Resin property value prediction device and resin property value prediction method

Abstract

To accurately calculate physical properties of a resin composition in short time.SOLUTION: A resin physical property value prediction device includes: an acquisition part for acquiring spectral data obtained by measuring a resin composition by a spectral sensor, in a process of performing a resin polymerization reaction by a batch type reaction tank; and a calculation part for calculating a resin physical property value of the resin composition having the spectral data acquired by the acquisition part, using the spectral data obtained by previously measuring the resin composition, and a learned model obtained by learning a resin physical property value as teacher data.SELECTED DRAWING: Figure 7

Description

[Technical field]

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

[0002] 2. Description of the Related Art Conventionally, in a polymerization reaction process using a batch reactor, a technique for controlling the process based on measured process data in order to maintain uniform quality of the product has been known (see, for example, Patent Document 1).

[0003] Furthermore, in a polymer production process, in order to reduce the time required for the entire process including sampling and inspection, a technique for calculating the quality of a polymer based on the Raman spectrum of the polymer is known (see, for example, Patent Document 2). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2001-106703 A [Patent Document 2] Special Publication No. 2021-521463 Summary of the Invention [Problem to be solved by the invention]

[0005] According to the process control method disclosed in Patent Document 1, sampling and physical property testing are required, and there is a problem in that it takes a long time to control the process. In addition, when using a technology such as that disclosed in Patent Document 2, the time may be shortened because a machine learning algorithm is used, but the prediction accuracy is about ±20%, and high-precision prediction is not possible. If the physical property values ​​of the resin composition do not satisfy the quality threshold due to the prediction accuracy, corrective measures such as reanalysis and readjustment of polymerization conditions are required. Corrective measures require cost and labor, and have the problem of causing a decrease in productivity, and in some cases, the polymerized product must be discarded. That is, according to conventional techniques, calculating the physical properties of a resin composition requires a lot of effort and time, and even when machine learning is used, there is a problem that the prediction accuracy is low.

[0006] Therefore, an object of the present invention is to provide a technique capable of calculating the physical properties of a resin composition with high accuracy in a short period of time. [Means for solving the problem]

[0007] As a result of intensive research aimed at solving the above problems, the inventors discovered that a machine learning algorithm can be used to calculate the physical properties of a resin composition with high accuracy in a short period of time, and thus completed the present invention.

[0008] That is, the present invention has the following aspects.

[0009] A resin physical property value prediction device according to one aspect of the present invention includes an acquisition unit that acquires spectral data obtained as a result of measuring a resin composition with a spectroscopic sensor in a process of performing a resin polymerization reaction in a batch-type reaction tank, and a calculation unit that calculates the resin physical property value of the resin composition having the spectral data acquired by the acquisition unit using a trained model trained with the spectral data of the resin composition in advance and the resin physical property value as teacher data. The calculation unit further includes a pre-processing unit that performs a smoothing process on the spectrum data acquired by the acquisition unit, and the pre-processing unit performs the smoothing process by local regression with a local width of less than 41. .

[0010] In the resin physical property prediction device according to one aspect of the present invention, the spectral data included in the teaching data is spectral data obtained by smoothing spectral data previously measured by a spectroscopic sensor. ,beforeResin physical property values ​​are calculated based on the spectrum data that has been smoothed by the pre-processing unit.

[0011] In addition, in a resin physical property prediction device according to one embodiment of the present invention, the spectral data included in the teacher data is spectral data obtained by differentially processing spectral data previously measured by a spectroscopic sensor, the pre-processing unit further performs differential processing of the spectral data acquired by the acquisition unit, and the calculation unit calculates the resin physical property value based on the spectral data that has been differentially processed by the pre-processing unit.

[0012] In addition, in a resin physical property prediction device according to one embodiment of the present invention, the pre-processing unit separately and independently performs the smoothing process on the spectral data acquired by the acquisition unit and the differential process on the spectral data acquired by the acquisition unit, and the calculation unit calculates resin physical properties based on either the spectral data on which the smoothing process has been performed by the pre-processing unit, the spectral data on which the differential process has been performed by the pre-processing unit, or the spectral data on which neither the smoothing process nor the differential process has been performed by the pre-processing unit.

[0013] In addition, in a resin physical property prediction device according to one embodiment of the present invention, the calculation unit performs a logarithmic transformation process on the resin physical property value separately and independently from the smoothing process and the differentiation process, and outputs either the resin physical property value that has been subjected to the logarithmic transformation process or the resin physical property value that has not been subjected to the logarithmic transformation process.

[0015] In addition, in the resin property prediction device according to one aspect of the present invention, the acquisition unit acquires spectral data measured after the sampled resin composition is allowed to cool for a predetermined period of time.

[0016] In addition, in the resin property prediction device according to one aspect of the present invention, the acquisition unit acquires spectral data measured at predetermined time intervals while a resin polymerization reaction is being carried out in the batch reaction tank.

[0017] In addition, in a resin property prediction device according to one aspect of the present invention, the resin composition is an acrylic resin, and the acquisition unit acquires spectral data obtained as a result of measuring the acrylic resin with a spectroscopic sensor.

[0018] In addition, in the resin property prediction device according to one aspect of the present invention, the spectroscopic sensor that measures the spectral data acquired by the acquisition unit is at least one of a near-infrared (NIR) spectroscopic sensor and a Raman spectroscopic sensor.

[0021] A method for predicting a resin physical property value according to one aspect of the present invention includes an acquisition step of acquiring spectral data obtained by measuring a sampled resin composition by a spectroscopic sensor in a process of performing a resin polymerization reaction in a batch-type reaction tank, and a calculation step of calculating the resin physical property value of the resin composition having the spectral data acquired by the acquisition step using a trained model trained using the spectral data obtained by measuring the resin composition in advance and the resin physical property value as training data. The calculation step further includes a pre-processing step of performing a smoothing process on the spectrum data acquired by the acquisition step, and the pre-processing step performs the smoothing process by local regression with a local width of less than 41. . Effect of the Invention

[0022] According to the present invention, a technique can be provided that can calculate the physical properties of a resin composition with high accuracy in a short period of time. [Brief description of the drawings]

[0023] [Figure 1] FIG. 2 is a diagram for explaining a resin property value prediction system according to the present embodiment. [Diagram 2] FIG. 2 is a diagram showing an NIR spectrum as an example of spectral data according to the present embodiment. [Diagram 3] FIG. 2 is a diagram showing a Raman spectrum as an example of spectrum data according to the embodiment. [Figure 4] FIG. 2 is a diagram for explaining online measurement according to the present embodiment. [Diagram 5] FIG. 2 is a functional configuration diagram for explaining an example of a learning stage according to the embodiment. [Figure 6] FIG. 13 is a diagram for explaining an example of a polymerization condition for creating teacher data according to the present embodiment. [Figure 7] FIG. 2 is a functional configuration diagram showing an example of a functional configuration of a resin physical property prediction device at a prediction stage according to the present embodiment. [Figure 8] FIG. 4 is a functional configuration diagram showing an example of the functional configuration of a calculation unit according to the present embodiment. [Figure 9] 10A to 10C are diagrams for explaining an example of a smoothing process according to the embodiment; [Figure 10] FIG. 2 is a diagram showing a schematic example of a prediction result of the resin property value prediction device according to the embodiment; [Figure 11] FIG. 2 is a diagram for explaining a first example of a prediction result of the resin property prediction device according to the present embodiment. [Figure 12] FIG. 11 is a diagram for explaining a second example of a prediction result of the resin property prediction device according to the embodiment. [Figure 13] FIG. 11 is a diagram for explaining a third example of a prediction result of the resin property value prediction device according to the embodiment. [Figure 14] FIG. 13 is a diagram for explaining a fourth example of a prediction result of the resin property value prediction device according to the present embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0024] [Outline of resin property prediction system] Hereinafter, a resin physical property prediction system 1 according to an embodiment of the present invention will be described with reference to the drawings. First, the prerequisites for the resin physical property prediction system 1 according to the present embodiment will be described. The resin physical property prediction system 1 is used to measure the properties of a resin composition in a resin polymerization process.

[0025] The resin composition according to the present embodiment broadly includes 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").

[0026] The physical properties of the resin composition according to the present embodiment broadly include NV (Non-Volatile) value, viscosity, residual monomer concentration, molecular weight, etc. In the following description, the physical properties of the resin composition may also be described as resin physical property values ​​or polymer liquid physical properties. An example of a method for measuring the resin physical property values ​​according to the present embodiment will be described below. 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.

[0027] The method for measuring the NV value (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.

[0028] 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.

[0029] The method for measuring the residual monomer concentration 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 monomer in the resin composition is measured using gas chromatography with an FID detector. The residual monomer concentration is measured from the total amount of all monomers.

[0030] The method for measuring the molecular weight is described below. First, tetrahydrofuran is added to the sampled resin composition to dilute it to 0.2 [wt%] to prepare a measurement sample. From this measurement sample, the molecular weight is measured as a polystyrene-equivalent molecular weight using gel permeation chromatography with a differential refractometer.

[0031] 1 is a diagram for explaining a resin property prediction system according to the present embodiment, and an overview of the resin property prediction system 1 will be described with reference to the drawing. The resin physical property prediction system 1 includes a prediction model PM. The prediction model PM is a trained model that is trained in advance to predict a resin physical property value RP based on spectrum data SD. In other words, the prediction model PM is a trained model that uses the spectrum data SD as an explanatory variable X and the resin physical property value RP as an objective variable Y.

[0032] The resin physical property prediction system 1 predicts a resin physical property RP based on spectrum data SD measured by either offline measurement or online measurement. Specifically, the resin physical property value RP may be the NV value, viscosity, residual monomer concentration, molecular weight, or the like.

[0033] The spectrum data SD is obtained from a sample of the resin composition using a NIR sensor (near infrared sensor) or a Raman spectroscopic sensor (Raman spectrophotometer). An example of the spectrum data SD obtained using the NIR sensor or the Raman spectroscopic sensor will be described with reference to Figs. 2 and 3.

[0034] 2 is a diagram showing an NIR spectrum as an example of the spectral data according to the present embodiment. With reference to the figure, an example of the spectral data SD acquired by the NIR sensor will be described. In the figure, the horizontal axis indicates the wavelength [nm (nanometers)], and the vertical axis indicates the transmittance [%]. FIG. 1 shows an example of the spectrum data SD measured immediately after the monomer dropping step was completed.

[0035] 2, when a NIR sensor is used, the spectral data SD indicates the transmittance at each wavelength. In this case, the prediction model PM uses the transmittance at each wavelength as an explanatory variable X to perform learning and prediction. In this case, the number of data included in one spectral data SD may be about 1300.

[0036] 3 is a diagram showing a Raman spectrum as an example of the spectral data according to the present embodiment. With reference to the figure, an example of the spectral data SD acquired by the Raman spectroscopic sensor will be described. The horizontal axis in the figure represents the Raman shift [cm -1 ], and the vertical axis represents the scattering intensity [arbitrary unit]. FIG. 1 shows an example of the spectrum data SD measured immediately after the monomer dropping step was completed.

[0037] 3, when a Raman spectroscopic sensor is used, the spectral data SD indicates the scattering intensity at each Raman shift. In this case, the prediction model PM uses the scattering intensity at each Raman shift as an explanatory variable X to perform learning and prediction. In this case, the number of data included in one piece of spectral data SD may be about 2900.

[0038] In the example shown in the figure, the horizontal axis represents the Raman shift [cm -1 ] and indicates the Stokes scattering observed in the longer wavelength region than the incident light, but the wavelength [nm] may be used as the horizontal axis instead of the Raman shift.

[0039] In the learning stage and prediction stage (inference stage), the spectrum data SD may be acquired by either offline measurement or online measurement.

[0040] Offline measurement refers to taking a sample from the resin composition undergoing reaction in a polymerization reaction vessel, and measuring the spectrum data SD after a predetermined period of time has passed. Specifically, the resin composition sampled from the polymerization reaction vessel is first placed in a glass vessel and allowed to cool to room temperature overnight. After cooling, an NIR sensor or Raman spectroscopic sensor is inserted into the sampled resin composition to measure the spectrum, thereby performing offline measurement.

[0041] The online measurement means taking a sample from the resin composition undergoing reaction in the polymerization reaction vessel and measuring the spectrum data SD in real time. Specifically, the online measurement is performed by installing an NIR sensor or Raman sensor in the polymerization reaction vessel, and measuring the spectrum at regular intervals while synthesizing the acrylic resin with the NIR sensor or Raman sensor inserted in the polymerization liquid.

[0042] 4 is a diagram for explaining the online measurement according to the present embodiment. An example of the sampling timing in the online measurement will be described with reference to the figure. The figure shows the change in temperature of the polymerization reactor over time, with the temperature of the polymerization reactor on the vertical axis and time on the horizontal axis.

[0043] 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, for example, 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. In this case, the temperature of the polymerization reaction tank may be the temperature of the oil bath that heats the reaction vessel.

[0044] First, at time t11, the temperature inside the reaction vessel is increased. At time t12, when the temperature reaches a predetermined reaction temperature (for example, 115° C.), the temperature of the oil bath is kept constant. From time t12 to time t14, a dripping step is performed in which the monomer and the polymerization initiator are dripped. Since this polymerization reaction is an exothermic reaction, the reaction vessel is cooled with cooling water or the like as necessary to maintain the temperature of the sample stored in the reaction vessel. The dripping step may last, for example, 5 hours. During the dripping step, for example, one sampling is performed. In the example shown in FIG. 4, sampling is performed at time t13.

[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. 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. 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.

[0046] After the dripping is completed, a holding step is performed from time t14 to time t18, 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. During the holding step, sampling is performed multiple times, for example, every hour. In the example shown in FIG. 4, sampling is performed a total of five times every hour during the period from time t14 to time t18. By performing multiple samplings during this holding step, for example, the viscosity of the resin composition can be measured. After the holding step, a dilution and cooling step is carried out from time t18 to time t19, 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 can be measured.

[0047] When using a prediction model PM trained on the basis of spectrum data SD obtained by offline measurement, it is preferable to make predictions on the basis of spectrum data SD obtained by offline measurement. Also, when using a prediction model PM trained on the basis of spectrum data SD obtained by online measurement, it is preferable to make predictions on the basis of spectrum data SD obtained by online measurement.

[0048] In addition, even if the prediction model PM has been trained based on the spectrum data SD obtained by offline measurement, it is possible to make predictions based on the spectrum data SD obtained by online measurement by applying temperature correction. In addition, even if the prediction model PM has been trained based on the spectrum data SD obtained by online measurement, it is possible to make predictions based on the spectrum data SD obtained by offline measurement by applying temperature correction.

[0049] The method of measuring the spectrum data SD is not limited to the above-mentioned examples of offline and online measurements, and may be other methods.

[0050] [Learning stage] An example of the learning stage will now be described with reference to FIGS. 5 is a functional configuration diagram for explaining an example of a learning stage according to the present embodiment. Learning of the prediction model PM will be explained with reference to the same figure. In the following explanation, the prediction model PM may be referred to as a learning device or a trained model 122, etc.

[0051] The prediction model PM is trained by supervised learning using training data TD. The training data TD is associated with the spectral data SD measured in advance by online or offline measurement and the resin physical property value RP. The prediction model PM predicts the resin physical property value RP according to the spectral data SD. That is, the prediction model PM is trained to predict the resin physical property value RP according to the spectral data SD, using the spectral data SD obtained as a result of measuring the resin composition sampled in the process of performing a resin polymerization reaction in a batch reaction tank with a spectroscopic sensor and the resin physical property value RP of the resin composition as the training data TD.

[0052] Here, the prediction model PM may try regression methods such as Ridge regression, Lasso regression, random forest, support vector regression, and partial least squares regression (PLS) to select the most appropriate method. That is, the prediction model PM first performs a first learning based on the same teacher data TD using multiple machine learning algorithms. Next, the prediction model PM selects one of the multiple results of the first learning. In this way, by trying learning using multiple machine learning algorithms, the optimal regression method can be selected.

[0053] The prediction model PM is trained using a plurality of pieces of training data TD. The plurality of pieces of training data TD may be data measured under different conditions. The plurality of pieces of teacher data TD used for learning are preferably the same in temperature or to a degree that can be regarded as the same. The degree that can be regarded as the same may be, for example, a degree that does not affect the resin physical property value RP.

[0054] 6 is a diagram for explaining an example of a superposition condition for creating teacher data according to the present embodiment. With reference to the drawing, an example of the superposition condition for creating a plurality of teacher data TD will be explained. Here, in order to train the prediction model PM, it is desirable to create multiple pieces of training data TD (if possible, a large amount of data). If multiple pieces of training data TD are created using the same polymerization conditions, overlearning may cause errors in the prediction stage. Therefore, in this embodiment, in order to prevent overlearning, multiple pieces of training data TD are created using basic conditions BC, and multiple pieces of training data TD are created using multiple improvement conditions IC with different conditions centered on the basic conditions BC.

[0055] In the example shown in FIG. 6, the basic conditions BC are a temperature of "115[°C]" and "5[h]" in the dripping process, and a temperature of "115[°C]" and "4[h]" in the hold process. To create the training data TD, for example, data for six lots is acquired under the basic conditions BC. In the measurement described with reference to FIG. 4, six pieces of data can be acquired per lot, so by acquiring data for six lots, a total of 36 pieces of data can be acquired. Note that, under the basic conditions, the liquid temperature may vary by ±1[°C] and the dripping time may vary by ±10[min] for each lot.

[0056] 6 also shows, as improved conditions IC, conditions obtained by changing the time and temperature of the dropping step and the holding step of the basic conditions BC. Specifically, eight conditions, improved conditions IC1 to IC8, are shown as improved conditions IC. In the improved condition IC1, the temperature in the dripping process is 110°C and 5 h, and the temperature in the hold process is 110°C and 4 h. In the improved condition IC2, the temperature in the dripping process is 120°C and 5 h, and the temperature in the hold process is 120°C and 4 h. In the improved condition IC3, the temperature in the dripping process is 115°C and 4 h, and the temperature in the hold process is 115°C and 4 h. In the improved condition IC4, the temperature in the dripping process is 115°C and 6 h, and the temperature in the hold process is 115°C and 4 h. In the improved condition IC5, the temperature in the dripping process is 115°C and 5 hours, and the temperature in the holding process is 110°C and 4 hours. In the improved condition IC6, the temperature in the dripping process is 115°C and 5 hours, and the temperature in the holding process is 120°C and 4 hours. In the improved condition IC7, the temperature in the dripping process is 110°C and 5 hours, and the temperature in the holding process is 115°C and 4 hours. In the improved condition IC8, the temperature in the dripping process is 120°C and 5 hours, and the temperature in the holding process is 115°C and 4 hours.

[0057] The prediction model PM is trained using data acquired under conditions for a total of 14 lots (i.e., 84 pieces of data) as training data TD, for example, using the basic conditions BC for 6 lots and the improved conditions IC for 8 lots as described above.

[0058] [Prediction stage] An example of the prediction stage will now be described with reference to Figures 7 to 9. 7 is a functional configuration diagram showing an example of the functional configuration of a resin property value prediction device at the prediction stage according to this embodiment. With reference to the same figure, an example of a resin property value prediction system 1 at the prediction stage will be described. The resin property value prediction system 1 includes a spectroscopic sensor SS, a resin property value prediction device 10, and a storage device 20. The resin property value prediction device 10 includes an acquisition unit 11, a calculation unit 12, and an output unit 13.

[0059] The spectroscopic sensor SS measures a resin composition sampled in a process of performing a resin polymerization reaction in a batch-type reaction tank. The acquisition unit 11 acquires spectral data SD obtained as a result of the measurement by the spectroscopic sensor SS. The spectroscopic sensor SS that measures the spectral data SD acquired by the acquisition unit 11 may be at least either a near-infrared (NIR) spectroscopic sensor or a Raman spectroscopic sensor.

[0060] Moreover, an example of the resin composition according to the present embodiment may be, specifically, an acrylic resin, more specifically, a solution containing an acrylic resin. That is, the acquisition unit 11 acquires spectral data obtained as a result of measuring a resin product, specifically, an acrylic resin, more specifically, a resin product in the form of a solution containing an acrylic resin, by the spectroscopic sensor SS.

[0061] When the resin property value prediction device 10 is used for offline measurement, the acquisition unit 11 acquires the spectrum data SD measured after the sampled resin composition is allowed to cool for a predetermined period of time. When the resin property prediction device 10 is used for online measurement, the acquisition unit 11 acquires the spectrum data SD measured at predetermined time intervals while the resin polymerization reaction is being carried out in a batch reaction tank.

[0062] The calculation unit 12 uses the prediction model PM to calculate the resin property value RP of the resin composition having the spectrum data SD acquired by the acquisition unit 11. The prediction model PM is a trained model that has been trained in advance using the spectrum data SD obtained by measuring the resin composition and the resin property value RP as teacher data. That is, the calculation unit 12 calculates the resin property value RP by machine learning.

[0063] The output unit 13 outputs information about the resin physical property value RP calculated by the calculation unit 12 to the storage device 20. The information about the resin physical property value RP stored in the storage device 20 may be transmitted to an information processing terminal (not shown) via a predetermined communication network or displayed on a display unit (not shown).

[0064] 8 is a functional configuration diagram showing an example of the functional configuration of the calculation unit according to the present embodiment. With reference to the same figure, an example of the functional configuration of the calculation unit 12 will be described. The calculation unit 12 includes a preprocessing unit 121 and a trained model 122. The trained model 122 is an example of a prediction model PM.

[0065] The preprocessing unit 121 performs preprocessing on the spectrum data SD acquired by the spectroscopic sensor SS. Here, the information included in the spectrum data SD may include information directly used to calculate the resin physical property value RP, and unnecessary information (noise) that may cause an error in calculating the resin physical property value RP. The preprocessing unit 121 performs preprocessing on the spectrum data SD to remove the unnecessary information and extract only the main information. When preprocessing is performed by the preprocessing unit 121, the trained model 122 is trained based on information obtained by preprocessing the spectrum data SD in the training stage as well as in the prediction stage.

[0066] An example of the pre-processing performed by the pre-processing unit 121 is a smoothing process. In other words, the pre-processing unit 121 performs a smoothing process on the spectrum data SD acquired by the acquisition unit 11, thereby performing pre-processing on the spectrum data SD. Specifically, the pre-processing unit 121 may perform the smoothing process by approximating to a polynomial. Preferably, the pre-processing unit 121 may perform the smoothing by local regression. In this case, by varying the local width (kernel width), necessary information may be lost or noise may not be completely removed. Therefore, the pre-processing unit 121 may regress the spectrum data SD by selecting an optimal local width.

[0067] When the smoothing process is performed by the pre-processing unit 121, the calculation unit 12 calculates the resin property value RP based on the spectral data SD smoothed by the pre-processing unit 121. When the smoothing process is performed by the pre-processing unit 121, the trained model 122 is trained based on information obtained by smoothing the spectral data SD in the training stage. In other words, the spectral data SD included in the teacher data TD may be spectral data SD obtained by smoothing the spectral data SD measured in advance by the spectroscopic sensor SS.

[0068] Fig. 9 is a diagram for explaining an example of the smoothing process according to this embodiment. An example of the smoothing process performed by the pre-processing unit 121 will be explained with reference to the figure. Fig. 9(A) shows the spectrum data SD before the smoothing process is performed, and Fig. 9(B) to Fig. 9(D) show the results after the smoothing process is performed on the spectrum data SD shown in Fig. 9(A) with different central widths.

[0069] The spectrum data SD shown in Fig. 9(A) has not been subjected to smoothing processing, and therefore contains a large amount of noise.

[0070] Fig. 9(B) is an example of a case where smoothing processing is performed by local regression with a local width of 21. As shown in the figure, as a result of performing local regression with a local width of 21, the noise shown in Fig. 9(A) is removed.

[0071] FIG. 9(C) is an example of a case where smoothing processing is performed by local regression with a local width of 41. As shown in the figure, as a result of performing local regression with a local width of 41, the noise shown in FIG. 9(A) is removed, but there are some areas where important information has been lost. For example, referring to FIG. 9(A) and FIG. 9(B), -1 However, referring to FIG. 9(C), it can be seen that one peak has been omitted.

[0072] Fig. 9(D) is an example of a case where smoothing processing is performed by local regression with a local width of 201. As shown in the figure, as a result of performing local regression with a local width of 201, although the noise shown in Fig. 9(A) is removed, a significant amount of the main information (peaks) is lost.

[0073] When the preprocessing unit 121 performs smoothing processing, it is possible to reduce noise, but there is a risk that extremely small peaks may disappear if they exist. Therefore, it is desirable to select a value for the local width that can reduce noise and does not cause important extremely small peaks to disappear. Specifically, it is preferable that the pre-processing unit 121 performs the smoothing process by local regression with a local width of less than 41. More preferably, it is preferable that the pre-processing unit 121 performs the smoothing process by local regression with a local width of about 21.

[0074] Furthermore, the pre-processing unit 121 may perform differentiation processing in addition to or instead of the smoothing processing of the spectrum data SD acquired by the acquisition unit 11. When the pre-processing unit 121 performs differentiation processing, the calculation unit 12 calculates the resin physical property value RP based on the spectrum data SD that has been subjected to differentiation processing by the pre-processing unit 121. The preprocessing unit 121 can extract spectral information from each of the overlapping peaks by performing differential processing, and can correct the baseline. When the preprocessing unit 121 performs differentiation, the spectral data SD included in the teacher data TD is spectral data SD obtained by differentially processing the spectral data SD previously measured by the spectroscopic sensor SS.

[0075] On the other hand, the pre-processing unit 121 may increase noise in the spectrum by performing differentiation. Therefore, depending on the spectrum data SD, there are cases where it is better to perform differentiation and cases where it is better not to perform differentiation. Therefore, the pre-processing unit 121 may perform smoothing on the spectrum data SD, and further perform differentiation separately and independently, and execute machine learning using the optimal spectrum data SD. In other words, the pre-processing unit 121 may separately and independently perform smoothing processing of the spectral data SD acquired by the acquisition unit 11 and differential processing of the spectral data SD acquired by the acquisition unit 11, and the calculation unit 12 may calculate the resin property value RP based on either the spectral data SD that has been smoothed by the pre-processing unit 121 or the spectral data SD that has been differentially processed by the pre-processing unit 121.

[0076] The preprocessing unit 121 may perform differentiation processing on the spectrum data SD that has been subjected to the smoothing processing.

[0077] As a representative method of the smoothing process and the differentiation process described above, the Savitzky-Golay method (hereinafter, referred to as the SG method) based on the least square method is known. The pre-processing unit 121 may perform the pre-processing by the SG method.

[0078] Furthermore, the calculation unit 12 may perform logarithmic conversion of the resin physical property value RP in addition to smoothing and / or differentiation of the spectrum data SD acquired by the acquisition unit 11. For example, when the relationship between the spectrum data SD and the resin physical property value RP complies with the Arrhenius type physical law, it is effective to perform logarithmic conversion of the resin physical property value RP. On the other hand, by performing logarithmic conversion of the resin physical property value RP, the prediction error of the resin physical property value RP may increase. In other words, when the logarithmic conversion of the resin physical property value RP is performed, the accuracy may improve or not. Therefore, the calculation unit 12 may output either the resin physical property value RP that has been subjected to the logarithmic conversion process or the resin physical property value RP that has not been subjected to the logarithmic conversion process as the resin physical property value RP. Note that the logarithmic conversion process of the resin physical property value is performed separately and independently from the smoothing process and the differentiation process.

[0079] In addition, when the spectrum data SD does not include noise, the resin physical property prediction system 1 may adopt a configuration that does not include the pre-processing unit 121.

[0080] [Prediction results] Next, an example of a prediction result by the resin physical property prediction system 1 will be described with reference to Fig. 10 to Fig. 14. Fig. 10 is a diagram showing a schematic example of a prediction result by the resin physical property prediction device according to this embodiment. The prediction result is plotted as an actual measurement on the horizontal axis and a predicted value by the resin physical property prediction device 10 on the vertical axis, with the error and the coefficient of determination R 2 11 to 14 show the results of combinations in which the type of sensor used was an NIR sensor and a Raman sensor, and in which the prediction model PM was trained using only the basic conditions BC as the training data TD, and in which the prediction model PM was trained using the basic conditions BC and improved conditions IC.

[0081] First, a specific example of the prediction result of the resin physical property prediction device will be described with reference to Fig. 10. Fig. 10(A) shows the error and the coefficient of determination R 2 This is an example of a case where the error and the coefficient of determination R 2 This is an example of a case where the accuracy is poor. The graph shown in FIG. 4 has predicted values ​​on the horizontal axis and measured values ​​on the vertical axis, so the closer the measurement results are to the linear line shown in the figure, the better the accuracy. In the example shown in FIG. 10(A), the error is ±0.49% and the coefficient of determination R 2 The coefficient of determination R is 0.996. Similarly, the further away the measurement result point is on the linear line shown in the figure, the worse the accuracy. In the example shown in FIG. 10(B), the error is ±5.5% and the coefficient of determination R 2 is -0.077.

[0082] 11 to 14 are diagrams for explaining an example of a prediction result of the resin physical property prediction device according to the present embodiment. The example shown in Fig. 11 is an example of a case where an NIR sensor is used and the prediction model PM is trained using only the basic conditions BC. The example shown in Fig. 12 is an example of a case where a Raman sensor is used and the prediction model PM is trained using only the basic conditions BC. The example shown in Fig. 13 is an example of a case where an NIR sensor is used and the prediction model PM is trained using the basic conditions BC and the improved conditions IC. The example shown in Fig. 14 is an example of a case where a Raman sensor is used and the prediction model PM is trained using the basic conditions BC and the improved conditions IC.

[0083] Figures 11 to 14 show the results of predicting Gardner viscosity [mPa s] as an example of a resin physical property value RP, NV value [wt%] as an example of a resin physical property value RP, residual monomer concentration, and molecular weight, respectively. Each result is indicated by "yes" or "no" to indicate whether smoothing, differentiation, and logarithmic transformation were performed.

[0084] First, with reference to FIG. 11, an example will be described in which an NIR sensor is used and the prediction model PM is trained using only the basic conditions BC. The Gardner viscosity [mPa s] was smoothed, differentiated, and logarithmically transformed before learning and prediction. In this case, the coefficient of determination R 2 was 0.958, with an error of ±8.3%. The NV value [wt%] was smoothed, differentiated, and logarithmically transformed before learning and prediction. In this case, the coefficient of determination R 2 was 0.996, with an error of ±0.49%. The residual monomer concentration was smoothed and differentiated, and learning and prediction were performed without logarithmic transformation. In this case, the coefficient of determination R 2 was 0.975, with an error of ±5.4 [wt%]. The molecular weight was smoothed, differentiated, and logarithmically transformed before learning and prediction. In this case, the coefficient of determination R 2 was 0.696, with an error of ±2.5%.

[0085] Next, with reference to FIG. 12, an example will be described in which a Raman sensor is used and the prediction model PM is trained using only the basic conditions BC. The Gardner viscosity [mPa s] was smoothed and logarithmically transformed, and learning and prediction were performed without differentiation. In this case, the coefficient of determination R 2 was 0.945, with an error of ±11.8%. The NV value [wt%] was smoothed and logarithmically transformed, and learning and prediction were performed without differentiation. In this case, the coefficient of determination R 2 was 0.992, with an error of ±0.85%. The residual monomer concentration was smoothed, and learning and prediction were performed without differentiation or logarithmic transformation. In this case, the coefficient of determination R 2 was 0.962, with an error of ±6.9 [wt%]. The molecular weight was smoothed and logarithmically transformed, and learning and prediction were performed without differentiation. In this case, the coefficient of determination R 2 was -0.077, with an error of ±5.5%.

[0086] Next, with reference to FIG. 13, an example will be described in which an NIR sensor is used and the prediction model PM is trained using the basic conditions BC and the improved conditions IC. The Gardner viscosity [mPa s] was smoothed, differentiated, and logarithmically transformed before learning and prediction. In this case, the coefficient of determination R 2 was 0.854, with an error of ±15.3%. The NV value [wt%] was smoothed, differentiated, and logarithmically transformed before learning and prediction. In this case, the coefficient of determination R 2was 0.990, with an error of ±1.04%. The residual monomer concentration was subjected to smoothing, differentiation, and logarithmic transformation before learning and prediction. In this case, the coefficient of determination R 2 was 0.934, with an error of ±7.0 [wt%]. The molecular weight was smoothed, differentiated, and logarithmically transformed before learning and prediction. In this case, the coefficient of determination R 2 was 0.054, with an error of ±5.3%.

[0087] Next, with reference to FIG. 14, an example will be described in which a Raman sensor is used and the prediction model PM is trained using basic conditions BC and improved conditions IC. The Gardner viscosity [mPa s] was smoothed and logarithmically transformed, and learning and prediction were performed without differentiation. In this case, the coefficient of determination R 2 was 0.820, with an error of ±13.7%. The NV value [wt%] was smoothed and logarithmically transformed, and learning and prediction were performed without differentiation. In this case, the coefficient of determination R 2 was 0.971, with an error of ±1.41%. The residual monomer concentration was smoothed, and learning and prediction were performed without differentiation or logarithmic transformation. In this case, the coefficient of determination R 2 was 0.981, with an error of ±4.9 [wt%]. The molecular weight was smoothed and logarithmically transformed, and learning and prediction were performed without differentiation. In this case, the coefficient of determination R 2 was 0.083, with an error of ±5.2%.

[0088] [Summary of the embodiment] According to the embodiment described above, the resin property value prediction device 10 includes the acquisition unit 11, which acquires the spectral data SD obtained by measuring the resin composition with the spectroscopic sensor SS in the process of performing a resin polymerization reaction in a batch-type reaction tank, and includes the calculation unit 12, which calculates the resin property value RP of the resin composition using the prediction model PM. That is, according to the resin property value prediction device 10, since it is not necessary to perform a property inspection work for each sample, it is possible to calculate the properties of the resin composition in a short time. Furthermore, according to the resin property value prediction device 10, since it is not necessary for an operator to measure the resin property value RP, it is possible to reduce the deviation in measurement accuracy.

[0089] According to the above-described embodiment, in the learning stage of the prediction model PM, the spectral data SD included in the teacher data TD is spectral data SD obtained by smoothing the spectral data SD measured in advance by the spectroscopic sensor SS. The calculation unit 12 included in the resin physical property value prediction device 10 includes a pre-processing unit 121, which performs a smoothing process on the spectral data SD acquired by the acquisition unit 11, and calculates the resin physical property value RP based on the spectral data SD smoothed by the pre-processing unit 121. The pre-processing unit 121 performs a smoothing process to remove noise included in the spectral data SD in advance. Therefore, according to the present embodiment, the noise included in the spectral data SD can be removed by the smoothing process, and therefore the physical properties of the resin composition can be calculated with high accuracy.

[0090] According to the above-described embodiment, in the learning stage of the prediction model PM, the spectral data SD included in the teacher data TD is the spectral data SD obtained by differentially processing the spectral data SD measured in advance by the spectroscopic sensor SS. The calculation unit 12 included in the resin physical property value prediction device 10 further performs differential processing of the spectral data SD acquired by the acquisition unit 11 by including the preprocessing unit 121, and calculates the resin physical property value RP based on the spectral data SD subjected to differential processing by the preprocessing unit 121. The preprocessing unit 121 extracts spectral information from overlapping peaks, etc. included in the spectral data SD by performing differential processing, and corrects the baseline. Therefore, according to the present embodiment, it is possible to extract spectral information from overlapping peaks, etc. included in the spectral data SD by differential processing, and correct the baseline, so that the physical properties of the resin composition can be calculated with high accuracy.

[0091] According to the above-described embodiment, the preprocessing unit 121 performs smoothing and differentiation on the spectrum data SD acquired by the acquisition unit 11 separately and independently, and calculates the resin physical property value RP based on either the spectrum data SD smoothed by the preprocessing unit 121 or the spectrum data SD differentiated by the preprocessing unit 121. In other words, since there are cases where an optimal result is obtained by performing smoothing and cases where an optimal result is obtained by performing differentiation, the calculation unit 12 performs smoothing and differentiation separately and independently, and then calculates the resin physical property value RP using either one of them. Therefore, according to the present embodiment, the resin physical property value RP can be calculated with high accuracy.

[0092] Moreover, according to the above-described embodiment, the pre-processing unit 121 performs the smoothing process by local regression with the local width being less than 41. Therefore, according to the resin physical property value prediction device 10, it is possible to perform the smoothing process within a range in which important minimum peaks are unlikely to disappear while removing noise contained in the spectrum data SD.

[0093] Further, according to the above-described embodiment, the acquisition unit 11 acquires the spectrum data SD measured after the sampled resin composition is allowed to cool for a predetermined time. That is, the resin property value prediction device 10 executes machine learning based on the offline measurement result. Therefore, according to the present embodiment, the resin property value RP can be calculated based on the offline measurement result. Since the resin property value prediction device 10 can calculate the resin property value RP based on the offline measurement result, the property evaluation and end point management can be completed in a shorter time than before. Furthermore, since the resin property value prediction device 10 can calculate the resin property value RP based on the offline measurement result, the measurement accuracy does not depend on the operator.

[0094] Further, according to the above-described embodiment, the acquisition unit 11 acquires the spectrum data SD measured at predetermined time intervals while the resin polymerization reaction is being performed in the batch-type reaction tank. That is, the resin property value prediction device 10 executes machine learning based on the results of online measurement. Therefore, according to the present embodiment, the resin property value RP can be calculated based on the results of online measurement. Since the resin property value prediction device 10 can calculate the resin property value RP based on the results of online measurement, the processing can be completed in a shorter time than in the past in the property evaluation and end point management. Furthermore, since the resin property value prediction device 10 can calculate the resin property value RP based on the results of online measurement, the measurement accuracy does not depend on the operator. Furthermore, since sampling is not performed during the resin polymerization reaction in the batch-type reaction tank in the online measurement, the sampling work can be reduced.

[0095] According to the embodiment described above, the acquisition unit 11 acquires the spectral data SD obtained as a result of measuring the resin product, which is an acrylic resin, by the spectroscopic sensor SS. Therefore, according to the embodiment, the resin physical property value prediction device 10 can calculate the resin physical property value RP of the acrylic resin. The acquisition unit 11 can also acquire spectral data SD obtained as a result of measuring the resin product in the form of a solution containing an acrylic resin by the spectroscopic sensor SS. Therefore, the resin property value prediction device 10 can calculate the resin property value RP of the resin product in the form of a solution containing an acrylic resin.

[0096] According to the above-described embodiment, the spectroscopic sensor SS that measures the spectral data SD acquired by the acquisition unit 11 is at least one of a near-infrared (NIR) spectroscopic sensor and a Raman spectroscopic sensor. Therefore, according to the present embodiment, the resin physical property value RP can be calculated based on the spectral data SD measured by the near-infrared spectroscopic sensor or the Raman spectroscopic sensor.

[0097] Furthermore, according to the above-described embodiment, the prediction model PM is trained to predict the resin physical property value RP corresponding to the spectral data SD obtained by measuring the resin composition with the spectroscopic sensor SS in a process of performing a resin polymerization reaction in a batch reaction tank, and the resin physical property values ​​of the resin composition are used as training data TD. Therefore, by being equipped with the prediction model PM, the resin physical property value prediction device 10 can calculate the resin physical property value RP based on the spectral data SD.

[0098] In addition, all or part of the functions of each unit of the resin property prediction system 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.

[0099] 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.

[0100] 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]

[0101] 1...resin property value prediction system, 10...resin property value prediction device, 11...acquisition unit, 12...calculation unit, 13...output unit, 121...preprocessing unit, 122...trained model, SS...spectroscopic sensor, 20...storage device, SD...spectral data, PM...prediction model, RP...resin property value, TD...teaching data

Claims

1. an acquisition unit that acquires spectral data obtained as a result of measuring a resin composition by a spectroscopic sensor in a process of performing a resin polymerization reaction in a batch reaction tank; a calculation unit that calculates a resin physical property value of the resin composition having the spectrum data acquired by the acquisition unit using a trained model trained using the spectrum data of the resin composition in advance and the resin physical property value as teacher data; The calculation unit further includes a pre-processing unit that performs a smoothing process on the spectrum data acquired by the acquisition unit, The pre-processing unit performs the smoothing process by local regression with a local width of less than 41. Resin property value prediction device.

2. The spectral data included in the teacher data is spectral data obtained by smoothing spectral data previously measured by a spectroscopic sensor, Resin physical property values ​​are calculated based on the spectrum data that has been smoothed by the pre-processing unit. The resin physical property value prediction device according to claim 1 .

3. The spectral data included in the teacher data is spectral data obtained by differentially processing spectral data previously measured by a spectroscopic sensor, The preprocessing unit further performs a differential process on the spectrum data acquired by the acquisition unit, The calculation unit calculates a resin physical property value based on the spectrum data that has been subjected to the differential processing by the preprocessing unit. The resin physical property value prediction device according to claim 2 .

4. the pre-processing unit performs the smoothing process on the spectrum data acquired by the acquisition unit and the differentiation process on the spectrum data acquired by the acquisition unit separately and independently; The calculation unit calculates a resin physical property value based on any one of the spectrum data that has been subjected to the smoothing process by the pre-processing unit, the spectrum data that has been subjected to the differentiation process by the pre-processing unit, and the spectrum data that has not been subjected to either the smoothing process or the differentiation process by the pre-processing unit. The resin physical property value prediction device according to claim 3 .

5. The calculation unit performs a logarithmic conversion process on the resin physical property values ​​separately from the smoothing process and the differentiation process, and outputs either the resin physical property values ​​that have been subjected to the logarithmic conversion process or the resin physical property values ​​that have not been subjected to the logarithmic conversion process. The resin physical property value prediction device according to claim 4.

6. The acquiring unit acquires spectrum data measured after the sampled resin composition is allowed to cool for a predetermined period of time. The resin property value prediction device according to any one of claims 1 to 5.

7. The acquisition unit acquires spectrum data measured at predetermined time intervals while a resin polymerization reaction is being carried out in the batch reaction tank. The resin property value prediction device according to any one of claims 1 to 5.

8. The resin composition is an acrylic resin, The acquisition unit acquires spectral data obtained as a result of measuring the acrylic resin by a spectroscopic sensor. The resin physical property value prediction device according to any one of claims 1 to 7.

9. The spectroscopic sensor that measures the spectral data acquired by the acquisition unit is at least one of a near-infrared (NIR) spectroscopic sensor and a Raman spectroscopic sensor. The resin property value prediction device according to any one of claims 1 to 8.

10. An acquisition step of acquiring spectral data obtained as a result of measuring a resin composition by a spectroscopic sensor in a process of performing a resin polymerization reaction in a batch reaction tank; A calculation step of calculating a resin physical property value of the resin composition having the spectrum data acquired by the acquisition step using a trained model trained using the spectrum data of the resin composition previously measured and the resin physical property value as teacher data, The calculation step further includes a pre-processing step of performing a smoothing process on the spectrum data acquired by the acquisition step, The pre-processing step performs the smoothing process by local regression with a local width of less than 41. A method for predicting resin physical properties.

Images