Resin property value prediction device and resin property value prediction method

Abstract

To accurately predict a resin physical property value when predicting the resin physical property value by using process data whose value changes over time.SOLUTION: A resin physical property value prediction device is the resin physical property value prediction device which predicts a resin physical property value from information about a reaction raw material and time-series process data during the polymerization reaction in a process where the resin polymerization reaction is performed by a batch type reaction tank and includes: a storage unit which stores information on the reaction raw material when manufactured in the past, the time-series process data during the polymerization reaction and a resin physical property value being the corresponding actual measurement data; a pre-processing unit which performs pre-processing on the time-series process data; a model creation unit which creates a learning model using the data subjected to the pre-processing as teacher data; and a calculation unit which calculates the resin physical property value by using a learned model created by the model creation unit on the basis of the blending amount of the raw material and process data in the reaction that is currently performed.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a resin physical property value prediction device and a resin physical property value prediction method. [Background technology]

[0002] Conventionally, there has been a technique for estimating the evaluation value of product quality in a multi-item batch production process in which a variety of products are produced by a similar process using a specified facility (see, for example, Patent Document 1). According to the technique described in Patent Document 1, process variables such as temperature, pressure, tensile speed, oxygen concentration, pH, and humidity are used as explanatory variables, and objective variables such as compressive strength, thickness, component ratio of the substance, temperature, brightness, viscosity, and molecular weight are estimated.

[0003] On the other hand, in a so-called batch plant that produces a wide variety of products in small quantities, the amount of data that can be acquired for each product is small, the plant is an unsteady process in which the state of the plant changes with time, and the product quality is affected by the entire history of the process. In a batch plant with such characteristics, it is difficult to control product quality. In order to perform accurate quality control of product quality, there is a technique for updating parameters of a machine learning model based on the error between predicted data and quality data (see, for example, Patent Document 2). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2022-32337 A [Patent Document 2] Patent Publication No. 2022-167027 Summary of the Invention [Problem to be solved by the invention]

[0005] Here, according to the above-mentioned technology, in a resin polymerization reaction process performed in a batch-type reactor, it may be possible to predict resin physical property values ​​from process data such as the reaction liquid temperature, the heat transfer medium temperature, and the amount of reaction raw material input using a machine learning algorithm. However, these process data have the characteristic that they are values ​​that change over time. Furthermore, the time required for each process varies from batch to batch. Therefore, when a machine learning model is created using the resin physical property values ​​obtained using such process data that varies over time as training data, it is not easy to improve the estimation accuracy of the resin physical property values.

[0006] Therefore, an object of the present invention is to provide a resin property value prediction device and a resin property value prediction method that can accurately predict resin property values ​​when predicting resin property values ​​using process data whose values ​​change over time. [Means for solving the problem]

[0007] (1) One aspect of the present invention is a resin property value prediction device that predicts resin property values ​​from information about reaction raw materials and time-series process data during a polymerization reaction in a process of performing a resin polymerization reaction in a batch-type reaction tank, the device being configured to predict resin property values ​​from information about the reaction raw materials and time-series process data during the ... Including raw material input The apparatus includes a storage unit that stores time-series process data during a polymerization reaction in association with resin physical property values ​​that are corresponding measured data, a pre-processing unit that performs pre-processing on the time-series process data, a model creation unit that creates a learning model using the pre-processed data as teacher data, and a calculation unit that calculates the resin physical property values ​​using the learned model created by the model creation unit based on the blending amounts of raw materials in a reaction currently in progress and the process data. The pre-processing performed by the pre-processing unit is a process of extracting mutual similarities as feature quantities for the plurality of time-series process data, and the model creation unit selects a method that reduces the difference from the measured physical property values ​​from among any one of an algorithm of Lasso regression (Least Absolute Shrinkage and Selection Operator), Ridge regression, or Partial Least Squares Regression (PLS regression), and updates regression parameters within the algorithm during the selection process. This is a resin property value prediction device.

[0009] ( 2 ) Also, one aspect of the present invention is the above ( 1 In the resin property value prediction device described in the above, the similarity is extracted by a DTW (Dynamic Time Warping) method.

[0010] ( 3 ) Also, one aspect of the present invention is the above ( 1 ) or ( 2 In the resin physical property prediction device described in the above, the pre-processing unit smoothes the time-series process data and then extracts the similarity as a feature.

[0012] ( 4 ) Also, one aspect of the present invention is the above (1) to ( 3 ), the resin composition for which the calculation unit calculates the resin physical property values ​​is an acrylic resin, and the resin physical property values ​​calculated by the calculation unit include at least one of viscosity, amount of non-volatile components, and amount of residual monomers.

[0015] ( 5 ) Also, one aspect of the present invention is a process for carrying out a resin polymerization reaction in a batch reaction tank, comprising: Using an information processing device A method for predicting resin physical properties from information about reaction raw materials and time-series process data during a polymerization reaction, comprising: The information processing device, Information regarding the reactants used in previous production; and Including raw material input a storage step of storing time-series process data during the polymerization reaction in association with corresponding measured data of resin physical properties; The information processing device, a pre-processing step of pre-processing the time-series process data; The information processing device, A model creation process for creating a learning model using the preprocessed data as training data; The information processing device, A calculation step of calculating a resin physical property value using the trained model created in the model creation step based on the blending amounts of raw materials in the reaction currently in progress and process data. the pre-processing performed by the pre-processing unit is a process of extracting mutual similarities as feature quantities for the plurality of time-series process data, and the model creation unit selects a method that reduces the difference from the measured physical property values ​​from among any one of an algorithm of Lasso regression (Least Absolute Shrinkage and Selection Operator), Ridge regression, or Partial Least Squares Regression (PLS regression), and updates regression parameters within the algorithm during the selection process. This is a method for predicting resin physical properties. Effect of the Invention

[0016] According to the present invention, it is possible to provide a resin physical property prediction device and a resin physical property prediction method that can accurately predict resin physical property values ​​when predicting resin physical property values ​​using process data whose values ​​change over time. [Brief description of the drawings]

[0017] [Figure 1] FIG. 2 is a functional configuration diagram showing an example of a functional configuration of a resin physical property prediction device according to the first embodiment. [Diagram 2] FIG. 4 is a diagram for explaining an example of sampling timing during a resin polymerization reaction according to the first embodiment. [Diagram 3] FIG. 4 is a diagram showing an example of a change over time in time-series process data according to the first embodiment. [Figure 4] 6 is a diagram showing a change in a dropping speed when the smoothing process according to the first embodiment is applied to the dropping speed. FIG. [Diagram 5] FIG. 4 is a diagram showing an example of conditions for acquiring teacher data in a learning stage of the resin property prediction device according to the first embodiment. [Figure 6] FIG. 4 is a first diagram showing a prediction result of the resin property prediction device according to the first embodiment. [Figure 7] FIG. 2 is a second diagram showing the prediction results of the resin property prediction device according to the first embodiment. [Figure 8] FIG. 11 is a functional configuration diagram showing an example of a functional configuration of a resin physical property prediction device according to a second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] [Embodiment] First, the resin physical property value prediction device and the resin physical property value prediction method according to the present embodiment are used to measure the properties of a resin composition in a resin polymerization process.

[0019] The resin composition targeted by the resin property value prediction device and the resin property value prediction method according to the present embodiment includes a wide range of polymers such as homopolymers and copolymers. The resin composition may be a thermoplastic resin or a thermosetting resin. The thermoplastic resin is not particularly limited, but examples thereof include polypropylene (PP), polyethylene (PE), ABS resin, polyvinyl chloride (PVC), acrylic resin, polyester resin, polystyrene resin (PS), urethane resin (PU), polyphenylene sulfide resin (PPS), and the like. In the following embodiment, as an example, a case will be described in which the resin composition is an acrylic resin and prediction is performed using a solution containing the acrylic resin (hereinafter, the form of the solution containing the resin composition is also referred to as a "resin product").

[0020] The resin property value prediction device and the resin property value prediction method according to the present embodiment are applied to a process in which a resin polymerization reaction is performed using a batch-type reaction tank. The resin property value prediction device and the resin property value prediction method according to the present embodiment predict resin property values ​​from information on reaction raw materials and time-series process data during the polymerization reaction. Examples of reaction raw materials used in a process in which a resin polymerization reaction is performed using a batch-type reaction tank include monomers, initiators, and solvents. Examples of time-series process data during the polymerization reaction include reaction liquid temperature, heat medium temperature, and raw material input amount, pressure, and the like. It is preferable that the time-series process data during the polymerization reaction be data that can be easily obtained from manufacturing equipment.

[0021] The properties of the resin composition to be predicted by the resin property value prediction device and the resin property value prediction method according to the present embodiment may specifically be polymer property values. The polymer property values ​​broadly include the amount of non-volatile components (NV; Non-Volatile), viscosity (Gardner viscosity), and the amount of residual monomers (residual monomer concentration), and it is preferable that at least one of these is included.

[0022] Hereinafter, an example of a method for measuring the resin physical properties according to the present embodiment will be described. Note that the physical properties of the resin composition according to the present embodiment are not limited to the above-mentioned example, and may be measured by a method other than the measurement method described below.

[0023] The method for measuring the amount of non-volatile components will be described. First, the sampled resin composition is placed in an aluminum dish whose weight has been measured in advance, and the weight is measured. Then, the sample is diluted to 20 wt% (weight percent) by adding a dilution solvent containing toluene and methanol, and heated in an oven heated to 108°C for one hour to remove the volatile components. The weight of the resin composition residue remaining on the aluminum dish after heating is measured, and the amount of non-volatile components is calculated from the difference from the weight before heating.

[0024] The method for measuring viscosity will be described below. First, a certain amount of the sampled resin composition is added to a Gardner bubble viscosity tube and sealed. The temperature of the resin composition is adjusted by immersing the sealed viscosity tube in a water bath controlled at 25°C. After that, the bubble rising speed is compared to determine the closest standard viscosity sample, and the value is converted into a numerical value using a standard viscosity conversion table.

[0025] The method for measuring the amount of residual monomers will be described. First, the sampled resin composition is diluted to 0.5 [wt%] using N,N-dimethylformamide to prepare a measurement sample. The amount of monomers in the resin composition is measured using gas chromatography with an FID detector. The amount of residual monomers is measured from the total amount of all monomers.

[0026] [Embodiment 1] Hereinafter, a resin property prediction device 1 according to a first embodiment of the present invention will be described with reference to FIGS. 1 to 9. FIG.

[0027] FIG. 1 is a functional configuration diagram showing an example of the functional configuration of the resin physical property prediction device according to the first embodiment. With reference to the same figure, an example of the functional configuration of the resin physical property prediction device 1 will be described. The resin physical property prediction device 1 includes a storage unit 10, a preprocessing unit 20, a model creation unit 30, an in-progress data acquisition unit 40, and a calculation unit 50. Each of these functional units is realized, for example, by using an electronic circuit. Furthermore, each functional unit may include a storage means such as a semiconductor memory or a magnetic hard disk device inside as necessary. Furthermore, each function may be realized by a computer and software.

[0028] The resin property value prediction device and the resin property value prediction method according to the present embodiment predict resin property values ​​by machine learning. Below, the configuration used in the learning stage and the configuration used in the inference stage will be described. First, the configuration used in the learning stage will be described. In the learning stage, a storage unit 10, a preprocessing unit 20, and a model creation unit 30 are used.

[0029] The storage unit 10 stores performance data of data measured in the past. The performance data is data in which an explanatory variable X and a corresponding objective variable Y are associated with each other. The explanatory variable X is process data used in the production of the resin composition, and specifically, information on the reaction raw materials when the resin composition was produced in the past and time-series process data during the polymerization reaction can be exemplified. The time-series process data during the polymerization reaction can be specifically the reaction liquid temperature, the heat medium temperature, and the amount of raw materials input. The objective variable Y is a resin property value that is actual measurement data measured from a sample sampled when the resin composition is produced using the corresponding process data, and specifically, the amount of non-volatile components, viscosity, and residual monomers can be exemplified.

[0030] The preprocessing unit 20 performs preprocessing on data input to the machine learning algorithm. The preprocessing is performed both in the learning stage and in the inference stage. In the learning stage, the preprocessing unit 20 may perform preprocessing on both the explanatory variable X and the target variable Y. The preprocessing according to this embodiment is applied to at least the time-series process data included in the explanatory variable X. In addition, the preprocessing according to this embodiment includes at least a synchronization process. Here, the time required for manufacturing the time-series process data may differ for each batch process. In addition, the sampling time, reaction time, raw material dropping time, etc. may also differ for each batch process. The synchronization process is a process for synchronizing the time-series process data by performing dynamic time warping.

[0031] An example of the synchronization process may be a process of extracting similarity between multiple pieces of time-series process data as a feature amount. In other words, the pre-processing performed by the pre-processing unit 20 may be a process of extracting mutual similarity between multiple pieces of time-series process data as a feature amount. In this case, the extracted similarity is passed to the model creation unit 30. The similarity may be between the multiple pieces of time-series process data or may be similarity with a predetermined sample data (golden sample).

[0032] More specifically, the similarity extracted by the synchronization process may be extracted by a dynamic time warping method (DTW method). The dynamic time warping method is a method used to measure the similarity between time series data, and a detailed description thereof will be omitted.

[0033] The pre-processing according to the present embodiment may include a smoothing process. The smoothing process is applied to the time-series process data included in the explanatory variable X among the data input to the machine learning algorithm. Here, the dripping speed of the raw material among the time-series process data is, for example, a value calculated based on a value measured by an operator using a balance to measure the dripping amount every 5 minutes and visually reading it. Therefore, the dripping speed of the raw material may be obtained by manual work by an operator, and may include an error. In addition, the reaction solution temperature may include a small fluctuation. The smoothing process is performed to reduce the influence on the measurement result based on such errors and fluctuations. The smoothing process may be performed by, for example, calculating a weighted moving average.

[0034] In this embodiment, the smoothing process does not necessarily have to be performed. For example, it may be better to perform the smoothing process on the non-volatile component amount and the residual monomer amount among the explanatory variables X, but it may be better not to perform the smoothing process on the viscosity. Therefore, the pre-processing unit 20 may perform the smoothing process on a predetermined explanatory variable X. When performing the smoothing process, the pre-processing unit 20 may extract the similarity as a feature amount after smoothing the time-series process data.

[0035] Moreover, the pre-processing according to the present embodiment may include a logarithmic transformation process. The logarithmic transformation may be performed on, for example, the objective variable Y. Here, depending on the type of the objective variable Y, it may be better to perform the logarithmic transformation. On the other hand, by performing the logarithmic transformation process on the resin physical property value, the prediction error of the resin physical property value may increase. That is, when the logarithmic transformation process on the resin physical property value is performed, the accuracy may be improved or not. Therefore, whether or not the logarithmic transformation process is applied may be determined according to the type of the physical property value.

[0036] The model creation unit 30 creates a learning model using as teacher data the data that has been preprocessed on the data stored in the storage unit 10. In other words, it can also be said that the model creation unit 30 creates a learning model by supervised learning.

[0037] The model creation unit 30 may create a learning model using a method such as Lasso regression (Least Absolute Shrinkage and Selection Operator), Ridge regression, or Partial Least Squares Regression (PLS regression). It is preferable that the model creation unit 30 creates a learning model using any suitable method among Lasso regression, Ridge regression, and Partial Least Squares Regression. Specifically, the model creation unit 30 may select a method that reduces the difference from the measured physical property value from any of the algorithms of Lasso regression, Ridge regression, and Partial Least Squares Regression, and update the regression parameters in the algorithm during the selection process.

[0038] Next, the configuration used in the inference stage will be described. In the inference stage, the learning model created by the model creation unit 30 in the learning stage, the progress data acquisition unit 40, and the calculation unit 50 are used.

[0039] The in-progress data acquisition unit 40 acquires process data of the currently in-progress manufacturing process in the inference stage. Examples of the process data acquired by the in-progress data acquisition unit 40 include information on the reaction raw materials and time-series process data during the polymerization reaction. Specifically, the time-series process data during the polymerization reaction may include the reaction liquid temperature, the heat medium temperature, and the amount of raw materials input.

[0040] The calculation unit 50 calculates the resin physical property values ​​based on the blending amounts of the raw materials in the reaction currently in progress and the process data, using the trained model created by the model creation unit 30. The resin composition for which the calculation unit 50 calculates the resin physical property values ​​may specifically be an acrylic resin.

[0041] 2 is a diagram for explaining an example of sampling timing during the resin polymerization reaction according to embodiment 1. With reference to the drawing, an example of sampling timing in the resin polymerization reaction step according to embodiment 1 will be explained. The drawing shows the change in temperature of the polymerization reaction tank over time, with the temperature of the polymerization reaction tank on the vertical axis and time on the horizontal axis.

[0042] The resin polymerization reaction is carried out using a polymerization reaction tank. The polymerization reaction tank is, for example, a reaction vessel (not shown), and the temperature of the sample stored in the reaction vessel is controlled by injecting steam or the like into the reaction vessel. In this case, the temperature of the polymerization reaction tank refers to the temperature of the reaction vessel itself or the temperature inside the reaction vessel. The temperature of the sample stored in the reaction vessel may also be controlled by submerging the reaction vessel in a predetermined temperature-controlled oil bath.

[0043] First, at time t11, the temperature inside the reaction vessel is increased. At time t12, the temperature reaches a predetermined reaction temperature T.

[0044] From time t12 to time t13, a dripping step is performed in which the monomer and the polymerization initiator are dripped. Since this polymerization reaction is an exothermic reaction, the temperature of the sample stored in the reaction vessel is maintained by introducing cooling water or lowering the oil bath temperature as necessary. The dripping step may take, for example, 5 hours. The time required for the dripping step may vary from batch to batch. During the dripping step, for example, one sampling is performed. In the illustrated example, the first sampling is performed once during the dripping step.

[0045] The monomer used in producing the acrylic resin is not particularly limited, and any known monomer can be used. Examples of the monomer include acrylic acid, alkyl acrylate, methacrylic acid, and alkyl methacrylate.

[0046] The polymerization initiator used in producing the acrylic resin is not particularly limited, and any known initiator can be used. Examples of the polymerization initiator that can be used include persulfates such as ammonium persulfate and potassium persulfate, hydrogen peroxide, azo compounds, and organic oxides such as (2-ethylhexanoyl)(tert-butyl)peroxide.

[0047] In the dropping step, radicals are generated in the initiation reaction by thermal decomposition of the polymerization initiator, etc. In the propagation reaction, the radicals generated in the initiation reaction react sequentially with monomers to generate polymer radicals, and the polymer radicals further react with the monomers to generate an acrylic resin as a resin composition.

[0048] After the dripping is completed, a holding step is performed from time t13 to time t14, in which the temperature is maintained at a predetermined temperature. This holding step is a period during which the remaining monomer and the polymerization initiator are consumed. The holding step may last, for example, four hours. The time required for the holding step may vary from batch to batch. During the holding step, sampling is performed multiple times, for example, every hour. In the illustrated example, sampling is performed a total of six times during the holding step. By performing multiple samplings during this holding step, the viscosity of the resin composition can be measured, for example.

[0049] After the holding step, a dilution and cooling step is carried out from time t14 to time t15, and the temperature inside the reaction vessel is cooled. By sampling after this dilution and cooling step, for example, the NV value of the resin composition is measured to confirm whether the product meets the specifications.

[0050] Fig. 3 is a diagram showing an example of the change over time of the time-series process data according to the first embodiment. With reference to the figure, an example of the change over time of the time-series process data obtained in the resin polymerization reaction described in Fig. 2 will be described. In the figure, the horizontal axis represents time, and the change over time of each time-series process data is shown. The vertical axis represents temperature [°C] for the reaction solution temperature, weight [g] for the cumulative amount of monomer-initiator mixture dripped, and weight per minute [g / min] for the amount of monomer-initiator mixture dripped.

[0051] Times t11 to t15 shown in the figure correspond to those in Figure 2. That is, from time t11 to time t12 is the process of increasing the temperature in the reaction vessel, from time t12 to time t13 is the dropping process, from time t13 to time t14 is the hold process, and from time t14 to time t14 is the dilution and cooling process. The amount of monomer-initiator mixture dropped and the cumulative amount of monomer-initiator mixture dropped are time-series process data that change only in the dropping process.

[0052] Here, the amount of monomer-initiator mixture dripped is the average value of the amount dripped every 5 minutes measured by an operator on a balance. The drip amount is measured by the operator visually reading the scale. Therefore, the amount of monomer-initiator mixture dripped may include a time error due to the measurement every 5 minutes and a measurement error due to the operator visually reading the scale. Therefore, in this embodiment, a smoothing process is performed by performing preprocessing by the preprocessing unit 20.

[0053] Here, the sampling timing is indicated by a circle in the figure. In the example shown in the figure, sampling is performed seven times at irregular intervals. The sampling timing differs for each batch. In addition, the reaction time and the raw material dropping time also differ for each batch. Therefore, in this embodiment, the time-series process data is synchronized by performing preprocessing by the preprocessing unit 20.

[0054] FIG. 4 is a diagram showing a change when the smoothing process according to the first embodiment is applied to the dripping speed. An example of the smoothing process will be described with reference to the figure. The smoothing process is performed to remove the error factors (time error based on the measurement cycle and measurement error caused by the operator performing the measurement) as described above. The error factors can be reduced by the smoothing process. Both FIG. 4(A) and FIG. 4(B) show the change in dripping speed with the horizontal axis representing time [min] and the vertical axis representing the dripping amount per minute [g / min]. FIG. 4(A) shows the dripping amount per minute [g / min] before the smoothing process is performed, and FIG. 4(B) shows the dripping amount per minute [g / min] after the smoothing process is performed.

[0055] The smoothing process may be performed by calculating a moving average. The moving average may be, for example, a weighted moving average. Note that the smoothing process may use any method capable of reducing error factors, such as a simple moving average. The smoothing process may also be called a smoothing process. As shown in the figure, it can be seen that the time series data is smoothed before and after the smoothing process.

[0056] Next, an example of a synchronization process of the time-series process data will be described.

[0057] The time-series process data described with reference to FIG. 3 differs for each batch (lot). Each batch is operated to meet the target temperature and dripping speed, but the process time differs slightly. Synchronization processing adjusts the process data with different lengths in this way. For example, the DTW method (dynamic time warping method) may be used for synchronization processing.

[0058] The DTW method (Dynamic Time Warping) can calculate the similarity between different batches of time-series process data. If two pieces of time-series process data are identical, the similarity is 0, and the more different the shapes of the time-series process data are, the greater the similarity becomes.

[0059] In actual processing, the similarity between the time-series process data is calculated in a brute-force manner. The model creation unit 30 uses the similarity calculated from the time-series process data in this way as part of the explanatory variable X to perform learning and inference.

[0060] Next, with reference to Figs. 5 to 7, the results of learning and inference (prediction) of resin physical property values ​​using the resin physical property value prediction device will be described.

[0061] 5 is a diagram showing an example of conditions for acquiring teacher data in the learning stage of the resin property prediction device according to the first embodiment. First, an example of conditions for acquiring teacher data in the learning stage of the resin property prediction device 1 will be described with reference to the same figure. In the illustrated example, data for eight lots, Run1 to Run8, are acquired as basic conditions. Specifically, in the basic conditions, the reaction is carried out while maintaining a predetermined temperature T in both the dropping process and the holding process. The dropping process is for 5 hours, and the holding process is for 4 hours.

[0062] Data for four lots, Run9 to Run12, are acquired under conditions where the temperature is varied up and down. Run9 is a condition where the reaction liquid temperature is increased, specifically, the temperature is increased by 5°C from the predetermined temperature T in both the dropping step and the hold step. Run10 to Run12 are a condition where the reaction liquid temperature is decreased. Specifically, in Run10, the temperature is decreased by 5°C from the predetermined temperature T in both the dropping step and the hold step. In Run11 and Run12, the temperature is decreased by 10°C from the predetermined temperature T in both the dropping step and the hold step. The resin physical property value prediction device 1 performs learning using the data acquired under such basic conditions and changed conditions as training data.

[0063] Fig. 6 is a first diagram showing prediction results by the resin physical property prediction device according to embodiment 1. Fig. 7 is a second diagram showing prediction results by the resin physical property prediction device according to embodiment 1. In Figs. 6 and 7, the horizontal axis indicates actual measured values ​​and the vertical axis indicates predicted values ​​by the resin physical property prediction device 1, and the error and the coefficient of determination R 2The machine learning algorithm used for calculating the NV value and viscosity is lasso, and the machine learning algorithm used for calculating the residual monomer concentration is ElasticNet. Note that the machine learning algorithms may be configured so that the model creation unit 30 appropriately selects an appropriate one.

[0064] Figures 6(A) to 6(C) show an example of inference using a learning model trained based only on the basic conditions described above. Figures 6(D) to 6(F) show an example of inference using a learning model trained based on the basic conditions described above and also on modified conditions that include temperature.

[0065] Figures 6(A) and 6(D) show the results of the NV value, Figures 6(B) and 6(E) show the results of the residual monomer concentration, and Figures 6(C) and 6(F) show the results of the viscosity. When estimating the NV value in Figures 6(A) and 6(D), smoothing was performed, but logarithmic transformation was not performed. In addition, Figures 6(B), 6(C), 6(E), and 6(F) were both smoothed and logarithmic transformed.

[0066] Looking at the NV values ​​shown in Figure 6(A), the average error rate is ±0.83% and the coefficient of determination R 2 The NV value shown in Figure 6(D) shows that the average error rate is ±1.33% and the coefficient of determination R 2 Looking at the residual monomer concentration shown in Figure 6(B), the average error rate is ±7.4% and the coefficient of determination R 2 The residual monomer concentration shown in Figure 6(E) shows that the average error rate is ±8.6% and the coefficient of determination R 2 Looking at the viscosity shown in Figure 6(C), the average error rate is ±15.4% and the coefficient of determination R 2 Looking at the viscosity shown in Figure 6(F), the average error rate is ±18.6% and the coefficient of determination R 2 is 0.83.

[0067] FIG. 7 shows the results of an experiment to see the change in prediction accuracy when smoothing is performed or not. FIG. 7(A) and FIG. 7(D) show the results of the NV value, FIG. 7(B) and FIG. 7(E) show the results of the residual monomer concentration, and FIG. 7(C) and FIG. 7(F) show the results of the viscosity. FIG. 7(A) to FIG. 7(C) are diagrams showing the results when smoothing is performed. FIG. 7(D) to FIG. 7(F) are diagrams showing the results when smoothing is not performed. FIG. 7(A) to FIG. 7(C) use the same results as FIG. 6(A) to FIG. 6(C). When estimating FIG. 7(A) and FIG. 7(D), i.e., the NV value, logarithmic transformation was not performed. Also, FIG. 7(B), FIG. 7(C), FIG. 7(E), and FIG. 7(F) all performed logarithmic transformation.

[0068] Looking at the NV values ​​shown in Figure 7(D), the average error rate is ±1.00% and the coefficient of determination R 2 is 0.99. Therefore, it can be said that the prediction accuracy of the NV value in Fig. 7(A), which has been smoothed, is higher than that in Fig. 7(D), which has not been smoothed. Looking at the residual monomer concentration in Fig. 7(E), the average error rate is ±13.3[%], and the coefficient of determination R 2 is 0.87. Therefore, it can be said that the prediction accuracy of the residual monomer concentration in Fig. 7(B), which has been smoothed, is higher than that in Fig. 7(E), which has not been smoothed. Looking at the viscosity shown in Fig. 7(F), the average error rate is ±12.8% and the coefficient of determination R 2 is 0.95. Therefore, it can be said that the prediction accuracy of the viscosity in FIG. 7(F), where no smoothing processing was performed, is higher than that in FIG. 7(C), where smoothing processing was performed.

[0069] That is, the prediction accuracy of the NV value and the residual monomer concentration is higher when the smoothing process is performed. Also, the prediction accuracy of the viscosity is higher when the smoothing process is not performed. Therefore, whether or not to perform the pretreatment by the pretreatment unit 20 may be determined depending on the type of the resin property value to be measured.

[0070] [Summary of embodiment 1] According to the embodiment described above, the resin property value prediction device 1 predicts the resin property value from information on the reacting raw materials and time-series process data during the polymerization reaction in a process of performing a resin polymerization reaction in a batch-type reaction tank. The resin property value prediction device 1 includes a storage unit 10, and stores information on the reacting raw materials and time-series process data during the polymerization reaction in association with the resin property value, which is the corresponding measured data. The resin property value prediction device 1 also includes a pre-processing unit 20, and performs pre-processing on the time-series process data. The pre-processing includes at least a synchronization process of the time-series process data. The resin property value prediction device 1 also includes a model creation unit 30, and creates a learning model using the pre-processed data as teacher data. The resin property value prediction device 1 also includes a calculation unit 50, and calculates the resin property value using the learned model created by the model creation unit 30 based on the blending amounts of raw materials in the reaction currently in progress and the process data. That is, according to this embodiment, after the synchronization process of the time-series process data, learning and inference are performed. Therefore, according to the resin physical property value prediction device 1, when predicting a resin physical property value using process data whose values ​​change over time, the resin physical property value can be predicted with high accuracy.

[0071] According to the embodiment described above, the pre-processing performed by the pre-processing unit 20 is a process of extracting mutual similarities between a plurality of pieces of time-series process data as feature quantities. According to the embodiment described above, the similarities are extracted by the DTW method. That is, according to the resin physical property value prediction device 1, since the time-series process data is learned based on the similarities, even if the sizes of the plurality of process data are different, the resin physical property value can be predicted with high accuracy.

[0072] According to the embodiment described above, the pre-processing unit 20 extracts the similarity as a feature value after smoothing the time-series process data. Therefore, according to the embodiment, even if the time-series process data includes an error, the resin physical property value can be predicted with high accuracy.

[0073] According to the embodiment described above, the model creation unit 30 selects a method that reduces the difference from the measured physical property value from among any of the algorithms of Lasso regression, Ridge regression, and partial least squares regression, and updates the regression parameters in the algorithm during the selection process. Therefore, according to the resin physical property value prediction device 1, inference can be performed using a learning model created using a more suitable machine learning algorithm. Therefore, according to the resin physical property value prediction device 1, when predicting the resin physical property value using process data whose value changes over time, the resin physical property value can be predicted with high accuracy.

[0074] According to the embodiment described above, the resin composition for which the calculation unit 50 calculates the resin physical property values ​​is an acrylic resin, and the resin physical property values ​​calculated by the calculation unit include at least one of the viscosity, the amount of non-volatile components, and the amount of remaining monomers. Therefore, according to the embodiment, when predicting the resin physical property values ​​such as the viscosity, the amount of non-volatile components, and the amount of remaining monomers using process data whose values ​​change over time, the resin physical property values ​​can be predicted with high accuracy.

[0075] [Embodiment 2] Next, a second embodiment will be described with reference to FIG. 8. In the second embodiment, a future (e.g., 30 minutes later) physical property value is predicted. That is, the second embodiment differs from the first embodiment in that a future physical property value is predicted from a resin polymerization reaction currently in progress. Also, in the second embodiment, a plurality of objective variables Y are calculated when the parameters of the explanatory variable X are varied, and it is determined what value of the explanatory variable X should be used to obtain a more suitable physical property value, and feedback may be provided to use suitable parameters.

[0076] 8 is a functional configuration diagram showing an example of the functional configuration of the resin property value prediction device according to the second embodiment. With reference to the same figure, an example of the functional configuration of the resin property value prediction device 1A according to the second embodiment will be described. The resin property value prediction device 1A differs from the resin property value prediction device 1 in that it further includes a candidate generation unit 60, a future property value calculation unit 70, and a process data determination unit 80. In the description of the resin property value prediction device 1A, the same components as those of the resin property value prediction device 1 are denoted by the same reference numerals, and description thereof may be omitted.

[0077] The candidate generating unit 60 generates process data for the current time and after in the resin polymerization reaction in progress. The process data for the current time and after generated by the candidate generating unit 60 is a candidate for future process data. Therefore, the process data for the current time and after generated by the candidate generating unit 60 can also be called future process data candidates. The candidate generating unit 60 can also generate multiple future process data candidates. The multiple future process data candidates generated by the candidate generating unit 60 are preprocessed by the preprocessing unit 20, and inference is performed using the learning model created by the model creating unit 30.

[0078] The future physical property value calculation unit 70 calculates the resin physical property value by inputting the blending amount of the raw materials and the process data generated by the candidate generation unit 60 into the learning model created by the model creation unit. In other words, it can be said that the resin physical property value calculated by the future physical property value calculation unit 70 is the resin physical property value when a resin polymerization reaction is performed using the future process data candidate. When a plurality of future process data candidates are generated by the candidate generation unit 60, the future physical property value calculation unit 70 calculates a plurality of resin physical property values. In this case, the candidate generation unit 60 may further generate future process data candidates based on the plurality of resin physical property values ​​calculated by the future physical property value calculation unit 70. The future process data candidates generated based on the plurality of resin physical property values ​​calculated by the future physical property value calculation unit 70 may be those that can obtain more suitable resin physical property values.

[0079] Moreover, the resin physical property value prediction device 1A further includes a process data determination unit 80, which determines whether or not a resin physical property value determined as a target value can be realized. Specifically, the process data determination unit 80 determines whether or not a resin physical property value determined as a target value can be realized based on a change in the physical property value calculated based on the future process data candidates generated by the candidate generation unit 60. The target value may be determined in advance and stored in the storage unit 10, or may be calculated based on the current resin physical property value. Furthermore, the candidate generation unit 60 may further raise or lower the process data based on a change in the physical property value obtained as a result of raising or lowering the process data.

[0080] [Summary of embodiment 2] According to the embodiment described above, the resin physical property value prediction device 1A includes the candidate generation unit 60, and generates process data from the current time onward in a resin polymerization reaction that is in progress. In addition, the resin physical property value prediction device 1A includes the future physical property value calculation unit 70, and calculates a resin physical property value by inputting the blending amounts of raw materials and the process data generated by the candidate generation unit 60 into a learning model created by the model creation unit 30. By employing such a configuration, the resin physical property value prediction device 1A can predict future (e.g., 30 minutes from now) physical property values.

[0081] According to the embodiment described above, the future physical property value calculation unit 70 calculates a plurality of resin physical property values, and the candidate generation unit 60 generates future process data candidates, which are candidates for future process data, based on the plurality of resin physical property values ​​calculated by the future physical property value calculation unit 70. The resin physical property value prediction device 1A further includes a process data determination unit 80, and determines whether or not a resin physical property value set as a target value can be realized based on the transition of the physical property value calculated based on the future process data candidates generated by the candidate generation unit 60. Therefore, according to the present embodiment, it is possible to swing the future process data candidates up and down and select a more suitable candidate, thereby achieving a target resin physical property value.

[0082] In addition, all or part of the functions of each unit of the resin property prediction device 1 in the above-mentioned embodiment may be realized by recording a program for realizing these functions in a computer-readable recording medium, reading the program recorded in the recording medium into a computer system, and executing it. Note that the "computer system" here includes hardware such as an OS and peripheral devices.

[0083] In addition, "computer-readable recording medium" refers to portable media such as magneto-optical disks, ROMs, and CD-ROMs, and storage units such as hard disks built into computer systems. Furthermore, "computer-readable recording medium" may also include those that dynamically hold a program for a short period of time, such as a communication line when transmitting a program over a network such as the Internet, and those that hold a program for a certain period of time, such as volatile memory inside a computer system that serves as a server or client in such a case. Furthermore, the above program may be one that realizes part of the above-mentioned functions, or may be one that can realize the above-mentioned functions in combination with a program already recorded in the computer system.

[0084] The configurations and combinations thereof in each embodiment are merely examples, and additions, omissions, substitutions, and other modifications of the configurations are possible without departing from the spirit of the present invention. Furthermore, the present invention is not limited to each embodiment, but is limited only by the scope of the claims. [Explanation of symbols]

[0085] 1... resin physical property value prediction device, 10... storage unit, 20... pre-processing unit, 30... model creation unit, 40... ongoing data acquisition unit, 50... calculation unit, 60... candidate generation unit, 70... future physical property value calculation unit, 80... process data determination unit

Claims

1. A resin physical property value prediction device for predicting resin physical property values ​​from information on reaction raw materials and time-series process data during a resin polymerization reaction in a batch-type reaction tank, comprising: a storage unit that stores time-series process data during a polymerization reaction, including information on the reactant materials and amounts of the reactants input, in association with corresponding measured data of resin physical properties; a pre-processing unit that performs pre-processing on the time-series process data; A model creation unit that creates a learning model using the preprocessed data as training data; A calculation unit that calculates a resin physical property value using the trained model created by the model creation unit based on the blending amounts of raw materials in the reaction currently in progress and process data; Equipped with The pre-processing performed by the pre-processing unit is a process of extracting a mutual similarity as a feature amount for a plurality of the time-series process data, The model creation unit selects a method that reduces a difference from an actually measured physical property value from among any one of an algorithm of Lasso regression (Least Absolute Shrinkage and Selection Operator), Ridge regression, or Partial Least Squares Regression (PLS regression), and updates a regression parameter in the algorithm during the selection process. Resin property value prediction device.

2. The similarity is extracted by a Dynamic Time Warping (DTW) method. The resin physical property value prediction device according to claim 1 .

3. the pre-processing unit smoothes the time-series process data and then extracts the similarity as a feature. The resin physical property value prediction device according to claim 1 or 2.

4. The resin composition for which the calculation unit calculates the resin physical property value is an acrylic resin, The resin physical property value calculated by the calculation unit includes at least one of viscosity, non-volatile component amount, and residual monomer amount. The resin physical property value prediction device according to claim 1 or 2.

5. A method for predicting resin physical property values ​​from information on reaction raw materials and time-series process data during a polymerization reaction in a process of performing a resin polymerization reaction in a batch reaction tank, comprising: a storage step in which the information processing device stores time-series process data during the polymerization reaction, including information on the reactant raw materials and amounts of the raw materials input, in association with corresponding measured data of resin physical properties; a pre-processing step in which the information processing device performs pre-processing on the time-series process data; a model creation process in which the information processing device creates a learning model using the preprocessed data as training data; A calculation step in which the information processing device calculates a resin physical property value using the trained model created by the model creation step based on the blending amounts of raw materials in the reaction currently in progress and process data; having The pre-processing step is a process of extracting a similarity between the plurality of pieces of time-series process data as a feature amount, The model creation step includes selecting a method that reduces the difference from the measured physical property values ​​from among any one of Lasso regression (Least Absolute Shrinkage and Selection Operator), Ridge regression, or Partial Least Squares Regression (PLS regression), and updating the regression parameters in the algorithm during the selection process. A method for predicting resin physical properties.

Images