Method for estimating rubber composition

Abstract

To provide a method for estimating a rubber composition more efficiently.SOLUTION: A method for estimating a rubber composition comprises the following steps (1) to (3): (1) acquiring analytical data obtained by qualitatively analyzing a target rubber composition using an analyzer; (2) creating input data based on analytical data; and (3) inputting the input data into a learned machine learning model and deriving an output from the learned machine learning model. Here, the output of the learned machine learning model corresponds to at least one of an abundance ratio of a compound in the target rubber composition and the physical properties of the target rubber composition.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a rubber composition estimation method, an estimation device, an estimation program, and a trained model generation method. [Background technology]

[0002] Patent Document 1 discloses a method for analyzing a rubber composition using pyrolysis GC-MS (gas chromatography-mass spectrometry). According to Patent Document 1, a rubber material is decomposed in a pyrolysis section, the decomposition products are separated in a gas chromatography section, and these are ionized with an electron beam in a mass spectrometry section and detected with a detector to obtain a total ion chromatogram and mass spectrum. According to Patent Document 1, the composition of the rubber material can be analyzed by qualitatively and quantitatively analyzing the peaks of the obtained total ion chromatogram and mass spectrum.

[0003] Non-Patent Document 1 is an example of a document disclosing quantitative analysis of chromatograms. Non-Patent Document 1 discloses a method for quantitatively analyzing rubber blended with at least two of SBR, NR, BR, and NBR using its pyrolysis chromatogram (pyrogram). In this method, standard samples are first prepared by blending two or three of the four polymers mentioned above in various proportions, and pyrograms are obtained for these standard samples. From each pyrogram, peaks of preselected pyrolysis products are identified, and a calibration curve is created using the relative area method. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2013-160599 [Non-patent literature]

[0005] [Non-Patent Document 1] "Quantitative Analysis of SBR Blend Rubber by Pyrolysis Gas Chromatography," Journal of the Society of Rubber Science and Technology of Japan, Vol. 55, No. 5, 1982 Summary of the Invention [Problem to be solved by the invention]

[0006] As disclosed in Non-Patent Document 1, a method of creating a calibration curve from analytical data of a standard sample and estimating the compounding ratio of an unknown rubber composition, i.e., the proportion of polymers present in the rubber composition, has been used conventionally. However, the calibration curve method requires complex calculations because the calibration curve equation is nonlinear rather than linear. Furthermore, the estimation accuracy may not be sufficient considering the effort required for the calculation. For this reason, a method for estimating the proportion of compounds present based on analytical data more efficiently and with sufficient accuracy, and ultimately a method for estimating the physical properties of a rubber composition, has been desired. This applies not only to quantification based on gas chromatograms, but also to quantification and estimation of physical properties based on analytical data containing information suggesting the proportion of compounds present, such as various chromatograms and spectroscopic data.

[0007] The present disclosure aims to provide a more efficient rubber composition estimation method, estimation device, program, and trained model generation method. [Means for solving the problem]

[0008] A method for estimating a rubber composition according to an aspect of the present disclosure includes the following steps (1) to (3): (1) Obtain analytical data by qualitatively analyzing the target rubber composition using an analytical device. (2) Creating input data based on the analysis data (3) Inputting input data into a trained machine learning model and deriving output from the trained machine learning model. The output of the trained machine learning model corresponds to at least one of the abundance ratio of the compound in the target rubber composition and the physical properties of the target rubber composition.

[0009] In the above aspect, the analysis data may include a plurality of data combining a first measurement value and a second measurement value for the first measurement value, and the input data may be one-dimensional data based on the second measurement value.

[0010] In the above aspect, creating the input data may include normalizing the second measurement value by a maximum value of the second measurement values ​​included in the analysis data.

[0011] In the above aspect, the analytical data may include at least one of a gas chromatogram obtained using a gas chromatograph, a liquid chromatogram obtained using a liquid chromatograph, and spectroscopic spectrum data obtained using a spectrometer.

[0012] In the above aspect, the analytical data may include a plurality of data combining time acquired using a gas chromatograph and signal intensity at that time, and creating the input data may include shifting the analytical data so that the time of the peak of signal intensity corresponding to the first pyrolysis product in the analytical data becomes the origin, and creating one-dimensional input data based on the shifted signal intensity.

[0013] In the above aspect, the analytical data may include a plurality of data combining time acquired using a gas chromatograph and signal intensities at those times, and creating the input data may include flattening signal intensity peaks corresponding to one or more specified pyrolysis products among the signal intensities of the analytical data, and creating one-dimensional input data based on the signal intensities with the flattened signal intensity peaks.

[0014] In the above aspect, the analytical data may include a plurality of data combining time acquired using a gas chromatograph and signal intensity at that time, and creating the input data may include shifting the analytical data and multiplying it by a constant in the time direction so that the time of the peak of signal intensity corresponding to the second pyrolysis product in the analytical data coincides with the time of the peak of signal intensity corresponding to the second pyrolysis product in the reference analytical data, and creating one-dimensional input data based on the signal intensity of the analytical data shifted and multiplied by a constant in the time direction.

[0015] In the above aspect, the output of the trained machine learning model may be a value that estimates the abundance ratio of one or more predetermined polymers contained in the target rubber composition.

[0016] In the above aspect, the machine learning model may be a convolutional neural network.

[0017] A method for estimating a rubber composition according to one aspect of the present disclosure includes a data acquisition unit, a memory unit, a data processing unit, and a derivation unit. The data acquisition unit acquires analytical data obtained by qualitatively analyzing a target rubber composition using an analytical device. The memory unit stores a trained machine learning model. The data processing unit creates input data based on the analytical data. The derivation unit inputs the input data to the trained machine learning model and derives an output from the trained machine learning model. The output of the trained machine learning model corresponds to at least one of the abundance ratio of a compound in the target rubber composition and a physical property of the target rubber composition.

[0018] A rubber composition estimation program according to an aspect of the present disclosure causes a computer to execute the following steps (1) to (3): (1) Obtain analytical data by qualitatively analyzing the target rubber composition using an analytical device. (2) Creating input data based on the analysis data (3) Inputting input data into a trained machine learning model and deriving output from the trained machine learning model. The output of the trained machine learning model corresponds to at least one of the abundance ratio of the compound in the target rubber composition and the physical properties of the target rubber composition.

[0019] A method for generating a trained model according to one aspect of the present disclosure includes the following steps (1) to (4): (1) Obtaining analytical data by qualitatively analyzing a known rubber composition using an analytical device. (2) Creating input data based on the analysis data. (3) Preparing learning data that combines the input data and correct answer data. (4) Using the learning data, adjust the parameters of the machine learning model so that when the input data for the target rubber composition is input, data corresponding to the correct answer data is output. The correct data represents at least one of the abundance ratio of compounds in the known rubber composition and the physical properties of the known rubber composition. [Effects of the Invention]

[0020] According to the above-mentioned viewpoints, a technique is provided for more efficiently estimating at least one of the abundance ratio and physical properties of compounds in a rubber composition from analytical data of the rubber composition. [Brief explanation of the drawings]

[0021] [Figure 1] FIG. 10 is a diagram illustrating an example of a method for acquiring analysis data. [Figure 2] FIG. 2 is a block diagram showing the electrical configuration of the estimation device. [Figure 3] FIG. 2 is a diagram illustrating the structure of a machine learning model according to an embodiment. [Figure 4] 10 is a flowchart showing the flow of an estimation process according to an embodiment. [Figure 5] 10 is a flowchart showing the flow of a learning process according to an embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0022] Hereinafter, a rubber composition estimation method, estimation device, estimation program, and trained model generation method according to an embodiment of the present disclosure will be described. According to this rubber composition estimation method, at least one of the abundance ratio and physical properties of compounds in a target rubber composition can be easily estimated using a machine learning model. Below, a qualitative analysis method for a rubber composition for obtaining analytical data will be described first. Next, a rubber composition estimation device and machine learning model using a machine learning model will be described. Furthermore, a method for creating input data to be input into the machine learning model and a method for generating a trained machine learning model will be described.

[0023] <1.Analysis equipment and analysis data> Fig. 1 is a diagram illustrating an example of a method for obtaining analytical data obtained by qualitatively analyzing a sample of a rubber composition 1 using an analytical device. In the example of Fig. 1, a known gas chromatograph 4 is used as the analytical device. As shown in Fig. 1, the gas chromatograph 4 includes a pyrolysis device 41 connected to a gas supply unit 40, a column 42 that separates pyrolysis products produced in the pyrolysis device 41, and a detector 43 that detects the pyrolysis products flowing in from the column 42 and outputs a detection signal.

[0024] The pyrolysis device 41 has a chamber. The pyrolysis device 41 heats a sample introduced into the chamber to produce pyrolysis products derived from the components contained in the sample. For example, polyisoprene (PI) mainly produces isoprene and limonene, polybutadiene (PB) mainly produces 1,4-butadiene and 4-vinylcyclohexene, and styrene-butadiene copolymer (SBR) mainly produces styrene, 1,4-butadiene, and 4-vinylcyclohexene. PI, PB, and SBR are polymers found in rubbers primarily used in tires. The following explanation will use as an example the estimation of a crosslinked rubber composition 1 containing these polymers.

[0025] The gas supply unit 40 supplies an inert gas into the chamber at a predetermined flow rate. As a result, the pyrolysis products generated in the chamber are sent to the column 42 with the inert gas as a carrier gas. The column 42 is a long, hollow pipe maintained at a predetermined temperature. The column 42 may be filled with a packing material. The inner wall of the column 42 may also have a liquid phase or a stationary phase. As a result, the pyrolysis products sent into the column 42 along with the carrier gas are adsorbed or distributed within the column 42 according to their type, resulting in differences in the time it takes for the products to pass through the column 42. The pyrolysis products that have passed through the column 42 reach the detector 43 sequentially with time differences according to their type.

[0026] The detector 43 may be, for example, but is not limited to, a flame ionization detector (FID). When the pyrolysis products that have passed through the column 42 reach the detector 43, the detector 43 combusts the pyrolysis products to generate ions. The detector 43 outputs an electrical signal based on the current generated in response to the ions. The signal strength of this electrical signal increases as the ion concentration increases. The detector 43 is electrically connected to a data processing device 44, which generates a gas chromatogram based on the electrical signal output by the detector 43. The data processing device 44 may be, for example, but is not limited to, a general-purpose computer.

[0027] A gas chromatogram is an example of analytical data, containing multiple data sets that combine time (as a first measurement value) and the signal intensity of the detector 43 at that time (as a second measurement value). The time is the time from when the sample is introduced into the chamber of the pyrolysis device 41 until it reaches the detector 43 and is detected. This time is also referred to as retention time, duration, or retention time. In a gas chromatogram chart, with time on the horizontal axis and signal intensity on the vertical axis, several signal intensity peaks appear graphically due to the separation of pyrolysis products in the column 42. For example, by using a standard sample with known components and identifying in advance the times of peaks appearing due to the pyrolysis products of that sample, it is possible to identify which pyrolysis product each peak originates from, enabling qualitative analysis.

[0028] The signal intensity of a gas chromatogram is proportional to the concentration of pyrolysis products at that time. Therefore, the gas chromatogram contains information suggesting the abundance ratio of compounds, such as polymers, from which the pyrolysis products are derived in a rubber composition 1. In conventional techniques, standard samples containing various blends of a specific type of polymer are prepared, and calibration curves are created from the gas chromatograms of these standard samples to estimate the abundance ratio of a specific polymer in a target rubber composition 1. For example, as described in Non-Patent Document 1, calibration curve methods are broadly divided into absolute calibration curve methods and relative area methods. The absolute calibration curve method is difficult to estimate with high accuracy because sample weighing errors affect the estimation accuracy, while the relative area method requires a complicated procedure. In addition, in pyrolysis gas chromatography analysis, the time (retention time) of the analytical data can vary with repeated analysis. However, the estimation method according to the present embodiment can accommodate sample weighing errors and time lags to some extent, enabling a simple method to estimate the abundance ratio of a polymer with sufficiently high accuracy. This is explained below.

[0029] <2. Rubber composition estimation device> [Estimation device] 2 is a block diagram showing the electrical configuration of the estimation device 2. The estimation device 2 is a general-purpose computer in terms of hardware, and is realized as, for example, a desktop personal computer, a laptop personal computer, a tablet, or a smartphone. The estimation device 2 is manufactured by installing a program 230 into the general-purpose computer from a computer-readable recording medium 231, such as a CD-ROM or a USB memory, or via a network.

[0030] The estimation device 2 includes a control unit 20, a display unit 21, an input unit 22, a storage unit 23, and a communication unit 24. These units 20 to 24 are connected to one another via a bus line 25 and are capable of communicating with one another. The display unit 21 can be configured with a liquid crystal display, an organic EL display, a plasma display, a liquid crystal element, or the like, and displays various information to the user. The input unit 22 can be configured with a mouse, a keyboard, a touch panel, or the like, and accepts user operations on the estimation device 2. The communication unit 24 communicates with networks, including the Internet, and external devices, and functions as an interface for sending and receiving data.

[0031] The control unit 20 may be configured with a central processing unit (CPU), a graphics processing unit (GPU), a read-only memory (ROM), a random access memory (RAM), etc. The control unit 20 reads and executes a program 230 stored in the storage unit 23, thereby virtually operating as an acquisition unit (data acquisition unit) 20A, a data processing unit 20B, a derivation unit 20C, and a learning unit 20D. The acquisition unit 20A acquires analytical data obtained by qualitatively analyzing a rubber composition 1 of unknown formulation. In this embodiment, the analytical data is numerical data of a gas chromatogram generated by a data processing device 44. Based on this data, thermal decomposition products corresponding to specific peaks can be identified, as described below. The data processing unit 20B creates input data based on the analytical data when applying the analytical data to a trained machine learning model 3A (hereinafter also simply referred to as "model 3A") described below. A method for creating the input data will be described below. The derivation unit 20C inputs input data to the model 3A and derives the output from the model 3A to estimate the abundance ratio of the polymer in the rubber composition 1. The learning unit 20D performs learning of the machine learning model 3 through a learning process described later, and generates a model 3A in which the parameters described later have been adjusted.

[0032] The storage unit 23 can be configured with a non-volatile storage device such as a hard disk or SSD (Solid State Drive). In addition to storing a program 230, the storage unit 23 also stores analysis data acquired by the acquisition unit 20A and input data created by the data processing unit 20B. The storage unit 23 also stores learning data used in the learning process described below, information defining the structure of the machine learning model 3, and information defining the model 3A (including parameters adjusted by learning).

[0033] <3. Machine learning model configuration> Next, with reference to FIG. 3, the configuration of the machine learning model 3 (and model 3A) that undergoes the learning process described below will be described. The machine learning model 3 is a machine learning model that receives input data created by the data processing unit 20B as input and outputs a numerical value corresponding to the abundance ratio (parts by mass) of the polymer in the rubber composition 1, and in this embodiment, is a convolutional neural network having a structure as shown in FIG. 3. In this embodiment, the input data is one-dimensional data in which a predetermined number of values ​​based on signal intensity (second measurement value) are arranged in a time series. Furthermore, the output in this embodiment is three values ​​that represent the estimated proportions (parts by mass) of the polymers PI, PB, and SBR present in the rubber composition 1.

[0034] As shown in FIG. 3, the machine learning model 3 has an input layer 30, followed by an intermediate layer 31, a flattened layer 32, multiple fully connected layers 33A and 33B, and a dropout layer 34. The input layer 30 is the earliest layer, and outputs data obtained by applying a predetermined weight to the input data. The weighting values ​​of each unit in the input layer 30 are adjusted by a learning process described below. The output from the input layer is input to a first intermediate layer 31A of the subsequent intermediate layer 31.

[0035] The intermediate layer 31 is a layer in which a first intermediate layer 31A, a second intermediate layer 31B, and a third intermediate layer 31C are arranged in this order from the input side to the output side, and extracts features of input data. The first intermediate layer 31A includes a convolutional layer 310A, a pooling layer 311A, and a dropout layer 312A. These layers are arranged in this order from the input side to the output side. Similarly, the second intermediate layer 31B and the third intermediate layer 31C include convolutional layers 310B and 310C, pooling layers 311B and 311C, and dropout layers 312B and 312C, arranged in this order. Therefore, in model 3A, the convolution, pooling, and dropout processes are repeated multiple times as the data passes through the first to third intermediate layers 31A to 31C.

[0036] In the convolution layer 310A, convolution processing in the time direction is performed using a predetermined number of one-dimensional convolution filters. Each filter is a one-dimensional array of values ​​for detecting and emphasizing certain features contained in the input, and includes a small number (window size) of values ​​relative to the number of inputs. In the convolution layer 310A, the dot product of each input value and each filter value is calculated. This calculation is repeated to scan the input over a predetermined width in the time direction. By repeating this process for each filter, one-dimensional data sets equal in number to the number of filters are output. The number of filters, window size, and scan width can be set as appropriate. The values ​​of each filter are adjusted by a learning process described below. The output from the convolution layer 310A is input to the next pooling layer 311A.

[0037] The pooling layer 311A ​​performs a pooling process on each set of one-dimensional data, generating the same number of sets of one-dimensional data. Pooling converts each set of one-dimensional data into new one-dimensional data by outputting a response value that represents a partial group of the one-dimensional data in each set. This pooling process reduces the size of each set of one-dimensional data. Furthermore, since the pooling process aggregates multiple values ​​contained in the one-dimensional data into a response value, the position sensitivity of the output is slightly reduced. Therefore, even if the position of the feature to be detected changes slightly in the input data, this change can be absorbed, and the output after the pooling process can be made closer to a constant value. In this embodiment, the response value is determined by the maximum value in the group. However, this is not limited to this method; for example, the average value of the values ​​contained in the group may be used. The output of the pooling layer 311A ​​is input to the next dropout layer 312A. The size of the group can be set as appropriate.

[0038] The dropout layer 312A is a layer that prevents overfitting by randomly selecting and invalidating the output of the pooling layer 311A ​​during training, and in this embodiment, it functions only during training of the machine learning model 3. This improves the robustness of model 3A.

[0039] The output from the first hidden layer 31A is input to the convolutional layer 310B of the second hidden layer 31B. The roles of the convolutional layer 310B, pooling layer 311B, and dropout layer 312B and the processes performed in these layers are the same as those of the convolutional layer 310A, pooling layer 311A, and dropout layer 312A, respectively, and therefore will not be described here. However, the parameters adjusted by learning and the pre-set hyperparameters for these layers differ between the first hidden layer 31A and the second hidden layer 31B. The output from the dropout layer 312B, the last layer of the second hidden layer 31B, is input to the convolutional layer 310C of the subsequent third hidden layer 31C.

[0040] The roles of the convolutional layer 310C, pooling layer 311C, and dropout layer 312C in the third hidden layer 31C and the processing performed in these layers are the same as those of the convolutional layer 310A, pooling layer 311A, and dropout layer 312A, respectively, and therefore will not be described here. However, the parameters adjusted by learning for these layers and the hyperparameters determined in advance by a human are different from those of the first hidden layer 31A and the second hidden layer 31B. The output from the dropout layer 312C, the final layer in the third hidden layer 31C, is input to the subsequent flattened layer 32.

[0041] The Flatten layer 32 is a smoothing layer that combines multiple sets of one-dimensional data output from the intermediate layer 31 into one one-dimensional data. The output from the Flatten layer 32 is input to the subsequent fully connected layer 33A.

[0042] The fully connected layer 33A is a layer in which all units are connected to all outputs of the flattened layer 32. The weighting coefficients and biases connecting each unit are adjusted by a learning process described below. In the fully connected layer 33A, feature information extracted by the intermediate layer 31 is compiled and condensed.

[0043] The subsequent dropout layer 34 is a layer that prevents overfitting by randomly selecting and invalidating the output of the fully connected layer 33A during training, and in this embodiment, it functions only during training of the machine learning model 3. The output side of the dropout layer 34 is connected to the second fully connected layer 33B.

[0044] The fully connected layer 33B is a layer in which all units are connected to all outputs of the dropout layer 34, and in this embodiment, it outputs three values. These three values ​​correspond to estimated values ​​of the abundance ratios of polyisoprene, polybutadiene, and styrene-butadiene polymer in rubber composition 1, respectively. In other words, the polymer blends of these polymers in rubber composition 1 are estimated based on the output from the fully connected layer 33B.

[0045] <4. How to create input data> Next, a method in which the data processing unit 20B creates input data based on the analytical data will be described. The data processing unit 20B of this embodiment performs the following processes [1] to [4] in order on the analytical data acquired by the data acquisition unit 20A to create input data. The analytical data is gas chromatogram data created by the data processing device 44. The following description will also include a step of identifying peaks of specific pyrolysis products from the analytical data as a preliminary step to the processes [1] to [4]. Note that the input data for the training data, which will be described later, is also created by a similar process.

[0046] [0. Qualitative analysis processing] The qualitative analysis process is a process of identifying peaks derived from pyrolysis products, such as butadiene (BD), isoprene (IP), 4-vinylcyclohexene (4VCH), styrene (Sty), and limonene (LI), from gas chromatogram data of the rubber composition 1, where a first measurement value represents time (retention time) and a second measurement value at that time represents signal intensity, and identifying the time of each peak. This qualitative analysis process is performed by measuring standard substances of the above compounds to identify peaks that match the confirmed retention times. In this embodiment, the rubber composition 1 is the target of estimation if it exhibits peaks derived from BD and at least one of IP, 4VCH, Sty, and LI.

[0047] [1. Origin adjustment process] Next, the analytical data is shifted in the time direction so that the time of the first peak derived from BD becomes the origin of time. This shifts the peaks of pyrolysis products other than BD in the time direction. BD is an example of the first pyrolysis product.

[0048] [2. Peak adjustment processing] Next, from the peaks derived from IP, 4VCH, Sty, and LI, the first peak present in both the reference analytical data and the analytical data of the target rubber composition 1 is selected. For example, if peaks derived from LI, Sty, and IP are present in the reference data and peaks derived from Sty and IP are present in the target analytical data, the peak derived from Sty is selected as the first peak. If peaks derived from 4VCH and IP are present in the target analytical data, the peak derived from IP is selected as the first peak. The reference analytical data is data that is randomly selected in advance from analytical data prepared for training the machine learning model 3 described below. The pyrolysis product of the peak selected here is an example of a second pyrolysis product. Then, the analytical data of the rubber composition 1 is multiplied by a constant in the time direction and further shifted in the time direction so that the time of the peak of the second pyrolysis product in the reference data matches the time of the peak of the second pyrolysis product in the analytical data to be estimated.

[0049] [3. Peak flattening processing] Next, the signal intensities of the peaks resulting from BD, IP, and Sty in the analytical data of rubber composition 1 are flattened, and these peaks are removed from the analytical data. In this process, both ends of the peak (start point and end point) are first determined. The start point of the peak is the point before the peak where the signal intensity is minimal and the time is greatest (closest to the peak's apex). The end point of the peak is the point after the peak where the signal intensity is minimal and the time is smallest (closest to the peak's apex). The start point can be determined, but is not limited to, as the first minimum found when tracing the analytical data back along the time axis from the peak's maximum value (apex). Similarly, the end point can be determined as the first minimum found when tracing the analytical data back along the time axis from the peak's maximum value (apex). The peak is then flattened by linear interpolation by connecting both ends of the peak with a line segment. This removes prominent features from the analytical data, making it easier to extract more subtle features using Model 3A.

[0050] [4. Normalization process] Furthermore, the signal intensities of the analytical data after the processing in step 3 are normalized. In this embodiment, the maximum value of the signal intensities after the peak flattening processing is identified, and normalization is performed by dividing all signal intensities included in the analytical data at that time by this maximum value. In other words, the signal intensities of the analytical data after the peak flattening processing are expressed as relative values, with the maximum value being 1.

[0051] By performing the above processes [0] to [4] on the analytical data, one-dimensional input data based on the signal intensity of the analytical data is created. These processes are performed to improve the estimation accuracy by model 3A. Note that, although not limited to this, the input data in this embodiment is data having 2048 values ​​equally spaced along the time axis. The number of values ​​of the input data can be unified, for example, by extracting data from a predetermined time t0 to a predetermined time t1 from the analytical data obtained by process [0] and interpolating this data at 2048 equally spaced points along the time axis.

[0052] <5. Operation of the estimation device> Next, an estimation method for estimating the abundance ratios of polymer, PI, PB, and SBR in a rubber composition 1, which is executed by the estimation device 2, will be described with reference to Fig. 4. The processing in Fig. 4 starts, for example, when a user inputs an instruction to the estimation device 2 to start the estimation processing and the instruction is accepted by the estimation device 2.

[0053] In step S1, the acquisition unit 20A acquires analytical data of the target rubber composition 1 and stores it in the storage unit 23 or RAM. The method for acquiring the analytical data is not particularly limited, and the analytical data may be acquired, for example, from the data processing device 44 via wired or wireless data communication, or may be acquired from a storage medium such as a USB memory.

[0054] In step S2, data processing unit 20B performs the above-mentioned processes [1] to [4] in order on the acquired analysis data to create input data. Data processing unit 20B stores the created input data as input data 233 in storage unit 23 or RAM.

[0055] In step S3, the derivation unit 20C inputs the input data 233 to the model 3A and derives the output from the model 3A. The output from the model 3A is three numerical values ​​that represent the abundance ratios of PI, PB, and SBR estimated in the rubber composition 1 as described above.

[0056] In step S4, the derivation unit 20C outputs the abundance ratios of PI, PB, and SBR derived in step S3. The manner of output is not particularly limited, but for example, a screen displaying these abundance ratios can be created and displayed on the display unit 21. This completes the estimation process.

[0057] 6. Model training method Next, a learning method for the machine learning model 3 performed by the estimation device 2 will be described with reference to FIG.

[0058] In the first step S10, analytical data is obtained for a rubber composition whose compound blend (abundance ratio) is known. More specifically, the above-mentioned gas chromatograph 4 is used to obtain a large number of gas chromatogram data for a large number of known rubber compositions of different types.

[0059] In the next step S11, qualitative analysis of each gas chromatogram is performed in the same manner as described above, and peaks originating from IP, 4VCH, Sty, and LI are identified.

[0060] Next, the above-mentioned processes [1] to [4] are performed in this order on a large amount of analytical data to create one-dimensional data of the same size as the input data for model 3A (step S12). In step S12, reference analytical data [2] is randomly selected from the large amount of analytical data. The one-dimensional data created in step S12 is data corresponding to the input data input to model 3A. For this reason, hereinafter, the one-dimensional data created in step S12 will also be referred to as "input data."

[0061] Next, a large amount of training data is created by combining the input data created in step S12 with correct answer data for this input data, and is stored in the storage unit 23 as training data 232 (step S13). The correct answer data is a set of multiple values ​​indicating the abundance ratios of multiple predetermined compounds in the rubber composition. The multiple predetermined compounds are compounds whose abundance ratios are to be estimated for the rubber composition 1 whose formulation is unknown, and in this embodiment, they are three types of polymers PI, PB, and SBR. The total of the correct answer data represents 100 parts by mass.

[0062] The above steps S10 to S13 are steps of preparing the training data 232. The step of creating the training data may be performed by the estimation device 2, or may be performed by a device other than the estimation device 2. When the training data is created by a device other than the estimation device 2, the created training data is stored as training data 232 in the storage unit 23 of the estimation device 2, thereby enabling the learning unit 20D to train the machine learning model 3. Hereinafter, the learning unit 20D starts training the machine learning model 3.

[0063] In step S14, the learning unit 20D divides the learning data 232 into first data for parameter adjustment and second data for accuracy verification. The first data is data used to adjust the parameters of the machine learning model 3, and the second data is data used to confirm the prediction accuracy of the machine learning model 3 after the parameter adjustment. The ratio between the first data and the second data can be determined as appropriate.

[0064] In step S15, the learning unit 20D sequentially selects a predetermined number of data from the first data, inputs the data to the machine learning model 3, and derives an output from the machine learning model 3. The data selected here may be one selected from a set of data created by equally dividing the first data in advance, or may be a predetermined number of data randomly selected from the first data.

[0065] In step S16, the learning unit 20D adjusts the parameters in each layer of the machine learning model 3 so that the error between the output in step S15 and the ground truth data combined with the input first data is minimized. The loss function for calculating the error can be, for example, the mean square error, but is not limited to this. The parameter adjustment method can be, for example, a known method such as the RMSProp method, but is not limited to this.

[0066] In step S17, the learning unit 20D determines whether all of the first data have been used once each (whether one epoch of learning has been completed) by repeating steps S15 to S17. If it is determined that one epoch of learning has not been completed (NO), the learning unit 20D returns to step S15 again and repeats the processes of steps S15 to S17 for the newly selected first data.

[0067] On the other hand, when it is determined in step S17 that learning for one epoch has ended (YES), the learning unit 20D further determines whether learning for a predetermined number of epochs has ended (step S18). When it is determined that learning for the predetermined number of epochs has not ended (NO), the learning unit 20D returns to step S15 again, selects a predetermined number of first data, and repeats the processes of steps S15 to S17. This allows learning to proceed to the next epoch. After this, the first data is selected using the same combination as in the first epoch.

[0068] On the other hand, when it is determined in step S18 that learning for the predetermined number of epochs has been completed (YES), learning unit 20D ends learning and stores the last updated parameters of machine learning model 3 in storage unit 23. This generates trained machine learning model 3A. This completes the learning process.

[0069] Between steps S17 and S18, i.e., after completing one epoch of learning, learning unit 20D may input the second data to machine learning model 3, derive an output, calculate the error between the output and the ground truth data, and appropriately verify the estimation accuracy of machine learning model 3 in the learning process. Then, learning of machine learning model 3 may be terminated when the estimation accuracy converges within an acceptable range.

[0070] <7. Features> According to the estimation method for a rubber composition 1 of the above embodiment, the abundance ratio of polymers in an unknown rubber composition 1 can be efficiently estimated without calculating a calibration curve. In the above embodiment, a convolutional neural network is used as the machine learning model 3. This increases robustness against time lags in the gas chromatogram, enabling more accurate estimation. Furthermore, in the above embodiment, processing is performed to align the time with the analysis data and to remove large peaks. This makes the feature quantities extracted by model 3A less susceptible to the influence of large peaks, allowing for more subtle features to be extracted from the input data and more accurate estimation.

[0071] <8. Variations> Although one embodiment of the present invention has been described above, the present invention is not limited to the above embodiment, and various modifications are possible without departing from the spirit of the present invention. The gist of the following modifications can be combined as appropriate.

[0072] (1) In the above embodiment, a CNN model was used as the machine learning model 3, but the machine learning model is not limited to this, and other machine learning models such as a model based on Partial Least Squares (PLS), a support vector machine (SVM), a neural network (NN) model, a K-NN model, clustering, k-means, a decision tree, etc., and models that combine these may also be used.

[0073] (2) In the above embodiment, tire rubber containing PI, PB, and SBR was used as an example of the target rubber composition 1, and the case of estimating the abundance ratios of PI, PB, and SBR was described as an example. However, the abundance ratios to be estimated are not limited to these polymers; the abundance ratios of other polymers or compounds expected to be present may also be estimated. That is, Model 3A may be configured to output values ​​estimating the abundance ratios of other compounds. For example, in the stage of preparing a training dataset, the pre-determined abundance ratios of IP, Sty, and BD can be used as correct answer data instead of or in addition to the PI, PB, and SBR blend, and machine learning Model 3 can be trained to configure Model 3A that outputs values ​​corresponding to the abundance ratios of IP, Sty, and BD. With this configuration, it is possible to estimate, for example, the abundance ratio of styrene in SBR. Furthermore, for example, constructing Model 3A by simultaneously training the polymers present in Rubber Composition 1 and the abundance ratios of the monomers that constitute these polymers is expected to further improve the accuracy of estimation.

[0074] (3) In the above embodiment, a tire rubber containing PI, PB, and SBR is given as an example of the target rubber composition 1, but the target rubber composition 1 is not limited to this, and the compounds whose abundance ratios are to be estimated are not limited to PI, PB, and SBR. The target rubber composition 1 may be one before crosslinking or one after crosslinking.

[0075] (4) In the above embodiment, gas chromatogram data is used as the analytical data, but the qualitative analysis method and analytical data are not limited thereto. For example, the qualitative analysis may be performed by liquid chromatography, and the analytical data may be liquid chromatogram data. Furthermore, the qualitative analysis may be performed using a spectrometer using a spectroscopic method such as absorption spectroscopy, Raman spectroscopy, or fluorescence spectroscopy, and the analytical data may be spectral data in which the first measurement value is the wavelength or frequency and the second measurement value is the intensity. In this case, the input data may be one-dimensional data based on the second measurement value. In other words, the analytical data is not particularly limited as long as it is data that allows for qualitative and quantitative analysis of the rubber composition 1.

[0076] (5) In the above embodiment, the input data to the machine learning model 3 and the model 3A was one-dimensional data based on analytical data. However, the input data may be two-dimensional data such as an image based on analytical data. More specifically, the input data may be an image of a gas or liquid chromatogram chart, an image of a spectral curve, or an image created based on these images.

[0077] (6) In the above embodiment, when creating input data based on analytical data, data processing steps [1] to [4] were performed. However, at least a part of these steps may be omitted, or the entirety of these steps may be omitted. For example, a predetermined number of time-series data based on analytical data may be used as input data, or only one of steps [1] to [4] may be performed on the analytical data. Furthermore, the peaks of pyrolysis products selected in [1. Origin Adjustment Process] and [2. Peak Adjustment Process] are not limited to those in the above embodiment, and may be changed as appropriate.

[0078] (7) The structures of machine learning model 3 and model 3A are not limited to those shown in Fig. 3. For example, any of intermediate layers 31A to 31C may be omitted, or additional intermediate layers may be added. Furthermore, the learning method of machine learning model 3 is not limited to that in the above embodiment, and may be modified as appropriate.

[0079] (8) In the above embodiment, the data processing device 44 and the estimation device 2 are configured as separate devices, but these devices may also be configured as an integrated device. Furthermore, the function of the learning unit 20D of the estimation device 2 may be assigned to a device separate from the estimation device 2, and the estimation device 2 may be manufactured by storing in the storage unit 23 data defining the trained model 3A generated by the separate device.

[0080] (9) The control unit 20 of the estimation device 2 may be configured to include a vector processor, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or other chips dedicated to artificial intelligence, in addition to a CPU or a GPU.

[0081] (10) The estimation device 2 may be configured to estimate the physical properties of the rubber composition 1 instead of or in addition to the abundance ratio of the compounds in the rubber composition 1. This is because gas chromatograms, liquid chromatograms, and spectroscopic spectra are data that can be subjected to qualitative and quantitative analysis, and it is possible to estimate the physical properties of the rubber composition 1 based on its composition. In this case, the machine learning model 3 can be configured to perform learning using training data that includes known physical properties of rubber compositions as ground truth data instead of or in addition to the abundance ratio of the compounds, and to output values ​​corresponding to the physical properties of the target rubber composition 1. Examples of physical properties to be estimated include glass transition temperature (Tg), viscosity, storage modulus, loss modulus, and loss tangent (tanδ). [Example]

[0082] Examples of the present disclosure will be described in detail below, but the present disclosure is not limited to these examples.

[0083] <Experiment 1> Three types of rubber samples R1 to R3 were prepared by kneading polymers (PI, PB, SBR), a vulcanization accelerator, and sulfur together and vulcanizing them at 170°C for 12 minutes. As shown in Table 1 below, rubber samples R1 to R3 each had a different blend (parts by mass) of the three polymers. 0.1 mg of each of these rubber samples R1 to R3 was weighed and subjected to a pyrolysis gas chromatograph (pyrolysis apparatus: model JPS-900, manufactured by Japan Analytical Industry Co., Ltd.; gas chromatograph: model GC-2025, manufactured by Shimadzu Corporation), and a gas chromatogram was prepared for each sample. [Table 1]

[0084] From the obtained gas chromatograms, the abundance ratios of PI, PB, and SBR were estimated using a conventional method (Comparative Example) and a machine learning model according to an embodiment of the present disclosure (Example), and the accuracy of the estimations was compared with the actual blend ratios. In the conventional method, the blends of PI, PB, and SBR were known, and the abundance ratios were estimated using a calibration curve created based on pyrolysis gas chromatograms of multiple samples with different blends. The calibration curve was created in accordance with JIS K 6231-2 2007. Specifically, the peak areas of the pyrolysis products BD, IP, and Sty were calculated. The ratios of each peak area to the total peak area of ​​all these pyrolysis products were then calculated and plotted against the known polymer blends. The calibration curve was then fitted to a line using the least squares method. The results are shown in Table 2 below, confirming that the estimation accuracy of the estimation method according to an embodiment of the present disclosure is higher than that of the conventional method. [Table 2] <Experiment 2> Several estimation methods according to the above embodiments were performed on rubber compositions with known blends of PI, PB, and SBR, to estimate the abundance ratios of PI, PB, and SBR in the rubber composition, and to verify the error from the known blend. In the estimation methods according to the above embodiments, in Method 1, the analytical data was used as input data without undergoing the data processing steps [1] to [4] above. In Method 2, the analytical data was used as input data only after normalization of signal intensity [4]. In Method 3, the peak flattening step [3] and the signal intensity normalization step [4] were performed in this order, and the input data was used. In Method 4, the analytical data was used as input data after undergoing the data processing steps [1] to [4] above.

[0085] In both of the above methods, the trained machine learning model used was the same: the convolutional neural network shown in Figure 3. The number of epochs for training the machine learning model was 50. Error validation was performed using 5-fold cross-validation. There were 25 different rubber composition compounding ratios, and the number of samples was 309.

[0086] The root mean square error (RMSE) was calculated for the abundance ratios of PI, PB, and SBR estimated by methods 1 to 4 compared with known blends, and the results are shown below. Method 1:0.580 Method 2:0.060 Method 3: 0.055 Method 4:0.044 These results confirmed that a certain degree of estimation accuracy can be ensured even if data processing steps [1] to [4] are omitted, and that the highest estimation accuracy is achieved when all of the data processing steps [1] to [4] are performed. In particular, it was confirmed that accuracy is significantly improved when signal strength normalization is performed. [Explanation of symbols]

[0087] 1. Rubber composition 2 Estimation device 3. Machine Learning Model 3A Pre-trained machine learning model 4. Gas chromatograph 5. Estimation System

Claims

1. Obtaining analytical data obtained by qualitatively analyzing the target rubber composition using an analytical device; creating input data based on the analysis data; inputting the input data into a trained machine learning model and deriving an output from the trained machine learning model; Including, an output of the trained machine learning model corresponds to at least one of the abundance ratio of the compound in the target rubber composition and the physical properties of the target rubber composition; Method for estimating rubber composition.

2. the analytical data includes at least one of a gas chromatogram obtained using a gas chromatograph, a liquid chromatogram obtained using a liquid chromatograph, and spectroscopic spectrum data obtained using a spectrometer; The method for estimating the rubber composition according to claim 1 .

3. the analysis data includes a plurality of data sets each combining a time acquired using a gas chromatograph and a signal intensity at that time; Creating the input data includes shifting the analytical data so that the time of a peak in signal intensity corresponding to a first pyrolysis product in the analytical data becomes the origin; generating one-dimensional input data based on the shifted signal intensities; A method for estimating the rubber composition according to claim 1 or 2.

4. the analysis data includes a plurality of data sets each combining a time acquired using a gas chromatograph and a signal intensity at that time; The step of creating the input data includes flattening peaks of signal intensities corresponding to one or more predetermined pyrolysis products among the signal intensities of the analysis data; generating one-dimensional input data based on the signal intensity whose peaks have been flattened; The method for estimating a rubber composition according to any one of claims 1 to 3.

5. the analysis data includes a plurality of data sets each combining a time acquired using a gas chromatograph and a signal intensity at that time; Creating the input data includes shifting the analytical data and multiplying it by a constant in the time direction so that the time of a peak in signal intensity corresponding to the second pyrolysis product in the analytical data coincides with the time of a peak in signal intensity corresponding to the second pyrolysis product in the reference analytical data; generating one-dimensional input data based on the signal intensity of the analysis data multiplied by a constant in the shift and time directions; The method for estimating a rubber composition according to any one of claims 1 to 4.

6. The output of the trained machine learning model is a value that estimates the abundance ratio of one or more predetermined polymers contained in the target rubber composition. The method for estimating a rubber composition according to any one of claims 1 to 5.

7. The machine learning model is a convolutional neural network. The method for estimating a rubber composition according to any one of claims 1 to 6.

8. a data acquisition unit that acquires analytical data obtained by qualitatively analyzing a target rubber composition using an analytical device; a memory unit that stores the trained machine learning model; a data processing unit that creates input data based on the analysis data; a derivation unit that inputs the input data into the trained machine learning model and derives an output from the trained machine learning model; Equipped with an output of the trained machine learning model corresponds to at least one of the abundance ratio of the compound in the target rubber composition and the physical properties of the target rubber composition; Rubber composition estimation device.

9. Obtaining analytical data obtained by qualitatively analyzing the target rubber composition using an analytical device; creating input data based on the analysis data; inputting the input data into a trained machine learning model and deriving an output from the trained machine learning model; on the computer, an output of the trained machine learning model corresponds to at least one of the abundance ratio of the compound in the target rubber composition and the physical properties of the target rubber composition; Rubber composition estimation program.

10. Obtaining analytical data obtained by qualitatively analyzing a known rubber composition using an analytical device; creating input data based on the analysis data; preparing learning data that is a combination of the input data and correct answer data; Using the learning data, adjusting parameters of a machine learning model so that when the input data of a target rubber composition is input, data corresponding to the correct answer data is output. Including, The correct answer data represents at least one of the abundance ratio of compounds in the known rubber composition and the physical properties of the known rubber composition. How to generate a trained model.

Images