, OIMDL - Output image data for learning, PCD - Production condition data, PCDF - Production condition data for prediction, PD - Physical property data, PDF - Physical property data for prediction, PM - Temporary machine learning model, POD - Temporary output data, PODE - Temporary output data for extraction, PR - Product, PRD - Physical property related data (related data).PRDE - Extraction-related physical property data (extraction-related data), PRDF - Prediction-related physical property data (prediction-related data), QD - Quality data, QFD - Quality prediction data, RM - Raw material, RM1, RM2 - First raw material, Second raw material, SC - Setting conditions, SP - Spectrum, SPD - Spectral data, SPIMD - Spectral image data, ST100, ST200, ST300, ST500, ST600, ST700, ST800 - Steps, ST1001~ST1006 - Steps (provisional learning steps), ST2001~ST2006 - Steps (extraction steps), ST3001~ST3006 - Steps (formal learning steps), TM - Model after learning, Φ - Reaction path.

Claims

1. A learning device comprising: The first acquisition unit acquires learning input data, which is provided to a machine learning model for predicting the quality of a product, and includes multidimensional property-related data with multiple items derived from multidimensional property data representing the physical properties of the product, and production condition data set in the production process of the product. The temporary learning unit provides the learning input data to the machine learning model for learning and outputs a temporary machine learning model. The extraction unit, using the temporary machine learning model and the production condition data, extracts high-contribution items from the multiple items of the multidimensional physical property correlation data, ensuring that the degree of contribution to improving the accuracy of quality prediction meets pre-set conditions; and The formal learning department selectively provides the production condition data and the multidimensional physical property related data of the high-contribution projects to the machine learning model for learning, and outputs the machine learning model as a learned model for practical application.

2. The learning device according to claim 1, wherein, The multidimensional physical property data includes spectral data detected by spectral analysis of the product.

3. The learning device according to claim 2, wherein, The multidimensional physical property correlation data are representative values ​​of intensity derived for each of the multiple intervals in which the spectral data is divided.

4. The learning device according to claim 1 or 2, wherein, The multidimensional physical property data includes image data obtained by photographing the product.

5. The learning device according to claim 1 or 2, wherein, The product is produced using a flow synthesis method.

6. The learning device according to claim 1 or 2, It has an export section that applies at least a portion of an autoencoder to the multidimensional physical property data to export the multidimensional physical property related data.

7. The learning device according to claim 6, wherein, The automatic encoder learns by being provided with multidimensional physical property data of the product, which indicates a quality level better than a preset grade. The exporting unit provides the multidimensional physical property data to the autoencoder and outputs output data, and exports the multidimensional physical property related data based on the difference data between the multidimensional physical property data provided to the autoencoder and the output data.

8. The learning device according to claim 6, wherein, The exporting unit provides the multidimensional physical property data to the autoencoder, outputs feature data from the encoder network of the autoencoder, and exports the multidimensional physical property related data based on the feature data.

9. The learning device according to claim 6, wherein, The multidimensional physical property data includes image data of the spectrum represented by spectral data detected by spectral analysis of the product.

10. The learning device according to claim 9, wherein, The exporting unit exports the multidimensional physical property correlation data for each of the multiple intervals that divide the spectral data.

11. An application device comprising: The second acquisition unit acquires the completed learning model output by the formal learning unit of the learning device according to any one of claims 1 to 10; The third acquisition section acquires multidimensional physical property related data for predicting the quality of products with unknown characteristics. The processing unit provides the learned model acquired by the second acquisition unit with the multidimensional physical property correlation data for the product with unknown quality acquired by the third acquisition unit, to predict the quality; and The output control unit controls the output of the quality prediction results obtained through the learned model.

12. A method for operating a learning device, comprising: The first acquisition step involves acquiring learning input data, which is provided to a machine learning model for predicting product quality and includes multidimensional property-related data with multiple items derived from multidimensional property data representing the physical properties of the product, and production condition data set in the production process of the product. The temporary learning step involves providing the input data to the machine learning model for learning and outputting a temporary machine learning model. The extraction step utilizes the temporary machine learning model and the production condition data to extract high-contribution items from the multiple items of the multidimensional physical property related data, ensuring that the degree of contribution to improving the accuracy of quality prediction meets pre-set conditions; and In the formal learning process, the production condition data and the multidimensional physical property related data of the high-contribution projects are selectively provided to the machine learning model for learning, and the machine learning model is output as a learned model for practical application.

13. A recording medium containing a program that is an operation program for a learning device, which enables a computer to function as a first acquisition unit, a temporary learning unit, a retrieval unit, and a formal learning unit, wherein... The first acquisition unit acquires learning input data, which is provided to a machine learning model for predicting the quality of a product, and includes multidimensional property-related data with multiple items derived from multidimensional property data representing the physical properties of the product, and production condition data set in the production process of the product. The temporary learning unit provides the learning input data to the machine learning model for learning and outputs a temporary machine learning model. The extraction unit, using the temporary machine learning model and the production condition data, extracts high-contribution items from the multiple items of the multidimensional physical property correlation data, ensuring that the degree of contribution to improving the accuracy of quality prediction meets pre-set conditions; and The formal learning department selectively provides the production condition data and the multidimensional physical property related data of the high-contribution projects to the machine learning model for learning, and outputs the machine learning model as a learned model for practical application.

Images