JPH06199904A - Method for changing continuous operation condition of continuous polymerization process - Google Patents

Abstract

PURPOSE:To predict process response in set operation procedures after starting operation conditions and as necessary correct the operation procedures. CONSTITUTION:Plural kinds of process responses are learned to neural networks 4 and 6 using a process response value of e.g. polymerization temperature at initial stage of process response as a key and actual measured response value is given to these neural networks 4 and 6 and a process response pattern corresponding to the measured response value is taken out from these neural networks 4 and 6 as the response prediction of process being in progress.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a standard operating procedure and a neural network for process control when continuously changing polymerization conditions in a continuous polymerization process, that is, a condition changing procedure and response dynamic characteristics divided into a plurality of stages. The present invention relates to a continuous operation condition changing method for a continuous polymerization process, which is controlled in real time using a neural network for estimating the temperature.

[0002]

2. Description of the Related Art Statistical analysis of production results in order to predict the state of a complicated production process such as a polymerization process allows a physical quantity (hereinafter referred to as an operation quantity) indicating the operation content of the process, for example, a flow rate of a stock solution, It has been proposed to create an arithmetic expression showing the relationship between the temperature of the heat medium and the like and the physical quantity (hereinafter referred to as the controlled quantity) indicating the control state of the production process, for example, the monomer concentration and the temperature in the process (special feature). (Fairness No. 3-401).

According to such a proposal, prior to the process operation, a virtual operation amount is given to the above-mentioned arithmetic expression, and the control amount is calculated, whereby the process state can be predicted.

Further, since this proposal has a drawback that the prediction processing time is long, a proposal is made to reduce the prediction processing time by using a neural network (Japanese Patent Application No.
179938) has also been made.

[0005]

When the process control method as described above is applied to the switching of the operating conditions of the polymerization process, such as the brand switching of the polymerization process, the above-mentioned proposal has the following problems.

(1) The relationship between the manipulated variable and the controlled variable for brand switching in the polymerization process is non-linear, and even if the regression analysis with the highest approximation accuracy in the statistical analysis is used, the manipulated variable and the controlled variable are controlled. Difficult to find a relationship. Moreover, a large amount of measurement data is required, and it takes time to predict the control state.

(2) The method of predicting (simulating) the controlled variable using a neural network is suitable for determining the initial operating conditions when the process switching operation is actually performed, but the actual process operation is simulated. If it deviates from the operation trajectory defined by, the correction cannot be made.

Therefore, conventionally, when a change in operating conditions of the polymerization process (for example, brand switching) is started, an operating operator predicts a control amount obtained as a response to an operation of a certain operation amount based on past experience of trial and error. However, it is necessary to manually perform a fine adjustment operation so that the target control amount is obtained after changing the operating conditions.

Not only is it difficult to automate such a process control method, but the operating operator cannot judge whether the operation amount or the operation procedure is appropriate unless the process approaches a steady state to some extent. . In addition, since each operator has different experience and know-how, it is impossible to uniquely determine the optimal operation amount. There is a problem that it is difficult to settle the process within.

Therefore, a first object of the present invention is to provide a continuous operating condition changing method for a continuous polymerization process, which is capable of predicting the dynamic response from the measurement result of the controlled variable of the plant after starting the changing of operating conditions. To do.

A second object of the present invention is to provide a continuous operation condition changing method for a continuous polymerization process, which can obtain a corrected operation amount such that a target response pattern can be obtained based on the above-mentioned result of dynamic response prediction. Especially.

[0012]

In order to achieve the first object of the present invention, in the invention of claim 1, the operating condition change of the continuous polymerization process is started in accordance with the standard operating procedure, and the progress of the polymerization reaction is controlled. Measure the representative physical quantity to represent, the response pattern of the representative physical quantity in the standard procedure of the measurement results, the continuous operation conditions of the continuous polymerization process to correct the operation amount specified in the standard operating procedure to match the response pattern of the measurement result In the changing method, the representative physical quantity at a plurality of points at different times for each dynamic characteristic is input, and a plurality of types of dynamic characteristics are learned in a neural network by giving a parameter value representing the dynamic characteristic as an output. After starting the operation condition change, the representative physical quantity of a plurality of points obtained as a measurement result is given to the input of the neural network,
It is characterized in that a parameter value showing a dynamic characteristic similar to the representative physical quantity of the plurality of points is extracted from the output of the neural network to predict the subsequent dynamic characteristic of the representative physical quantity.

In order to achieve the second object of the present invention, the invention of claim 2 is, in addition to the invention of claim 1, determined in advance as an estimated response pattern obtained as a result of the prediction, and the standard operating procedure. If there is no match with the standard response pattern corresponding to, the subsequent operation based on the difference between the predicted parameter value (first parameter value) obtained as a result of the prediction and the standard parameter value corresponding to the standard operation procedure. It is characterized in that the quantity is relatively corrected.

[0014]

According to the first aspect of the present invention, the neural network stores parameters for determining a plurality of types of dynamic responses (patterns) using the representative physical quantity at a plurality of times as a key. From the representative physical quantity collected from the actual measurement result, the parameters that determine the response pattern similar to the response pattern currently in progress are extracted from the learning results of the neural network, so that even after starting the operation, the actual The response can be predicted.

According to the second aspect of the present invention, a suitable corrected operation amount is acquired by correcting the operation amount based on the relative relationship between the predicted amount and the standard amount for the parameter indicating the response content.

[0016]

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

In this embodiment, a specific example of the polymerization temperature control at the time of changing the brand of the synthetic rubber polymerization process will be shown.

FIG. 2 shows a configuration example of the target plant. Here, SC is the operation amount of the valve 21 when the brand is changed, and is instructed by the plant control device 10.
HT is a polymerization temperature that is a representative physical quantity indicating the progress of brand change. The polymerization temperature HT is measured by the temperature sensor TIC and input to the plant control device 10. θ is the average residence time in the reaction tank. The plant control device 10 displays the measurement result of the state quantity of the reaction tank 20 on the workstation 11
And the plant components are controlled by the operation amount instructed from the workstation. The workstation 11 receives the measurement result of the controlled variable from the plant control device 10 at regular intervals, and based on the measurement result,
A new manipulated variable instruction is given to the plant control device 10.

The circuit configuration of the workstation 11 of FIG. 2 is shown in FIG.

The microprocessor (MPU) 1 inputs the physical property measurement result of the plant measuring equipment from the input / output interface (I / O) 2 according to a built-in software program, and determines the subsequent control contents in correspondence with the measurement result. , The plant control device 10 is instructed via the I / O 2 the operation amount of equipment for process control. Further, when the operating conditions of the continuous polymerization process are changed, the neural network 4 is used in the initial stage, and the control amount (polymerization temperature) of the polymerization reaction actually collected at a plurality of time points is changed from this control amount which will change from now on. Estimate the parameters that determine the dynamics. Further, the operation amount is automatically adjusted so that the response pattern calculated from this estimation result matches the predetermined response pattern as the control target. When the process response is estimated by the neural network 4 or 6, the input / output values for the neural networks 4 and 6 are stored in the buffers 3 and 5 and read / written by the MPU 1.

The neural networks 4 and 6 receive the input data from 1 via the buffers 3 and 5 and send the output data obtained as a result of the processing on the input data to the MPU 1 via the buffers 3 and 5. It is also possible to realize this neural network processing by software processing of the MPU 1.

The neural network used in this embodiment is a hierarchical network consisting of three layers as shown in FIG. 3, and the back propagation method is used as the learning algorithm of the neural network. The model of the plant trained by the neural network is
It has the configuration shown in FIG. 4 and executes the operation of the form of formula- (1). In this example, attention was paid to the fact that the actual response characteristic of this plant is an intermediate response characteristic between the first-order lag system and the second-order lag system, and a two-stage perfect mixing tank array model as shown in FIG. 4 was assumed. Here, K is the gain of the response, that is, the gain of the control signal for controlling the actual device. T is the time constant of the response,
α is the ratio of the primary response to the secondary response, and t is time. In this model, K is set in advance at the time of changing the operating conditions, so that there are two parameters to be identified during the changing of the operating conditions, T and α.

Therefore, a plurality of sets of T and α are prepared in advance in the training data for learning of the neural network, and the formula-
A curve (specifically, a plurality of coordinate values on the curve) as shown in FIG. 5 determined from (1) is given as an input. T and α
The number of sets of is determined in consideration of the estimation accuracy required for the neural network and the accuracy of the data that can be obtained from the plant. Specifically, as the input layer of the neural network, the response relating to the polymerization temperature at several points at fixed time intervals after the start of the response, and the output layer the time constant T of the response at that time.
And the ratio α is given. Here, the response of the polymerization temperature is a value obtained by normalizing the actual change of the polymerization temperature with a target temperature change width.

In the present embodiment, considering the timing of collecting the measurement data from the plant, the measurement data at 0.04θ intervals (in this example, the polymerization temperature) is given as an input. Also, considering the target settling time and response delay time,
Neural network (NNA) 4 that requires input of time-series data for points and neural network (NNB) 6 that requires input of time-series data for 12 points 2
Selectively use a type of network.

Regarding the standard operating procedure, after determining the desired settling time and the number of operating stages, the operating amount at each time is calculated using the two-stage perfect mixing tank array model represented by the equation (1). The time-series operation amount is the standard operation procedure. In the present embodiment, the target settling time was set to 1.5θ, and it was found that the settling time could be settled within the desired time by four-step operation, and as shown by the thick line in FIG. A standard operation procedure consisting of step 1, step 2, step 3, and step 4 in units of time) is determined in advance.

Steps 1 and 2 of the standard operating procedure
Is an operation that greatly accelerates the transition speed from the operating condition before change to the operating condition after change by performing an operation that greatly exceeds the target operation amount, and step 3 prevents overshoot by reducing the operation amount. It is an operation. In step 4,
Match the target operation amount. The process can be settled within a desired time by appropriately selecting the operation amount of each of the three steps of Step 1 to Step 3.

Furthermore, by correcting the standard operating procedure determined in this way according to the actually measured control amount and operating the process, individual differences among operators can be eliminated, and operator experience and know-how can be obtained. Can be unified. The procedure of the actual condition changing operation will be described with reference to FIG.

First, the workstation 11 instructs the program control device 10 about the amount of operation in step 1 according to the standard operation procedure of FIG. 6, and starts the operation of switching the brand. From here, the MPU 1 starts measuring the dead time and then determines the dead time. During this period, a measurement signal of the polymerization temperature is sent from the plant control device 10 at regular time intervals and stored in the workstation 11. Here, the dead time is defined as the time required from the start of changing the operating condition to the actual change of the control amount.

When the dead time is less than 0.24θ, the 12-input neural network 6 is selected, and the sequentially obtained measurement results are the 12-input neural network 6.
Sent to. When the measurement results for 12 points are obtained, M
The PU1 reads the output values (α and T in the equation of FIG. 4) of the 12-point input neural network 6 and uses the α and T to obtain the second and third response values y1 determined by the equation of FIG. The operation amount to be set in the step is determined. Specifically, MP
U1 executes the procedure of FIG. 8 and the gain K depends on the magnitude of y1.
Is determined, and the gain value currently set as the second and third steps is corrected by a minute value ΔK. Then, the corrected gain value Knew and other fixed parameter values are substituted into the equation (1) of FIG. 4 to obtain the response value yn.
Calculate ew in MPU1. Next, the MPU 1 calculates a difference from the target response value y ₅ (fixed value) obtained when the MPU 1 is operated along the trajectory of the target operation amount in FIG. 4, and this difference is within the threshold, that is, the corrected response. When the value Ynew and the target response value y _S are substantially equal to each other, the operation amount of each plant equipment in the second step and the third step, for example, the raw material flow rate and the like are determined from this gain value K by a predetermined arithmetic expression such as proportional (relative) operation Determined by the formula.

If the difference is not within the threshold value, the gain value currently set in the second and third steps is further corrected by ΔK, and the corrected response value is equal to the target response value. The correction of the gain value K is repeated until.

When the correction operation amount is determined in this way, the MPU 1 determines the end point (0.5θ) of one step in FIG.
When the time has elapsed), the second step manipulated variable for correction is instructed to the program control device 10 to correct the actual manipulated variable of the plant. Then, at the end of the second step (1.0
At the time when θ has elapsed, the operation amount in the third step is corrected by the instruction from the MPU 1, and when the third step ends (1.5θ
At the elapse time), the fourth step manipulated variable is set to a predetermined target value.

The above processing is a procedure when the dead time is within 0.24θ. 0.24θ <dead time <0.38
In the case of θ, the 6-input neural network 4 is selected as the neural network, and the same processing as described above is performed.

On the other hand, 0.38θ <dead time <0.74θ
In the case of, the 12-input neural network 6 is selected, but the operation amount of the second step is not the correction operation amount obtained by the above calculation, but the predetermined standard operation amount of the second step ( In FIG. 7, 1.215S, S is the difference between the manipulated variables before and after the switching) is used for the actual manipulation.
Only for the step, the correction operation amount obtained by the above calculation is used for the actual operation (see FIG. 7).

Further, 0.74θ <dead time <0.88θ
In the case of, the 6-input neural network 4 is selected as the neural network, the standard operation amount of the second step is used in the second step, and the correction operation amount is used in the third step ( (See FIG. 7). Also,
When 0.88θ <dead time, the second and third steps are operated according to the standard operation amount, and the target operation amount is set to 1.5θ, and the condition changing operation is ended.

To summarize the above, in this embodiment, when control results at a plurality of time points, for example, the measurement results of the polymerization temperature are collected after the condition change is started, the response of the polymerization temperature is calculated by using the neural network as a parameter (T, Predict in the form of α). If the predicted result does not reach the predetermined target response (y _S ), the current standard manipulated variable trajectory (the manipulated variables in the second and third steps in FIG. 6) is converted into an appropriate model formula (equation in FIG. 4- (1)) is used for correction to predict the control amount,
Find the manipulated variable trajectories whose prediction results match the target response. Since the actual manipulated variable is corrected in real time according to the manipulated variable thus found, the controlled variable of the process reaction can be settled within a desired time to a preset target value as shown in FIG.

Further, by selecting a neural network having different prediction accuracy according to the dead time and varying the number of corrections of the operation amount, the correction operation amount is calculated without lowering the operation control accuracy of the process. Save time,
Absorb the delay in settling time due to dead time.

For reference, the prediction result and the actual response when the response is predicted by using the neural network are shown in FIGS. 10A and 10C, and the regression analysis (least squares method) is performed under the same conditions. (B) and (d) of FIG. 10 schematically show the prediction result and the actual response when the response is predicted by (4). It can be seen that the approximation accuracy is better when the neural network is used, as the curve of the actual response becomes non-linear ((c), (d)).

The above example is an example (corresponding to the invention of claim 2) of predicting the dynamic response of the controlled variable from the measurement result of the controlled variable and automatically correcting the manipulated variable so as to obtain the target response. .
Only by using the dynamic response prediction function of the system (the invention of claim 1), for example, it is possible to determine whether or not the operation procedure manually initialized by the operator is correct.

[0039]

As described above, according to the present invention, the response of the control state of the process can be predicted and the automatic correction can be performed even after the operation is started, and the quality of the control content can be improved. As a result, it is possible to automate complicated condition change control.

[Brief description of drawings]

1 is a block diagram showing a main configuration of a workstation 11 of FIG.

FIG. 2 is a configuration diagram showing a configuration of a plant control system to which the present invention is applied.

FIG. 3 is a configuration diagram showing a configuration example of a neural network.

FIG. 4 is an explanatory diagram showing a two-stage perfect mixing tank row model.

FIG. 5 is a characteristic diagram showing an example of learning teacher data for a neural network used for response pattern estimation (prediction).

FIG. 6 is a diagram showing an operation procedure of the present embodiment.

FIG. 7 is a flowchart showing a control procedure of the workstation 11.

FIG. 8 is a flowchart showing a control procedure of the workstation 11.

9 is a diagram showing the response contents of the control amount to the operation procedure of FIG.

FIG. 10 is a diagram showing prediction results and actual responses of a neural network and regression analysis.

[Explanation of symbols]

 1 MPU 2 I / O 3,5 Buffer 4,6 Neural Network 10 Plant Controller 20 Reactor 21 Valve

Claims (2)

[Claims] 1. A change in operating conditions of a continuous polymerization process is started according to a standard operation procedure, a representative physical quantity indicating the progress of the polymerization reaction is measured, and a response pattern of the representative physical quantity in the standard procedure indicated by the measurement result and the measurement result. In the continuous operation condition changing method of the continuous polymerization process for correcting the operation amount specified in the standard operation procedure for matching the response pattern of the above, the representative physical quantities at a plurality of points at different times for each dynamic characteristic are input, A neural network is made to learn a plurality of types of dynamic characteristics by giving a parameter value representing the characteristics as an output, and after starting the change of the operating conditions, the representative physical quantities at a plurality of points obtained as a measurement result are calculated by the neural network. A parameter value showing dynamic characteristics similar to the representative physical quantity of the plurality of points is given to the input and output of the neural network. Continuous operation condition changing method for continuous polymerization process, characterized in that to predict the dynamics of subsequent of the representative physical quantities by extracting from. 2. A predicted parameter value (first parameter) obtained as a result of the prediction when an estimated response pattern obtained as a result of the prediction does not match a predetermined standard response pattern corresponding to the standard operation procedure. Value) and a standard parameter value corresponding to the standard operating procedure, based on the difference, the subsequent operation amount is relatively corrected, continuous operating conditions changing method of the continuous polymerization process according to claim 1, characterized in that .