JPH063243A - Method for estimating physical property of rubber polymer - Google Patents

Abstract

PURPOSE:To shorten the estimation processing time of physical properties by making a neutral network learn the correlation between the operating state of a synthetic rubber polymerizing process and the physical property of the reaction control factor of the polymer, and delivering, from the measurement value under actual operation, the corresponding physical property on the basis of the correlation. CONSTITUTION:To input elements A-1-A-6 forming the input layer of a neural network of three-layer perceptron type is given the information showing the operating state of a rubber polymerizing process such as the concentration and temperature of a feeding monomer and the heating medium temperature of a cooling device. To output elements C-1-C-5 forming the output layer is given the physical information forming the reaction control factor of the polymer such as Mooney viscosity, solution viscosity, and reaction representative temperature. Then, the correlation between the operation state of the input layer and the physical properties of the output layer is preliminarily learned by the neural network, and the operating state of the polymerizing process under operation is measured, and inputted to the input layer, whereby the corresponding physical property value can be provided from the output layer.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention relates to a method for polymerizing at least one kind of physical property which is a reaction control factor selected from Mooney viscosity, solution viscosity, vinylation rate and block styrene rate of a polymer in a synthetic rubber polymerization step. The present invention relates to a method for estimating physical properties of a rubber-like polymer, which is estimated from a measured value indicating the state of a vessel and / or a value estimated by another arithmetic device (hereinafter referred to as a state quantity of a polymerization vessel).

[0002]

2. Description of the Related Art Generally, important factors that determine the quality of synthetic rubber include Mooney viscosity, solution viscosity, vinylation rate, and block styrene rate. It is necessary to uniformly control these values in the polymerization process. Is the key point.

In controlling a conventional polymerization process, first, a sample solution is manually sampled from the process, a room containing a measuring instrument is selected, and a polymer component is taken out from the sample solution by some method, and JIS K6300 or ASTM 16 is used.
44 and the JIS K
The measured value is obtained by applying a solution viscosity measuring instrument such as 2283 or ASTM D445, a vinylation rate measuring instrument such as JIS K0115, and a block styrene percentage measuring instrument such as JIS K0117.

The value thus measured is fed back to the process, and when the measured value deviates from the specification value or is likely to deviate from the specification value, adjustment is performed according to a predetermined procedure so as to reach the specification value. The disadvantage of this method is that it takes 1-2 hours from the step of sampling to the measurement result because the sample processing step is complicated.
Therefore, not only the reaction is delayed in control, but also the process can usually be reviewed only once every 2 to 4 hours, which results in production of a product out of specification for a long time. In addition, the amount of work for measuring physical properties such as Mooney viscosity is complicated and the operator's load is large.

[0005]

In order to improve these, a method for estimating Mooney viscosity, bound styrene content and polymerization addition rate by combining an automatic sampling device, gel permeation chromatography (GPC) and multiple correlation equation ( Japanese Patent Publication No. 3-401), but even with this method, it takes about 1 hour until the GPC result is obtained, and there is a delay in the control of the reaction, and the process is only performed once an hour. It is inevitable that products that are out of specifications will be manufactured for a long time, such as being unable to review. Furthermore, the controllability of the Mooney viscosity remains at the target value ± 5 points.

Therefore, conventionally, the following points have been desired for this type of estimation method.

To review the process in real time.

Between the process status and the measurement result
Eliminate the two-hour delay.

To improve the control accuracy of physical properties.

To reduce the amount of work for obtaining physical properties such as Mooney viscosity.

Therefore, an object of the present invention is to solve the above-mentioned problems, and thereby to immediately respond to a change in the process situation,
Another object of the present invention is to provide a method for estimating the physical properties of a rubber-like polymer, which can eliminate the delay in reaction control, suppress the out-of-specification product manufacturing time to the necessary minimum, and achieve an improvement in quality.

[0012]

To achieve the above object, the present invention provides a neural network input layer with predetermined first information indicating an operating condition of a synthetic rubber polymerization process, Between the operating condition and the physical property, the physical property which is a reaction control factor of the rubber-like polymer in the polymerization step is shown, and the second information corresponding to the first information is given to the output layer of the neural network. The neural network is learned in advance by the neural network to measure the operation status of the polymerization step under operation, and the measurement result is input to the input layer of the neural network as the first information. The second information corresponding to the measurement result is obtained from the output layer of the, and the physical properties that are reaction control factors of the rubber-like polymer during the reaction are estimated. It is characterized in.

[0013]

In the present invention, attention is paid to the fact that the neural network has a function of storing the correlation of information, and the correlation between the operating state of the synthetic rubber polymerization process and the physical property as the reaction control factor of the rubber-like polymer is shown. Let the neural network learn in advance. Next, the neural network derives the physical properties that are the reaction control factors corresponding to this measurement result from the driving situation measured under actual driving based on the above correlation.

[0014]

Embodiments of the present invention will now be described in detail with reference to the drawings.

A typical flow of the synthetic rubber polymerization process is as follows:
The continuous polymerization process consists of stirring tanks arranged in a single stage or multiple stages in series, and the monomer, catalyst and solvent for dilution
Supply to the stirrer tank of the stage. As an example, a one-tank type flow is shown in FIG.

Since the physical property estimation system using the neural network according to the present invention is basically applied to the estimation only within the data range used for the preparation, there are frequent changes in future facilities and changes in operating conditions. At that time, it is necessary to collect data for parameter determination anew and add it to the existing data or select it again and again to determine the parameter (this is called re-learning).
Therefore, the number of calculations is reduced as much as possible when re-learning, and a neural network is created for each combination of a synthetic rubber brand and a polymerizer so that it does not affect unrelated brands and polymerizers. The details of the created neural network are shown below.

(1) Configuration of Neural Network (1) -1. Estimated item amount and required item amount used for estimation All or a part of the main reaction operation amount and operation index (first information indicating the operation status of the present invention) shown in Table 2 are input, and various amounts shown in Table 1 A neural network for outputting (the second information indicating the physical property which is the reaction control factor of the rubber-like polymer of the present invention) alone or in combination is prepared.

[0018]

[Table 1]

[0019]

[Table 2]

The items A to J are brands according to operating conditions,
It is an operation amount that is unique to each polymerizer and is adjusted to obtain the required physical properties or to secure the production amount, and is the minimum control item content necessary to obtain the physical properties. The temperature of the K in the polymerization vessel is related to the manipulated variable, and these values are used as a guide for the operation. The agitator motor input current value of L is an amount that is incidentally determined as a result of the reaction. When the items K and L are added to the items A to J, the prediction accuracy of physical properties is improved. From these item quantities, the most suitable combination (1)
-2 Determined by the method described below.

(1) -2. Configuration of Neural Network As a type of neural network, a characteristic type such as a hierarchical type perceptron, a mutual coupling type Hopfield, or a recurrent network having an arbitrary feedback coupling is devised, but in the present invention, Use a perceptron. The perceptron type is
It has been mathematically proved that any function can be approximated. The perceptron type neural network used in the present invention is composed of at least three layers, and the elements of each layer are connected to each other. However, elements in the same layer, or elements in the input layer and elements in the output layer are not coupled (see FIG. 1).

Here, the input layer is one of the required item amounts in Table 1
Each one corresponds to one input element. In the output layer, each of the estimated item quantities in Table 1 corresponds to one output element. The middle class functions to relate the estimated item quantity with the required item quantity in a non-linear manner. To summarize the operations of the operator to determine the configuration of the neural network, the output layer is determined (select from the estimated item amount).

Optimal combination of input layers (selected from candidates of required item amount).

Determination of the optimum number of intermediate layers and the number of elements forming the intermediate layers.

Determination of optimum values of two coefficients of a sigmoid function which is a non-linear operation function performed by the elements constituting the intermediate layer and the output layer.

Determining the optimum value of the learning coefficient that indicates the degree of update of parameter modification for optimization.

For the following items A and B,
This will be described in detail in (1) -3.

A trial experiment by numerical calculation for determining these is called “learning”. Details of learning will be described in detail in (1) -4.

A. Layer composition The layer composition is as follows.

Input layer (layer that receives input data) ... 1 Intermediate layer (layer that is appropriately selected so that target output data is the best) ... 1 or 2 or more Output layer (layer that outputs output data) ... 1 or The number of intermediate layers is determined so that the error between the target output data and the teacher data is the best.

In the present invention, the neural network is constructed by the three-layer perceptron type which is the minimum construction which is guaranteed that any function can be approximated.

B. Number of elements (neurons) that make up each layer Input layer The elements that make up the input layer correspond to the factors related to the reaction and mainly depend on the brand of the product. The number is of a nature that is dictated by the type of factors that primarily affect the reaction and product properties. In the present invention, brands that do not contain styrene as a component of synthetic rubber are listed in Table 1.
D and E are not used, and styrene values are not included in A, B and C. Further, G is unnecessary for brands that do not need to be vinylized, and H is not necessary for some brands. Also, K,
It is possible to estimate the estimated item amount without L, but there is an effect of increasing the estimation accuracy.

The inventor tried the optimum combination of the elements of the input layer by preliminarily narrowing down the candidates from the necessary item amounts in Table 2 by a method such as statistical analysis and considering the learning result. Through experiments, we have created a highly reliable neural network with good accuracy of the required estimator. The structure of the neural network is shown in FIGS.

As a result, the type of the input layer differs for each brand and each polymerization vessel.

Intermediate Layer The number of elements forming the intermediate layer is largely related to the estimation accuracy of the neural network and the estimation accuracy for unknown data. Generally, when the number of elements is large, the estimation accuracy for learning data is high, but the estimation accuracy for unknown data is high. Becomes worse, and the opposite effect appears when the number of elements is small. Therefore, the present inventor a) creates a highly reliable neural network by selecting the number of elements for which the estimation accuracy of both is highest by the following method. As a result, the number of elements constituting the intermediate layer depends on the brand of the product and the polymerization vessel. As a result, in the case of the present invention, the accuracy of the estimated Mooney viscosity is determined so that both the teacher data and the verification data are in good balance, and it differs depending on the brand and the polymerizer. Evaluation indicators for decisions are explained in the next section.

Since the estimation accuracy of the neural network may change depending on the initial value of the parameter, the number of elements in the intermediate layer is at least 5 in the range of 2 to 20 in order to determine the optimum number of elements. Choose more than cases and learn.

A range of the number of elements to be tried and the number thereof can be efficiently determined by roughly determining the following relational expression between the number of elements in the input layer and the number of data sets that can be used for learning. And we were able to create an accurate neural network.

0.5 × the number of sets of learning data ≧ the number of elements in the input layer × the number of elements in the intermediate layer + (the number of elements in the intermediate layer + the number of elements in the output layer) In the above equation, the first term on the right side represents each element. The sum of the parameters (weights) of the connection is shown, and the second term on the right side shows the sum of the thresholds in the sigmoid function of each element.

(B) Learning is performed for at least 5 cases by randomly changing the initial value of each parameter for one case of the number of elements in the intermediate layer, and (c) the optimum combination of parameters for the number of elements of the intermediate layer is shown in. The selection was made according to the selection criteria, and this was used as a representative example of the number of intermediate layer elements. D) The optimum number of intermediate layer elements was determined by comparing these representative examples.

By simply determining the optimum number of elements in the intermediate layer,
At least 25 or more trial experiments, that is, learnings, are performed per combination of one input layer.

The evaluation indexes used for optimizing the number of elements in the intermediate layer are as follows.

(A) Select a network having the smallest least squared error between the learning data and the verification data or at a level close to the least squared error.

The output elements forming the output layer can be set only for the type of data to be estimated.

In the case of the present invention, each network has one or two or more physical properties.

(1) -3. Output function of each element constituting the intermediate layer and the output layer for expressing nonlinearity The output function of each element has a property of being appropriately selected according to the purpose, and the following sigmoid function is used in the present invention.

[0046]

[Equation 1]

I: Input value from each element of the input layer to each element of the intermediate layer or input value from each element of the intermediate layer to each element of the output layer.

Θ0: A value that determines the shape of the sigmoid function, that is, a value for extracting the features of the data set. If this value is too large, the linearity will increase and the estimation accuracy will deteriorate. If this value is too small, the convergence of various parameters will be poor. In the present invention, the result θ0 = 1 obtained by searching a value for which the parameter is easily determined and the estimation accuracy is good by the trial experiment is used.

Α: A threshold value for adjusting the degree of non-linearity between the input value I and the output value. Similar to the parameter (weight) that determines the strength of coupling with each element, it is automatically determined at the time of learning so that the least square error between the teacher data and the output value is minimized.

(1) -4. Learning Method and Learning Data Determining the optimum combination of the weight between each element and the threshold value of the sigmoid function is called learning, and in the present embodiment, the usual back-propagation method was used as the determination method.

The learning algorithm is not limited to the backpropagation, but since the backpropagation is a calculation based on the steepest descent method, it is more stable than other optimization calculation methods, and thus it is adopted. . The sample data used to determine the combination of the optimum parameters is referred to as teacher data, and in the present embodiment, the combination of the data of the input layer and the data of the output layer at the same time (the time when the physical properties are measured) is regarded as one set of data. Collected the set.

Generally, there is no clear guideline for the number of data sets for learning and verification necessary for improving the accuracy of estimation of physical properties, but it is preferable to use at least 100 or more data sets in this embodiment. Was needed. In this embodiment, the maximum number of products per brand is 700 sets or more. Since the range of learning data is the range of application of this neural network, a data group covering the operating conditions of the process throughout the year was collected and used as learning data and verification data.

From the collected data, the learning data used for learning and the verification data used for verifying the created network are selected. At this time, only the data set confirmed that the operating condition is sufficiently stable is selected. By selecting and using for learning, we were able to create a neural network with good estimation accuracy and high reliability. Also, when selecting learning data from the collected data, basically, it is randomly selected, but a set of data including the maximum value and the minimum value of each input layer element is included in the learning data to create the neural network. Was able to widen the range of application. The estimation ability (generalization) of the created neural network for unknown data was verified with a part of the collected data (about 20% in the case of the present invention).

When the parameters (weights) are modified, the degree of update is usually obtained quickly by using the learning coefficient η and the learning habit term coefficient α.

These coefficients take any positive value,
Normally, it is good to use a value between 0 and 1.0.

If the learning coefficient η is increased, the learning proceeds rapidly, but if it is too large, it vibrates.

As a result of the trial experiment, in the present invention, 0.2 to 0.
5 seems to be good. Further, the present invention is devised to gradually reduce the size in the course of learning.

The school habit term coefficient α is used to suppress the vibration of the solution. However, if this value is too large, the vibration will be large and the convergence will be slow, and if it is too small, a local solution will easily fall.

As a result of the trial experiment, in the present invention, 0.7 to
0.9 seems to be good.

[0060]

[Equation 2]

Learning n + 1-th Parameter Correction Amount η: Learning Coefficient Error with Teacher Signal in Element α: Learning Habitual Term Coefficient Element Output Value Learning n-th Parameter Correction Amount (2) Configuration of Physical Property Estimation System Then, this neural network is executed on the arithmetic unit 14 such as the workstation of FIG. 2 so that the Mooney viscosity that reflects the change of the process condition in real time can be obtained in real time without human intervention. The necessary input information shown in Table 1 has a system configuration in which the DSC (Distributed Control System) 11 is loaded online to the workstation via the network 13.

[0062]

EFFECTS OF THE INVENTION By obtaining the estimated amount of Mooney viscosity in real time, which reflects the process status, it is possible to eliminate the delay in reaction control and eliminate the operator's sampling and sample processing operations. The effect occurs.

By using the neural network, the accuracy is higher than the estimation by the multiple regression equation, the estimation accuracy is similar to the measurement accuracy of the measuring instrument, and the online real-time estimation is possible.
Compared with the measurement by sampling, the efficiency of personnel required for analysis can be improved.

A product having extremely stable physical properties can be obtained.

Out-of-specification products can be reduced as much as possible.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing a configuration of a neural network according to an embodiment of the present invention.

FIG. 2 is a configuration diagram showing an example of a system configuration of a polymerization process.

FIG. 3 is a diagram showing configuration information of the neural network of the present embodiment.

FIG. 4 is a diagram showing the configuration information of the neural network of the present embodiment.

[Explanation of symbols]

 1 Stirring tank 2 Mixing tank 3 Mixer 4 Pump 5 Automatic control valve 6 Flow meter 7 Thermometer 8 Ammeter 9 Pressure gauge 10 Level meter 11 DCS (Distributed control system) 12 Computer gateway 13 Ethernet 14 Workstation (arithmetic device)

Claims (1)

[Claims] 1. The operation status of the synthetic rubber polymerization step is shown.
Predetermined first information is given to the input layer of the neural network, physical properties that are reaction control factors of the rubber-like polymer relating to the polymerization step are shown, and second information corresponding to the first information is given to the neural network. The neural network is made to learn in advance the correlation between the operating condition and the physical property by giving the result to the output layer, and the operating condition of the polymerization step under operation is measured, and the measurement result is used as the first result. By inputting as information to the input layer of the neural network, the second information corresponding to the measurement result is obtained from the output layer of the neural network, and the physical properties that are reaction control factors of the rubber-like polymer during the reaction are acquired. A method for estimating physical properties of a rubber-like polymer, which comprises estimating.