A method and system for controlling the operation of a reaction vessel used in resin production.

Abstract

This invention relates to the field of control, and in particular to a method and system for controlling the operation of a reactor used in resin production. The method involves acquiring real-time operating parameters of the reactor, iteratively training a fuzzy neural network model using a hierarchical collaborative hybrid optimization strategy. The antecedent membership function of the fuzzy neural network model is composed of a Gaussian mixture model, and the consequent is a nonlinear polynomial function of the input variables. The training aims to minimize the comprehensive objective function. The optimization strategy involves using an improved quantum particle swarm optimization algorithm for global search to determine the Gaussian mixture model parameters, and using a recursive least squares algorithm for local search after each iteration to determine the consequent coefficients. During training, importance indices are calculated based on the average activation degree of the fuzzy rules and their contribution to the prediction error. Rules with importance consistently below a preset threshold are removed. After training, control commands for controlling the heating system power and material feed rate of the reactor at the next moment are generated and executed based on the real-time parameters.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of operation control, and particularly relates to a reaction kettle operation control method and system for resin production. BACKGROUND

[0002] The resin polymerization reaction is a typical complex chemical process with multivariable, strong nonlinearity and large time delay. There is a close coupling relationship between key parameters such as kettle temperature, pressure, stirring rate and material viscosity. A slight fluctuation of any parameter may have a significant impact on the reaction process and product performance. The traditional control method, such as proportional-integral-derivative control, although simple in structure and widely used, has difficulty in parameter setting when facing such a complex nonlinear time-varying system as the resin reaction kettle, and is difficult to adapt to changing working conditions, often resulting in low control accuracy and large overshoot.

[0003] Fuzzy neural network (FNN) combines the ability of fuzzy logic to handle uncertain information and the self-learning and nonlinear approximation ability of neural network. Without establishing an accurate mechanism model, the input-output relationship of the process can be constructed by learning historical operation data. Most FNN models only pursue data fitting accuracy during training, and are black box models. The prediction results may be contrary to the basic physical and chemical laws, resulting in poor model generalization ability and insufficient reliability under unexperienced working conditions. The training algorithm of traditional FNN such as gradient descent method is easy to fall into local optimum, and the network structure such as fuzzy rule number usually depends on artificial experience setting, which is difficult to balance between model accuracy and computational complexity. Redundant rules not only increase the computational burden, but also may affect the performance of the model. SUMMARY

[0004] In order to improve the accuracy of the reaction kettle operation control for resin production, the application provides a reaction kettle operation control method for resin production, comprising the following steps:

[0005] Obtaining real-time operation parameters of the reaction kettle for resin production, the real-time operation parameters including kettle temperature, kettle pressure, stirring rate and material viscosity;

[0006] The fuzzy neural network model is iteratively optimized and trained by using a hierarchical collaborative hybrid optimization strategy, the antecedent membership function of the fuzzy neural network model is composed of a Gaussian mixture model, and the consequent is a nonlinear polynomial function of the input variable; a weighted prediction error is calculated based on an asymmetric penalty cost function, the asymmetric penalty cost function imposes a large penalty on the error of the predicted output higher than the target output; a physical consistency deviation is obtained by comparing the output of the fuzzy neural network model with the predicted value of the simplified mechanism model; the iterative optimization and training aims to minimize a comprehensive objective function, the comprehensive objective function is the weighted sum of the weighted prediction error and the physical consistency deviation of the fuzzy neural network model; the hybrid optimization strategy is: using an improved quantum particle swarm optimization algorithm, performing global search according to the comprehensive objective function to determine the parameters in the Gaussian mixture model, and after each global iteration, performing local search using a recursive least square algorithm to determine the coefficients in the nonlinear polynomial of the consequent;

[0007] In the iterative optimization and training process, a rule importance index is calculated according to the average activation degree of each fuzzy rule and the contribution degree to the weighted prediction error, and rules with an importance index continuously lower than a pruning threshold are removed;

[0008] The obtained real-time operation parameters are input into the trained fuzzy neural network model to obtain control instructions for controlling the heating system power and the material feeding rate of the reaction kettle at the next moment.

[0009] Preferably, the calculation formula of the comprehensive objective function J is:

[0010] , wherein is the weighted prediction error, is the physical consistency deviation, and α is a weight coefficient with a value between 0 and 1.

[0011] Preferably, the weighted prediction error is calculated based on an asymmetric penalty cost function, the asymmetric penalty cost function imposes a large penalty on the error of the predicted output higher than the target output, and the method comprises the following steps:

[0012] The prediction error of each sampling moment ;

[0013] When e≥0, the error cost of the sampling moment is ;

[0014] When e<0, the error cost of the sampling moment is , wherein λ is a penalty factor greater than 1.

[0015] The weighted prediction error is the average value of the error costs of all sampling moments.

[0016] Preferably, the physical consistency deviation is obtained by comparing the output of the fuzzy neural network model with the predicted value of the simplified mechanism model, comprising:

[0017] The reaction rate model based on the Arrhenius equation and the reaction kettle heat transfer model based on the law of conservation of energy are established as the simplified mechanism model;

[0018] During the training process, the real-time operating parameters and the corresponding control amounts are simultaneously used as the inputs of the fuzzy neural network model and the simplified mechanism model, and the predicted outputs of the next time reaction kettle state are obtained respectively;

[0019] The root mean square error of the predicted output of the fuzzy neural network model and the predicted output of the simplified mechanism model within a preset time window is calculated as the physical consistency deviation.

[0020] Preferably, the improved quantum particle swarm optimization algorithm is used to perform global search according to the comprehensive objective function to determine the parameters in the Gaussian mixture model, and after each global iteration, the recursive least squares algorithm is used to perform local search to determine the coefficients in the consequent nonlinear polynomial, comprising:

[0021] In a single iteration, the quantum particle swarm optimization algorithm is executed, the center and width parameters in the Gaussian mixture model are used as particle positions, the comprehensive objective function is used as the fitness function, the particle position is updated, and a set of optimized antecedent parameters is obtained;

[0022] The antecedent parameters are fixed, and the recursive least squares algorithm is used to perform one-time solution update on all coefficients in the consequent nonlinear polynomial.

[0023] Preferably, the rule importance index is calculated according to the average activation degree of each fuzzy rule and the contribution degree to the weighted prediction error, comprising:

[0024] The importance index of the jth rule is , wherein is the average activation degree of the rule on all training samples, is the sum of the weighted prediction errors attributed to the rule, and ε is a preset small positive number used to prevent the denominator from being zero.

[0025] Preferably, the rules with importance indexes continuously below the pruning threshold are removed, comprising:

[0026] The rule importance pruning threshold and the persistence judgment threshold are obtained;

[0027] If the importance index of a rule is lower than the pruning threshold for the iteration number reaching the persistence judgment threshold, the rule is removed from the fuzzy neural network model.

[0028] In another aspect, the application further provides a reaction kettle operation control system for resin production, comprising the following modules:

[0029] A parameter acquisition module acquires real-time operation parameters of the reaction kettle for resin production, wherein the real-time operation parameters include in-kettle temperature, in-kettle pressure, stirring rate and material viscosity.

[0030] An iterative optimization training module adopts a hierarchical collaborative hybrid optimization strategy to perform iterative optimization training on the fuzzy neural network model, wherein the antecedent membership function of the fuzzy neural network model is composed of a Gaussian mixture model, and the consequent is a nonlinear polynomial function of input variables; a weighted prediction error is calculated based on an asymmetric penalty cost function, wherein the asymmetric penalty cost function imposes a large penalty on the error of the predicted output being higher than the target output; a physical consistency deviation is obtained by comparing the output of the fuzzy neural network model with the predicted value of the simplified mechanism model; the iterative optimization training aims to minimize a comprehensive objective function, wherein the comprehensive objective function is the weighted sum of the weighted prediction error and the physical consistency deviation of the fuzzy neural network model; the hybrid optimization strategy is as follows: an improved quantum particle swarm optimization algorithm is used to perform global search according to the comprehensive objective function to determine the parameters in the Gaussian mixture model, and after each global iteration, a recursive least squares algorithm is used to perform local search to determine the coefficients in the nonlinear polynomial of the consequent.

[0031] A rule removal module calculates a rule importance index according to the average activation degree of each fuzzy rule and the contribution degree to the weighted prediction error, and removes rules with an importance index continuously lower than a pruning threshold.

[0032] An instruction acquisition and execution module inputs the acquired real-time operation parameters into the trained fuzzy neural network model, obtains control instructions for the heating system power and the material feeding rate of the reaction kettle at the next time point, and executes the control instructions.

[0033] Preferably, the calculation formula of the comprehensive objective function J is as follows:

[0034] wherein is the weighted prediction error, is the physical consistency deviation, and a is a weight coefficient with a value between 0 and 1.

[0035] Preferably, the weighted prediction error is calculated based on an asymmetric penalty cost function, the asymmetric penalty cost function imposing a large penalty on the error of the prediction output being higher than the target output, comprising:

[0036] the prediction error of each sampling time ;

[0037] when e≥0, the error cost of the sampling time is ;

[0038] when e<0, the error cost of the sampling time is , wherein λ is a penalty factor greater than 1;

[0039] the weighted prediction error is the mean of the error costs of all sampling times.

[0040] Preferably, the physical consistency deviation is obtained by comparing the output of the fuzzy neural network model with the prediction value of the simplified mechanism model, comprising:

[0041] establishing a reaction rate model based on the Arrhenius equation and a reaction kettle heat transfer model based on the law of conservation of energy as the simplified mechanism model;

[0042] in the training process, the real-time operating parameters and the corresponding control amounts are simultaneously used as the inputs of the fuzzy neural network model and the simplified mechanism model, and the prediction outputs of the next time of the reaction kettle state are obtained respectively;

[0043] the root mean square error of the prediction output of the fuzzy neural network model and the prediction output of the simplified mechanism model in a preset time window is calculated as the physical consistency deviation.

[0044] Preferably, the improved quantum particle swarm optimization algorithm is used to perform global search according to the comprehensive objective function to determine the parameters in the Gaussian mixture model, and after each global iteration, the recursive least squares algorithm is used to perform local search to determine the coefficients in the consequent nonlinear polynomial, comprising:

[0045] in a single iteration, the quantum particle swarm optimization algorithm is executed, the center and width parameters in the Gaussian mixture model are used as the particle positions, the comprehensive objective function is used as the fitness function, the particle position is updated, and a set of optimized antecedent parameters is obtained;

[0046] the antecedent parameters are fixed, and the recursive least squares algorithm is used to perform one-time solution update on all coefficients in the consequent nonlinear polynomial.

[0047] Preferably, the rule importance index is calculated according to the average activation degree of each fuzzy rule and the contribution degree to the weighted prediction error, comprising:

[0048] the importance indicator of the jth rule is wherein is the average activation of the rule over all training samples, is the sum of weighted prediction errors attributed to the rule, and ε is a preset small positive number for preventing the denominator from being zero.

[0049] Preferably, the rules whose importance indicators continuously remain below the pruning threshold are removed, including:

[0050] obtaining a rule importance pruning threshold and a persistence judgment threshold;

[0051] After each iteration optimization, if the number of iterations in which the importance indicator of a rule is below the pruning threshold reaches the persistence judgment threshold, the rule is deleted from the fuzzy neural network model.

[0052] The present application can find the optimal parameters of the fuzzy neural network by adopting the hierarchical collaborative hybrid optimization strategy, combining the global search ability of the quantum particle swarm algorithm with the local fast convergence characteristics of the recursive least squares algorithm, and improving the prediction accuracy of the model. The physical consistency deviation based on the mechanism model is introduced into the training objective function, so that the model is constrained by the physical law during the learning process, ensuring the rationality and generalization ability of the model output, avoiding the control decisions that may violate the physical common sense of the pure black box model, and enhancing the reliability of the model under complex working conditions. Moreover, by introducing the asymmetric penalty cost function, the overshoot of the key process parameters can be targetedly inhibited, so that the control strategy is more suitable for the safety risk consideration in actual production, and the safety of the system is improved. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is a flowchart of a specific embodiment;

[0054] Figure 2 is a schematic diagram of a reaction kettle control;

[0055] Figure 3 is a schematic diagram of pruning rules. DETAILED DESCRIPTION

[0056] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present description. However, it will be apparent to one skilled in the art that the present description can be practiced without the specific details set forth in this description. In other instances, well-known methods, procedures, components, and circuits have not been described in detail as not to unnecessarily obscure aspects of the present description.

[0057] The terminology used in this disclosure of one or more embodiments is for the purpose of describing particular embodiments only and is not intended to be limiting of this disclosure of one or more embodiments. As used in this disclosure and the appended claims, the singular forms "a," "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0058] It is to be understood that, although the terms first, second, etc. can be used herein to describe various information, these terms are not intended to denote a temporal sequence, but are used only to distinguish one piece of information from another. For example, a first can be termed a second, and, similarly, a second can be termed a first, without departing from the scope of this disclosure of one or more embodiments. As used herein, the word "if" can be interpreted to mean "when" or "upon" or "in response to determining" depending on the context.

[0059] In one embodiment, the present application proposes a reaction kettle operation control method for resin production, as shown in the following steps: Figure 1

[0060] S1, acquiring real-time operation parameters of the reaction kettle for resin production, the real-time operation parameters including kettle temperature, kettle pressure, stirring rate and material viscosity;

[0061] Through the thermocouple temperature sensor, pressure transmitter, motor speed sensor and online viscometer installed on the reaction kettle, the real-time data of the kettle temperature, kettle pressure, stirring rate and material viscosity are collected via the distributed control system (DCS) or programmable logic controller (PLC) at a fixed sampling period, for example, once per second, and the data are transmitted to the upper computer through the industrial Ethernet, as shown in the following figure: Figure 2

[0062] ​​S2, a hierarchical collaborative hybrid optimization strategy is used to iteratively optimize and train the fuzzy neural network model. The antecedent membership function of the fuzzy neural network model is composed of a Gaussian mixture model, and the consequent is a nonlinear polynomial function of the input variables. The weighted prediction error is calculated based on an asymmetric penalty cost function, which imposes a large penalty on the error where the predicted output is higher than the target output. The physical consistency deviation is obtained by comparing the output of the fuzzy neural network model with the predicted value of the simplified mechanism model. The iterative optimization training aims to minimize the comprehensive objective function, which is the weighted sum of the weighted prediction error and the physical consistency deviation of the fuzzy neural network model. The hybrid optimization strategy is as follows: using an improved quantum particle swarm optimization algorithm, a global search is performed based on the comprehensive objective function to determine the parameters in the Gaussian mixture model. After each global iteration, a recursive least squares algorithm is used to perform a local search to determine the coefficients in the consequent nonlinear polynomial.

[0063] A fuzzy neural network model is constructed, where the antecedent of each fuzzy rule adopts a Gaussian mixture model, with parameters such as center and width to be determined; the consequent adopts a second-order nonlinear polynomial function of the input variables, for example... Where T is temperature, P is pressure, V is viscosity, and coefficient to To be determined. Define the comprehensive objective function. ,in Weight the prediction error of the model. For physical consistency deviation, w1 and w2 are weighting coefficients. Calculation When calculating, an asymmetric penalty cost function is preferred. For example, when the model predicts a temperature higher than the target temperature, the error penalty coefficient is set to 1.5, and when it predicts a temperature lower than the target temperature, it is set to 1.0. At this point, the predicted temperature from the fuzzy neural network is compared with the predicted temperature from a simplified heat transfer reaction model based on the Arrhenius equation and the law of conservation of energy. The square of the difference between the two is then calculated. During training, the outer loop employs an improved quantum particle swarm optimization algorithm, using all center and width parameters of the Gaussian mixture model as particle positions for global optimization. In each iteration of the quantum particle swarm algorithm, for each particle's represented antecedent parameter, the inner loop calls a recursive least squares algorithm, minimizing... Following the principle of [missing information], all optimal coefficients of the consequent nonlinear polynomial are quickly analyzed. The entire process is guided by minimizing the comprehensive objective function J, iterating repeatedly until convergence.

[0064] S3, in the iterative optimization training process, the importance index of each fuzzy rule is calculated according to the average activation degree of each fuzzy rule and the contribution degree to the weighted prediction error, and the rules with the importance index continuously lower than the pruning threshold are removed;

[0065] After the model training every fixed number of iterations, for example, 20 times, the importance index of each fuzzy rule in the network is calculated. The importance index is composed of two parts: one part is the average activation strength of the rule, that is, the average triggering degree of the rule on the whole training data set, which represents the output energy; the other part is the inverse of the average contribution degree of the rule to the weighted prediction error of the model when the rule is activated. The weighted sum of the two parts is the final importance index. A pruning threshold is set, for example, 0.05, and an observation counter is set. If the importance index of a rule is continuously lower than the pruning threshold for 5 times, the rule is considered to be a redundant or harmful rule, and is permanently deleted from the model structure, including its corresponding antecedent Gaussian mixture model parameters and consequent polynomial coefficients.

[0066] S4, input the obtained real-time operating parameters into the trained fuzzy neural network model to obtain control instructions for controlling the heating system power and the material feeding rate of the reaction kettle at the next time.

[0067] The trained fuzzy neural network model with a simplified structure is deployed on an industrial control computer. At each control time, for example, every 10 seconds, the real-time collected values of the four parameters, i.e., the temperature in the kettle, the pressure in the kettle, the stirring rate and the material viscosity, are taken as an input vector and sent to the fuzzy neural network model. The model is calculated forwardly, and two specific numerical values are output, one representing the target power of the heating system in the next control period, with the unit of kilowatt, and the other representing the target rate of the material feeding in the next control period, with the unit of kilogram per hour. The two control instruction values are issued to the field execution mechanism through the DCS system, for example, the thyristor power regulator for controlling the heating jacket and the frequency converter for controlling the feeding pump, to realize the closed-loop control of the reaction kettle.

[0068] In an optional embodiment, the calculation formula of the comprehensive target function J is:

[0069] wherein is the weighted prediction error, is the physical consistency deviation, and a is a weight coefficient with a value between 0 and 1.

[0070] For example, in the temperature prediction task of a polymerization reaction kettle, represents the difference between the model predicted temperature and the actually measured temperature, ​The model prediction results are compared with the degree of violation of basic physical principles such as the law of conservation of energy. If a is set to a higher value, such as 0.9, the optimization process will mainly focus on reducing the prediction error , and the model will be closer to the historical measurement data. This setting is suitable for scenarios where the process mechanism is relatively obscure, but there is a large amount of high-quality historical data.

[0071] On the contrary, if a is set to a lower value, such as 0.2, the focus of the optimization process will shift to reducing the physical consistency deviation , ensuring that the output of the model is physically reasonable and interpretable, even if this may slightly sacrifice the fitting accuracy of certain historical data points. As an example of an actual optimization step, assume that the model weighted prediction error is 0.05, and the physical consistency deviation is 0.12. If a is set to 0.6, the integrated objective function value J is 0.078. The next iteration of the optimization algorithm will adjust the model parameters to seek a further reduction in this integrated objective value, resulting in an accurate and reliable prediction model.

[0072] In an optional embodiment, the weighted prediction error is calculated based on an asymmetric penalty cost function that imposes a large penalty on errors where the prediction output is higher than the target output, comprising:

[0073] The prediction error at each sampling time ;

[0074] When e ≥ 0, the error cost at the sampling time is ;

[0075] When e < 0, the error cost at the sampling time is , where λ is a penalty factor greater than 1;

[0076] The weighted prediction error is the average of the error costs at all sampling times.

[0077] In industrial scenarios, the risks or costs associated with overestimating and underestimating predicted values are not equal. For example, in the prediction of the content of a key ingredient in production, is the target content, is the model predicted content. If the model underestimates the content, i.e., e is greater than 0, it may cause the production process to be extended, increasing energy consumption, but the product quality is still guaranteed. However, if the model overestimates the content, i.e., e is less than 0, it may cause the production to be terminated prematurely, resulting in unqualified product batches, causing significant economic losses and safety risks.

[0078] To cope with this situation, a penalty factor λ greater than 1 can be set, for example λ equals 4. Suppose at a certain time, the target content is 95%, the model predicts 93%, then the error e is 2, and the error cost is 4. At another time, the target content is 95%, the model predicts 97%, then the error e is -2, and the error cost is 16. It can be seen that the same 2% prediction deviation, overestimation produces four times the cost of underestimation. By calculating the mean of the error cost of all times on the entire training data set, the training will tend to produce more conservative predictions to avoid the serious consequences of overestimation.

[0079] In an optional embodiment, the physical consistency deviation is obtained by comparing the output of the fuzzy neural network model with the predicted value of the simplified mechanism model, comprising:

[0080] A reaction rate model based on the Arrhenius equation and a reaction kettle heat transfer model based on the law of conservation of energy are established as the simplified mechanism model;

[0081] During the training process, the real-time operating parameters and the corresponding control amounts are simultaneously used as inputs of the fuzzy neural network model and the simplified mechanism model, respectively, to obtain the predicted outputs of both for the next time reaction kettle state;

[0082] The root mean square error of the predicted output of the fuzzy neural network model and the predicted output of the simplified mechanism model within a preset time window is calculated as the physical consistency deviation.

[0083] When simulating a chemical reaction process, a simplified model based on recognized chemical and physical principles is constructed. The simplified model may include the Arrhenius equation describing the relationship between reaction rate and temperature, and the energy conservation equation describing the heat exchange of the reaction kettle. Although this mechanism model may not be accurate due to inaccurate parameters or not considering all factors, it correctly reflects the basic dynamic trend of the system.

[0084] In each step of model training, a set of same inputs, such as the current temperature of 350 Kelvin and the feed rate of 0.5 kg per second, are provided to the fuzzy neural network model and the simplified mechanism model. The fuzzy neural network model may predict the temperature of the next minute to be 351.2 Kelvin, and the simplified mechanism model calculates the temperature of the next minute to be 350.8 Kelvin according to the physical law. The difference between the two predicted values, i.e. 0.4 Kelvin, constitutes the physical consistency deviation at the current time. By calculating the root mean square error of these differences within a time window, for example, 10 consecutive sampling points, the physical consistency deviation is obtained This deviation value will be added as a penalty term to the total objective function, forcing the prediction results of the fuzzy neural network not to conflict seriously with the basic physical law.

[0085] In an optional embodiment, the step of using an improved quantum particle swarm optimization algorithm to perform a global search based on the comprehensive objective function to determine the parameters in the Gaussian mixture model, and after each global iteration, using a recursive least squares algorithm to perform a local search to determine the coefficients in the consequent nonlinear polynomial, includes:

[0086] In a single iteration, the quantum particle swarm optimization algorithm is executed, using the center and width parameters in the Gaussian mixture model as particle positions, and the comprehensive objective function as the fitness function to update the particle positions, thereby obtaining a set of optimized antecedent parameters;

[0087] With the antecedent parameters fixed, the recursive least squares algorithm is used to solve and update all coefficients in the consequent nonlinear polynomial in one operation.

[0088] The antecedent parameters of the fuzzy rules, namely the center and width of the Gaussian membership function, determine the partitioning of the input space. These parameters have a highly nonlinear relationship with the final output, making their optimization a complex, multi-peaked global search problem. Therefore, the quantum particle swarm optimization algorithm, with its powerful global exploration capabilities, is employed to handle these parameters. For example, for a model with 10 rules, each corresponding to 2 antecedent parameters, the quantum particle swarm algorithm treats a vector containing these 20 parameters as a particle and searches in 20-dimensional space for the particle position that minimizes the comprehensive objective function J.

[0089] Once the quantum particle swarm optimization algorithm completes one iteration and finds a set of currently optimal antecedent parameters, these parameters are temporarily fixed. At this point, the consequent parameters of the model, i.e., the polynomial coefficients of each rule output, form a linear relationship with the total output of the model. For this type of linear parameter solving problem, the recursive least squares algorithm is a computationally extremely fast and efficient local search method. Using the recursive least squares algorithm, based on the fixed antecedent parameters, all consequent polynomial coefficients are solved precisely in one go, quickly fine-tuning the model's output. This iterative and collaborative approach of first globally searching the nonlinear part and then quickly analytically solving the linear part greatly improves the convergence speed and solution quality compared to using a single algorithm to optimize all parameters.

[0090] In an optional embodiment, the step of calculating the rule importance index based on the average activation degree of each fuzzy rule and its contribution to the weighted prediction error includes:

[0091] The importance index for rule j is: ,in The average activation of the rule across all training samples. ε is a preset small positive number to prevent denominator from being zero.

[0092] A good rule should be activated frequently enough to cover a large enough range of input conditions, and should also produce small prediction errors when activated. The numerator of the formula represents the generality or coverage of the rule. For example, after processing 1000 training samples, if the average activation of rule 5 is 0.7, and the average activation of rule 8 is only 0.05, it indicates that rule 5 is applicable to a wider range of input conditions than rule 8. The denominator of the formula represents the accuracy of the rule, i.e., the total weighted prediction error caused by the rule. A rule might have a high activation, but if it always introduces large errors in prediction, its contribution is negative. Suppose rule 5 has a high activation, but its associated total weighted prediction error is as high as 25.3; rule 8 has a low activation, but it is very accurate, with an associated total weighted prediction error of only 0.1. If ε is set to 0.001, the importance index of rule 5 is about 0.028, and the importance index of rule 8 is about 0.5. Through this index, it is clear that rule 8, although not frequently used, is much more valuable than rule 5, which frequently makes errors.

[0093]

[0094] In an optional embodiment, the removing rules whose importance indexes are continuously below the pruning threshold value comprises:

[0095] obtaining a rule importance pruning threshold value and a continuity judgment threshold value;

[0096] after each iteration, if the number of iterations in which the importance index of a rule is below the pruning threshold value reaches the continuity judgment threshold value, the rule is removed from the fuzzy neural network model.

[0097] Suppose the rule importance pruning threshold value is set to 0.1, and the continuity judgment threshold value is set to 10 iterations. Any rule whose importance index is below 0.1 is marked as a potential useless rule, but it is only removed if it is so marked for 10 consecutive iterations, as shown in the following table. Figure 3

[0098] ​​​​​​​Suppose that after the 50th iteration of the model training, the importance indicator of the 3rd rule is calculated to be 0.08, which is lower than 0.1, a counter for the 3rd rule is started and is 1. After the 51th iteration, the importance indicator of the 3rd rule becomes 0.12, which is higher than the threshold, the counter is cleared. In the next 10 consecutive iterations from the 60th to the 69th, the importance indicator of the 3rd rule continuously is lower than 0.1, the counter for the 3rd rule is accumulated from 1 to 10. When the counter reaches the persistence judgment threshold 10, it is confirmed that the 3rd rule performs poorly under multiple model parameter configurations, and is a redundant or harmful rule that stably exists. At this time, the automatic deletion operation is performed to completely remove the 3rd rule from the model structure of the fuzzy neural network.

[0099] Those skilled in the art will appreciate that embodiments of the application can be supplied as a method, a system, or a computer program product. Accordingly, the application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can be embodied in the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk memory, Compact Disc Read-Only Memory (CD-ROM), optical memory, and the like) having computer-usable program code embodied therein.

[0100] The present application is described in reference to the flowchart and / or block diagram of the method, apparatus (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as a combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device that implements the flow Figure 1 The function specified in one flow or multiple flows and / or blocks. Figure 1 The means for implementing the function specified in one flow or multiple flows and / or blocks.

[0101] The above only describes the embodiments of the present application and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of claims of the present application.

Claims

1. A method of controlling the operation of a reaction vessel for producing a resin, characterized by, The method comprises the following steps: obtaining real-time operation parameters of a reactor for resin production, the real-time operation parameters comprising temperature in the reactor, pressure in the reactor, stirring rate and material viscosity; iteratively optimizing and training a fuzzy neural network model by using a hierarchical collaborative hybrid optimization strategy, a membership function of an antecedent of the fuzzy neural network model being constituted by a Gaussian mixture model, and a consequent being a nonlinear polynomial function of input variables; calculating a weighted prediction error based on an asymmetric penalty cost function, the asymmetric penalty cost function imposing a large penalty on an error of a predicted output being higher than a target output; obtaining a physical consistency deviation by comparing an output of the fuzzy neural network model with a predicted value of a simplified mechanism model; the iteratively optimizing and training aiming to minimize a comprehensive objective function, the comprehensive objective function being a weighted sum of the weighted prediction error of the fuzzy neural network model and the physical consistency deviation; the hybrid optimization strategy being: using an improved quantum particle swarm optimization algorithm to perform global search according to the comprehensive objective function to determine parameters in the Gaussian mixture model, and using a recursive least square algorithm to perform local search to determine coefficients in the nonlinear polynomial of the consequent after each global iteration; in the iteratively optimizing and training, calculating a rule importance index according to an average activation degree of each fuzzy rule and a contribution degree of the rule to the weighted prediction error, and removing a rule whose importance index continuously falls below a pruning threshold; inputting the obtained real-time operation parameters into the trained fuzzy neural network model to obtain control instructions for a heating system power and a material feeding rate of the reactor at a next time point and executing the control instructions.

2. The method of claim 1, wherein, a calculation formula of the comprehensive objective function J is: wherein is the weighted prediction error, is the physical consistency bias, and a is a weighting factor with a value between 0 and 1.

3. The method of claim 1, wherein, the weighted prediction error is calculated based on an asymmetric penalty cost function, the asymmetric penalty cost function imposing a large penalty on an error of a predicted output being higher than a target output, and comprising: prediction error at each sampling instant ; When e ≥ 0, the error cost of the sampling time is ; When e < 0, the error cost at the sampling time is where λ is a penalty factor greater than 1. the weighted prediction error is a mean value of error costs at all sampling time points.

4. The method of claim 1, wherein, the physical consistency deviation is obtained by comparing an output of the fuzzy neural network model with a predicted value of a simplified mechanism model, and comprising: a reaction rate model based on an Arrhenius equation and a reactor heat transfer model based on an energy conservation law are established as the simplified mechanism model; in the training process, the real-time operation parameters and corresponding control amounts are simultaneously used as inputs of the fuzzy neural network model and the simplified mechanism model to respectively obtain predicted outputs of the two models for a state of the reactor at a next time point; a root mean square error of the predicted output of the fuzzy neural network model and the predicted output of the simplified mechanism model within a preset time window is calculated as the physical consistency deviation.

5. The method of claim 1, wherein, the improved quantum particle swarm optimization algorithm is used to perform global search according to the comprehensive objective function to determine parameters in the Gaussian mixture model, and the recursive least square algorithm is used to perform local search to determine coefficients in the nonlinear polynomial of the consequent after each global iteration, and comprising: In a single iteration, a quantum particle swarm optimization algorithm is performed, center and width parameters in the Gaussian mixture model are taken as particle positions, the comprehensive objective function is taken as a fitness function, particle position updating is performed, and a set of optimized antecedent parameters is obtained; Fixed antecedent parameters, all coefficients in the nonlinear polynomial of the consequent are solved and updated once by using a recursive least square algorithm.

6. The method of claim 1, wherein, The rule importance index is calculated according to the average activation degree of each fuzzy rule and the contribution degree to the weighted prediction error, and the rule importance index is calculated. The importance indicator of the jth rule is wherein is the average activation of the rule over all training samples, is the sum of weighted prediction errors attributed to the rule, and ε is a pre-set small positive number to prevent the denominator from being zero.

7. The method of claim 1, wherein, The rule is removed when the importance index is continuously lower than the pruning threshold value. A rule importance pruning threshold value and a persistence judgment threshold value are obtained. After each iteration optimization, if the iteration number of a rule whose importance index is lower than the pruning threshold value reaches the persistence judgment threshold value, the rule is deleted from the fuzzy neural network model.

8. A control system for a reaction vessel used in resin production, characterized in that, The following modules are included: A parameter acquisition module acquires real-time running parameters of a reaction kettle for resin production, the real-time running parameters including an in-kettle temperature, an in-kettle pressure, a stirring rate and a material viscosity; An iterative optimization training module adopts a hierarchical collaborative hybrid optimization strategy to perform iterative optimization training on a fuzzy neural network model, an antecedent membership function of the fuzzy neural network model is composed of a Gaussian mixture model, and a consequent is a nonlinear polynomial function of an input variable; a weighted prediction error is calculated based on an asymmetric penalty cost function, the asymmetric penalty cost function applies a large penalty to an error in which a predicted output is higher than a target output; a physical consistency deviation is obtained by comparing an output of the fuzzy neural network model with a predicted value of a simplified mechanism model; the iterative optimization training aims to minimize a comprehensive objective function, the comprehensive objective function is a weighted sum of the weighted prediction error of the fuzzy neural network model and the physical consistency deviation; the hybrid optimization strategy is that an improved quantum particle swarm optimization algorithm is used to perform global search according to the comprehensive objective function to determine parameters in the Gaussian mixture model, and a recursive least square algorithm is used to perform local search after each global iteration to determine coefficients in the nonlinear polynomial of the consequent; A rule removal module calculates a rule importance index according to an average activation degree of each fuzzy rule and a contribution degree to the weighted prediction error, and removes a rule whose importance index is continuously lower than a pruning threshold value; An instruction acquisition and execution module inputs the acquired real-time running parameters to the trained fuzzy neural network model, obtains control instructions for a next time heating system power and material feeding rate of the reaction kettle, and executes the control instructions.

9. The system of claim 8, wherein, The calculation formula of the comprehensive objective function J is: wherein is the weighted prediction error, is the physical consistency bias, and a is a weighting factor with a value between 0 and 1.

10. The system of claim 8, wherein, The weighted prediction error is calculated based on an asymmetric penalty cost function, the asymmetric penalty cost function applies a large penalty to an error in which a predicted output is higher than a target output, and the weighted prediction error is a mean value of error costs of all sampling time points. prediction error at each sampling instant ; When e ≥ 0, the error cost of the sampling time is ; When e < 0, the error cost at the sampling time is where λ is a penalty factor greater than 1. The weighted prediction error is a mean value of error costs of all sampling time points.

Images