Intelligent proportion control method and system for antistatic agent synthesis process

Abstract

The invention provides an intelligent proportion control method and system for an antistatic agent synthesis process, and the method comprises the steps: fuzzifying real-time parameters through real-time and historical process parameters by using an asymmetric membership function constructed based on data distribution skewness and kurtosis; historical data samples are mapped into graph theory nodes, communities are divided through a community discovery algorithm to generate fuzzy rules, initial weights are set, and an initial rule base is constructed; using a recursive least square method to identify rule consequent parameters, combining redundancy rules according to cosine similarity, and combining a particle swarm optimization algorithm to optimize antecedent parameters; and inputting the fuzzification real-time parameters into the optimized fuzzy neural network, calculating the activation intensity of the rule, adjusting the weighted average weight based on the information entropy of the current activation intensity, and obtaining the proportion control quantity of each component after defuzzification.

Description

Technical Field

[0001] This application belongs to the field of control, and in particular relates to an intelligent proportioning control method and system for the synthesis process of antistatic agents. Background Technology

[0002] Antistatic agents, as an important class of functional additives, are used to eliminate static electricity accumulation on material surfaces, ensuring production safety and product performance. The quality and performance of antistatic agent products depend on the ratio of raw material components during synthesis. In actual production, various factors such as reaction temperature, pressure, catalyst activity, and batch-to-batch variations of raw materials can affect the optimal ratio, making ratio control a crucial step in ensuring product quality stability and consistency. Experience-based manual adjustment is highly subjective, difficult to achieve precise control, labor-intensive, and results in poor product quality consistency. PID controllers, as linear controllers, have poor control performance and cannot cope with changes in operating conditions. Fuzzy neural networks combine the ability of fuzzy logic to handle uncertainty problems with the self-learning ability of neural networks, providing a new approach to intelligent control of the antistatic agent synthesis process. However, fuzzy neural networks have the following shortcomings:

[0003] 1. Symmetrical membership functions such as Gaussian and triangular types are usually used, which do not match the skewed characteristics that process parameter data often exhibit in actual industrial processes;

[0004] 2. Standard particle swarm optimization algorithms are prone to getting stuck in local optima in the later stages of optimization and are difficult to adapt to the characteristics of the process.

[0005] 3. In the weighted average defuzzification method, the weights of each rule are usually fixed, which makes it impossible to evaluate and adjust the contribution of different rules to control decisions based on real-time operating conditions, thus limiting the flexibility of the controller. Summary of the Invention

[0006] To address the problems mentioned in the background art, this invention proposes an intelligent proportioning control method for the synthesis process of antistatic agents, comprising the following steps:

[0007] The real-time and historical process parameters of the antistatic agent synthesis process are obtained; based on the skewness and kurtosis of the data distribution of the historical process parameters, an asymmetric membership function is constructed for the input variables, and the real-time process parameters are fuzzified using the asymmetric membership function;

[0008] Data samples from the historical process parameters are mapped to graph theory nodes. Associations are constructed based on the Euclidean distance between the data samples in the feature space. A community detection algorithm is used to divide the graph theory nodes into multiple communities. The core features of each community are extracted to generate a fuzzy rule. The initial weight of the fuzzy rule is set according to the modularity of the community, thereby constructing an initial fuzzy rule library.

[0009] For the fuzzy rules in the initial fuzzy rule base, the recursive least squares method is used to identify the consequent parameters, and redundant fuzzy rules are merged based on the cosine similarity between the consequent parameter vectors; the particle swarm optimization algorithm is used to optimize the antecedent parameters.

[0010] The real-time process parameters after the above fuzzification process are input into the fuzzy neural network obtained by the above optimization, and the activation intensity of each fuzzy rule is calculated. Based on the information entropy of the activation intensity of all current fuzzy rules, the weight of each fuzzy rule in the weighted averaging process is adjusted. The result after weighted averaging is defuzzified to obtain the proportion control amount of each component of the antistatic agent.

[0011] Optionally, the construction of asymmetric membership functions for the input variables based on the skewness and kurtosis of the data distribution of the historical process parameters includes:

[0012] An asymmetric Gaussian membership function is used, with the center of the function set as the mean of historical process parameters. The relative magnitudes of the left and right standard deviations are determined based on the skewness S, and the overall scale of the standard deviation is adjusted based on the kurtosis K and the reference kurtosis value.

[0013] Optionally, the step of extracting the core features of each community to generate a fuzzy rule includes:

[0014] The Louvain algorithm is used to divide the graph theory nodes into communities. For each community, the arithmetic mean of all data samples in the community is calculated on each feature dimension. The resulting mean vector is used as the core feature of the community and as the center of the membership function of the newly generated fuzzy rule antecedent.

[0015] Optionally, setting the initial weights of the fuzzy rules based on the modularity of the community includes:

[0016] Calculate the contribution of each partitioned community i to the overall network modularity. and use the formula Assign values ​​to the initial weights, where Let be the initial weight of the i-th fuzzy rule. The sum of contributions to all community modules.

[0017] Optionally, the fuzzy rule for merging redundancy based on cosine similarity between consequent parameter vectors includes:

[0018] Calculate the cosine similarity between the consequent parameter vectors of any two rules in the initial fuzzy rule base. When the cosine similarity is greater than a preset similarity threshold, merge the two rules. The consequent parameter of the merged new rule is the arithmetic mean of the consequent parameters of the original two rules, and the antecedent parameter is the weighted average of the initial weights of the original two rules.

[0019] Optionally, the inertia weights and acceleration constants of the particle swarm optimization algorithm are adjusted based on the information entropy of the activation intensity of all fuzzy rules in the previous control cycle, including:

[0020] Calculate the normalized activation intensity of all fuzzy rules in the previous control cycle. Based on the formula Calculate the information entropy H, and use the following formula to calculate the inertia weight for the current cycle. and acceleration constant , : , , ;

[0021] Where a, b, c, d, and e are preset parameters.

[0022] Optionally, calculating the activation strength of each fuzzy rule includes:

[0023] For each input variable, the asymmetric membership function is used to calculate the membership degree between the variable and the membership function center of each fuzzy rule antecedent, and the membership degrees of all input variables in a single rule are algebraically multiplied to obtain the activation strength of the fuzzy rule.

[0024] Optionally, adjusting the weights of each fuzzy rule in the weighted averaging process based on the information entropy of the activation strengths of all current fuzzy rules includes:

[0025] The activation intensities of all current fuzzy rules are normalized to obtain the normalized activation intensities. Calculate the information entropy H1 of the current activation strength, and introduce an exponential factor that depends on entropy. Where N is the total number of rules. The preset positive adjustment coefficient is used; the formula is employed. Calculate the intermediate weights, and then... Normalization is performed to obtain the weights used for the weighted average. .

[0026] Furthermore, this invention also relates to an intelligent proportioning control system for the synthesis process of antistatic agents, comprising the following modules:

[0027] The acquisition module is used to acquire real-time and historical process parameters of the antistatic agent synthesis process; based on the skewness and kurtosis of the data distribution of the historical process parameters, an asymmetric membership function is constructed for the input variables, and the real-time process parameters are fuzzified using the asymmetric membership function;

[0028] The module is used to map data samples in the historical process parameters to graph theory nodes, build associations based on the Euclidean distance of the data samples in the feature space, and use a community detection algorithm to divide the graph theory nodes into multiple communities; extract the core features of each community to generate a fuzzy rule, and set the initial weight of the fuzzy rule according to the modularity of the community, thereby constructing an initial fuzzy rule library;

[0029] The adjustment module is used to identify consequent parameters for the fuzzy rules in the initial fuzzy rule base using the recursive least squares method, and to merge redundant fuzzy rules based on the cosine similarity between consequent parameter vectors; and to optimize the antecedent parameters using the particle swarm optimization algorithm.

[0030] The calculation module is used to input the real-time process parameters after the above fuzzification process into the fuzzy neural network obtained by the above optimization, calculate the activation intensity of each fuzzy rule; adjust the weight of each fuzzy rule in the weighted averaging process based on the information entropy of the activation intensity of all current fuzzy rules; and defuzzify the result after weighted averaging to obtain the proportion control amount of each component of the antistatic agent.

[0031] Preferably, the construction of asymmetric membership functions for input variables based on the skewness and kurtosis of the data distribution of the historical process parameters includes:

[0032] An asymmetric Gaussian membership function is used, with the center of the function set as the mean of historical process parameters. The relative magnitudes of the left and right standard deviations are determined based on the skewness S, and the overall scale of the standard deviation is adjusted based on the kurtosis K and the reference kurtosis value.

[0033] Preferably, the step of extracting the core features of each community to generate a fuzzy rule includes:

[0034] The Louvain algorithm is used to divide the graph theory nodes into communities. For each community, the arithmetic mean of all data samples in the community is calculated on each feature dimension. The resulting mean vector is used as the core feature of the community and as the center of the membership function of the newly generated fuzzy rule antecedent.

[0035] Preferably, setting the initial weights of the fuzzy rules based on the modularity of the community includes:

[0036] Calculate the contribution of each partitioned community i to the overall network modularity. and use the formula Assign values ​​to the initial weights, where Let be the initial weight of the i-th fuzzy rule. The sum of contributions to all community modules.

[0037] Preferably, the fuzzy rule for merging redundancy based on cosine similarity between consequent parameter vectors includes:

[0038] Calculate the cosine similarity between the consequent parameter vectors of any two rules in the initial fuzzy rule base. When the cosine similarity is greater than a preset similarity threshold, merge the two rules. The consequent parameter of the merged new rule is the arithmetic mean of the consequent parameters of the original two rules, and the antecedent parameter is the weighted average of the initial weights of the original two rules.

[0039] Preferably, the inertia weights and acceleration constants of the particle swarm optimization algorithm are adjusted based on the information entropy of the activation intensity of all fuzzy rules in the previous control cycle, including:

[0040] Calculate the normalized activation intensity of all fuzzy rules in the previous control cycle. Based on the formula Calculate the information entropy H, and use the following formula to calculate the inertia weight for the current cycle. and acceleration constant , : , , ;

[0041] Where a, b, c, d, and e are preset parameters.

[0042] Preferably, calculating the activation strength of each fuzzy rule includes:

[0043] For each input variable, the asymmetric membership function is used to calculate the membership degree between the variable and the membership function center of each fuzzy rule antecedent, and the membership degrees of all input variables in a single rule are algebraically multiplied to obtain the activation strength of the fuzzy rule.

[0044] Preferably, adjusting the weights of each fuzzy rule in the weighted averaging process based on the information entropy of the activation strengths of all current fuzzy rules includes:

[0045] The activation intensities of all current fuzzy rules are normalized to obtain the normalized activation intensities. Calculate the information entropy H1 of the current activation strength, and introduce an exponential factor that depends on entropy. Where N is the total number of rules. The preset positive adjustment coefficient is used; the formula is employed. Calculate the intermediate weights, and then... Normalization is performed to obtain the weights used for the weighted average. .

[0046] This invention proposes an intelligent proportioning control method for the synthesis process of antistatic agents. By constructing an asymmetric membership function that reflects the skewness of actual data distribution, the accuracy of the model's representation of process parameter characteristics is improved. A community discovery algorithm is used to automatically explore and generate a structurally sound initial fuzzy rule base from historical data. Feedback information from the previous control cycle is used to adjust the search strategy of the optimization algorithm, improving the globality and convergence efficiency of parameter optimization. Furthermore, the contribution weight of each rule to the result is adjusted according to the activation status of each rule under the current operating conditions, making the control output closer to the actual process requirements, thereby improving the accuracy of proportioning control and the stability of the synthesis process. Attached Figure Description

[0047] Figure 1 A flowchart of the first embodiment; Figure 2 This is a schematic diagram of an asymmetric Gaussian membership function constructed based on skewness and kurtosis; Figure 3 This is a schematic diagram of the division and extraction of node communities in graph theory. Detailed Implementation

[0048] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0049] In the first embodiment, the present invention proposes an intelligent proportioning control method for the synthesis process of antistatic agents, such as... Figure 1 This includes the following steps:

[0050] S1, Obtain real-time and historical process parameters for the synthesis of antistatic agents; Based on the skewness and kurtosis of the data distribution of the historical process parameters, construct an asymmetric membership function for the input variables, and use the asymmetric membership function to fuzzify the real-time process parameters;

[0051] Specifically, a distributed control system (DCS) collects and stores process parameters such as reaction temperature, reaction pressure, stirring rate, catalyst concentration, and flow rate of each raw material in the antistatic agent synthesis reactor, forming a historical database. The system also acquires the current process parameters in real time. For each input variable in the historical database, such as reaction temperature, the skewness and kurtosis statistics of the data sample are calculated. Based on the calculated skewness and kurtosis, an asymmetric Gaussian membership function is constructed for the input variable of reaction temperature. This function has different width and shape parameters on both sides to accurately match the asymmetric distribution characteristics of the historical data. The real-time collected reaction temperature value, for example, 85℃, is input to multiple corresponding asymmetric membership functions, such as low temperature, medium temperature, and high temperature. The membership degree of the real-time temperature value to the three fuzzy sets is calculated, completing the fuzzification process.

[0052] In an optional embodiment, constructing an asymmetric membership function for the input variables based on the skewness and kurtosis of the data distribution of the historical process parameters includes:

[0053] An asymmetric Gaussian membership function is used, with the center of the function set as the mean of historical process parameters. The relative magnitudes of the left and right standard deviations are determined based on the skewness S, and the overall scale of the standard deviation is adjusted based on the kurtosis K and the reference kurtosis value.

[0054] Historical data on a given input variable, such as reaction temperature, is collected, assuming 1000 data points are gathered. The statistical characteristics of this data set are calculated, yielding a mean μ of 85℃, skewness S of 1.2, and kurtosis K of 4.5. The center c of the asymmetric Gaussian membership function is then set to the mean μ, i.e., 85℃. Since the skewness S is a positive value of 1.2, indicating a right-tailed distribution, the right-hand standard deviation of the membership function is... It should be greater than the left standard deviation. The relative magnitude of the left and right standard deviations can be set as a function related to skewness, for example... ,Right now Equal to 2.2 times This is to reflect the asymmetry in data distribution.

[0055] The reference value for kurtosis in a normal Gaussian distribution is 3, while the calculated kurtosis K is 4.5, which is greater than 3. This indicates that the actual data distribution is more keeled than a standard normal distribution. The overall size of the standard deviation needs to be scaled. For example, a baseline standard deviation can be calculated based on the sample variance. The initial value is 5. Next, a scaling factor based on kurtosis is entered. This factor can be the reference kurtosis value of 3 divided by the actual kurtosis value of 4.5, approximately 0.67. Multiplying the base standard deviation by this scaling factor yields the adjusted scaling baseline, approximately 3.35. By simultaneously solving the relative relationship determined by skewness and the scaling baseline determined by kurtosis, the left and right standard deviations can be obtained. For example, the solution might be... Approximately 2.1, Approximately 4.6. Therefore, the asymmetric Gaussian membership function of this input variable is completely determined, as follows: Figure 2 .

[0056] In an optional embodiment, the step of fuzzifying the real-time process parameters using the asymmetric membership function includes:

[0057] The real-time collected reaction temperature value, such as 85℃, is input into multiple corresponding asymmetric membership functions, such as low temperature, medium temperature, and high temperature, to calculate the membership degree of the real-time temperature value to the three fuzzy sets, thus completing the fuzzification.

[0058] For example, the parameters of the asymmetric Gaussian membership function for a given reaction temperature are the center of the intermediate temperature function. Left standard deviation right-hand standard deviation Simultaneously, the parameters of the low-temperature and high-temperature asymmetric Gaussian membership functions are defined, with the low-temperature function center being... skewness kurtosis Calculated , High-Temperature Function Center skewness kurtosis Calculated , The formula for calculating the asymmetric Gaussian membership function is: When At that time, membership degree

[0059] ;when hour, Real-time temperature Substitute into the three functions respectively to calculate: Substitute into the intermediate temperature function, because ,have to Substituting into the low-temperature function, Calculated Substituting into the high-temperature function, Calculated Ultimately obtained The membership degrees of the fuzzy sets for low temperature, medium temperature, and high temperature are 0.05, 1.0, and 0.007, respectively, thus completing the fuzzification.

[0060] S2, map the data samples in the historical process parameters to graph theory nodes, construct associations based on the Euclidean distance of the data samples in the feature space, and use a community detection algorithm to divide the graph theory nodes into multiple communities; extract the core features of each community to generate a fuzzy rule, and set the initial weight of the fuzzy rule according to the modularity of the community, thereby constructing an initial fuzzy rule library;

[0061] Each multidimensional data sample in the historical database, such as a combination of temperature, pressure, and flow rate at a given moment, is considered a graph theory node. The Euclidean distance between any two nodes in the multidimensional feature space is calculated. When the distance between two nodes is less than a preset threshold, an edge is established between them, thus constructing an undirected graph. The Louvain community detection algorithm is used to partition this graph. This algorithm iteratively optimizes the modularity index, aggregating closely connected data nodes into the same community, resulting in multiple data communities. For each partitioned community, the centroid of all data nodes within that community is calculated, i.e., the average value of each dimension of the features. This centroid vector constitutes the antecedent of the fuzzy rule; for example, if the temperature is 80℃ and the pressure is 1.5MPa, the output is Y. The modularity value of the community is used as the initial weight of the corresponding fuzzy rule. The higher the modularity of the community, the clearer the data pattern, and the higher the initial weight of the generated rule. Figure 3 .

[0062] In an optional embodiment, the extraction of core features for each community to generate a fuzzy rule includes:

[0063] The Louvain algorithm is used to divide the graph theory nodes into communities. For each community, the arithmetic mean of all data samples in the community is calculated on each feature dimension. The resulting mean vector is used as the core feature of the community and as the center of the membership function of the newly generated fuzzy rule antecedent.

[0064] Construct a graph network where each node represents a historical process data sample. Assume there are 2000 samples, each containing three feature dimensions: temperature, pressure, and reactant concentration. The weight of an edge between any two nodes is determined by their similarity in the feature space, for example, by using a Gaussian kernel function based on the Euclidean distance between the feature vectors of the two samples. A higher weight edge indicates that the two process data samples represent highly similar operating conditions.

[0065] After constructing the graph network, the Louvain community discovery algorithm is applied. This algorithm iteratively optimizes the network's modularity index, dividing the network into several tightly connected node clusters, i.e., communities. The iterative process involves two phases: first, traversing each node and assigning it to neighboring communities to maximize local modularity gain; second, treating each community as a supernode, reconstructing the network, and repeating the first phase of optimization until the modularity no longer increases, ultimately outputting several tightly connected node clusters, i.e., communities.

[0066] Suppose the algorithm identifies 5 communities, named respectively. arrive For example, communities It may contain 350 data samples, which are highly clustered in the feature space, collectively representing a specific type of high-temperature and high-pressure production conditions. Core features are extracted for each community. For each community... Calculate the arithmetic mean of the 350 samples contained in the community across three dimensions: temperature, pressure, and reactant concentration. If the result is a vector [120.5, 2.8, 0.75], then this vector is considered the community. The core feature vectors are used as the membership function centers for the antecedent of the first fuzzy rule. Similarly, core feature vectors are calculated for the remaining four communities, resulting in a total of five initial fuzzy rules.

[0067] In an optional embodiment, setting the initial weights of the fuzzy rules based on the modularity of the community includes:

[0068] Calculate the contribution of each partitioned community i to the overall network modularity. and use the formula Assign values ​​to the initial weights, where Let be the initial weight of the i-th fuzzy rule. The sum of contributions to all community modules.

[0069] After using the Louvain algorithm to partition the network into communities, the algorithm outputs a modularity Q-value for the entire network. This value is the sum of the modularity contributions of all communities. Modularity represents the tightness of connections within a community compared to random expectations. Each community i contributes its own modularity to the overall network. It can be calculated separately, reflecting the strength of the connections within the community and its correlation with the overall network structure. The higher a community's modularity contribution value, the more important the community's structure is within the entire network.

[0070] Assuming that the modularity contribution values ​​of the five communities in the previous embodiment are calculated as follows: =0.18, =0.15, =0.11, =0.09, and =0.07. Calculate the sum of the contribution values ​​of all community modules, i.e. =0.60. An initial weight is assigned to each fuzzy rule according to the formula. The weight of the first rule is... The weight is 0.30. This is the weight of the second rule. The initial weight is 0.25. Similarly, the initial weights for all five rules are calculated to be 0.30, 0.25, 0.183, 0.15, and 0.117. These weights reflect the relative importance of rules derived from each community; the community with the greater its modularity contribution, the more important its corresponding rules are in the initial model.

[0071] S3, For the fuzzy rules in the initial fuzzy rule base, the recursive least squares method is used to identify the consequent parameters, and redundant fuzzy rules are merged based on the cosine similarity between the consequent parameter vectors; the particle swarm optimization algorithm is used to optimize the antecedent parameters, wherein the inertia weight and acceleration constant of the particle swarm optimization algorithm are adjusted according to the information entropy of the activation intensity of all fuzzy rules in the previous control cycle;

[0072] Specifically, the linear parameters of the consequent of each rule in the rule base are identified. Using a recursive least squares algorithm, the consequent parameter matrix is ​​iteratively updated using historical data samples to minimize the error between the predicted and actual outputs. After identification, the cosine similarity between the consequent parameter vectors of any two fuzzy rules is calculated. If the similarity value is higher than a set threshold, such as 0.98, the two rules are considered to have similar effects and are redundant, thus merging them into a new rule. During merging, the consequent parameter is taken as the arithmetic mean of the consequent parameters of the original two rules, and the antecedent parameter is a weighted average of the initial weights of the original two rules. A particle swarm optimization algorithm is used to optimize the antecedent parameters of the rules, i.e., the center and width of the asymmetric membership function. The optimization objective is to systematically minimize the overall output error of the fuzzy neural network model on the entire historical dataset by adjusting the parameters of the asymmetric Gaussian membership function in the antecedent of each fuzzy rule. During the optimization process, the information entropy of the activation intensity of all rules in the previous control cycle is calculated. If the information entropy value is large, it indicates that the system is in an uncertain state. In this case, the inertia weight and global learning factor of the particle swarm algorithm are increased to enhance the global search capability. If the information entropy value is small, it indicates that the system is in a stable state. In this case, the inertia weight is reduced and the local influence of the learning factor is increased to accelerate local convergence.

[0073] To simplify the fuzzy rule base and eliminate functionally similar rules, in an optional embodiment, the merging of redundant fuzzy rules based on the cosine similarity between consequent parameter vectors includes:

[0074] Calculate the cosine similarity between the consequent parameter vectors of any two rules in the initial fuzzy rule base. When the cosine similarity is greater than a preset similarity threshold, merge the two rules. The consequent parameter of the merged new rule is the arithmetic mean of the consequent parameters of the original two rules, and the antecedent parameter is the weighted average of the initial weights of the original two rules.

[0075] Assuming the model is a TSK fuzzy model, the rule consequent is a linear combination of the input variables, in the form of: Each rule's consequent consists of a parameter vector. To define, the antecedent is determined by parameters of the asymmetric membership function, such as the center c and the left-hand standard deviation. right-hand standard deviation Definition. For a two-input system, the consequent vector of rule A. Weight Antecedent: Temperature ,pressure Rule B successor vector Weight Antecedent: Temperature ,pressure The cosine similarity of the consequent of ) is calculated as follows: Above the threshold Therefore, redundancy is determined and the rules are merged into a new rule C. Its parameters are generated by a weighted average: the consequent vector is the arithmetic mean of the two. The antecedent parameters are calculated using an initial weighted average, and the temperature center is... Where w1 and w2 are the initial weights of the original two rules, and the left standard deviation is... right standard deviation Pressure center Left standard deviation right standard deviation The initial weight of the new rule is set to... Iterate this process until the similarity of all rule pairs is below the threshold, thus obtaining a streamlined rule base.

[0076] In an optional embodiment, the inertia weights and acceleration constants of the particle swarm optimization algorithm are adjusted based on the information entropy of the activation intensity of all fuzzy rules in the previous control cycle, including:

[0077] Calculate the normalized activation intensity of all fuzzy rules in the previous control cycle. Based on the formula Calculate the information entropy H, and use the following formula to calculate the inertia weight for the current cycle. and acceleration constant , : , , ;

[0078] Where a, b, c, d, and e are preset parameters.

[0079] The parameters of the particle swarm optimization algorithm are adjusted to optimize the antecedent parameters of the fuzzy system. At the end of each control cycle, the activation intensity of all fuzzy rules within that cycle is obtained. Assume there are four rules with activation intensities of 0.7, 0.6, 0.1, and 0.05. These intensities are normalized so that their sum equals 1, yielding the normalized activation intensity. The vector is approximately [0.48, 0.41, 0.07, 0.03].

[0080] The information entropy H is calculated using the normalized activation intensity described above. When multiple rules are activated with similar intensities, such as 0.48 and 0.41 in this example, the uncertainty is high, and the calculated information entropy H value is also large; let's assume H is 1.6. Conversely, if only one rule is activated, the uncertainty is low, and the H value will be small. The calculated entropy value H is then substituted into the given formula to update the inertia weight ω and acceleration constant of the particle swarm optimization algorithm. and The preset parameters a, b, c, d, and e can be set to 0.4, 0.5, 2.5, 2.0, and 0.5 respectively. When the entropy H is large, such as 1.6, the fractional... The value of ω will approach 1, causing ω to increase. Decrease Increasing the entropy H makes particles more inclined to explore the global search space, helping them escape local optima and adapt to the current uncertain state of the system. When the entropy H is small, ω decreases. Increase Decreasing the size of the particle encourages it to converge toward its individual optimal and global optimal positions.

[0081] S4. Input the real-time process parameters after the above fuzzification process into the fuzzy neural network obtained by the above optimization, and calculate the activation intensity of each fuzzy rule; adjust the weight of each fuzzy rule in the weighted averaging process based on the information entropy of the activation intensity of all current fuzzy rules; defuzzify the result after weighted averaging to obtain the proportion control amount of each component of the antistatic agent.

[0082] Specifically, the fuzzified real-time process parameters are input into the optimized fuzzy neural network. The activation intensity of each fuzzy rule at the current moment is obtained by multiplying the membership degrees of the input parameters in the membership functions of each rule's antecedents. The information entropy of the activation intensities of all rules is calculated. Based on this information entropy value, the contribution weight of each rule in the decision-making process is adjusted. When the information entropy is low, it indicates that a few rules dominate the current operating condition, so the weights of the rules with high activation intensities are further increased. When the information entropy is high, it indicates that multiple rules are working together, so the weights of each rule are appropriately balanced to achieve a smooth transition. The consequent output of each rule is multiplied by the adjusted weights and summed, then divided by the sum of all weights. Defuzzification is performed using a weighted average method to obtain a precise value, which is the optimal ratio control amount of each component of the antistatic agent under the current operating condition. This optimal ratio control amount is then output to the actuator.

[0083] The fuzzy neural network described is a TSK-type fuzzy neural network adapted to the synthesis process of antistatic agents. Its structure follows a sequence of fuzzification layer → rule layer → inference layer → defuzzification layer. The fuzzification layer constructs an asymmetric Gaussian membership function based on the skewness and kurtosis of historical process parameters, converting real-time temperature, pressure, and other parameters into fuzzy membership degrees. The rule layer stores fuzzy rules automatically generated and simplified using a graph theory community discovery algorithm. Each rule includes antecedent membership function parameters, consequent linear parameters, and initial weights. The inference layer calculates the activation intensity of each rule through membership degree products and dynamically adjusts the rule weights based on information entropy. The defuzzification layer uses a weighted average method for defuzzification, generating precise proportioning control quantities. The optimization process involves: first, clustering similar working conditions from historical data using community discovery to generate an initial rule base; then, using recursive least squares to identify consequent linear parameters online; merging redundant rules based on the cosine similarity of consequent parameter vectors; and finally, employing a particle swarm optimization algorithm to adaptively adjust the inertia weights and acceleration constants based on the information entropy of the activation intensity from the previous control cycle, thus optimizing the antecedent membership function parameters.

[0084] In an optional embodiment, calculating the activation strength of each fuzzy rule includes:

[0085] For each input variable, the asymmetric membership function is used to calculate the membership degree between the variable and the membership function center of each fuzzy rule antecedent, and the membership degrees of all input variables in a single rule are algebraically multiplied to obtain the activation strength of the fuzzy rule.

[0086] This describes how to calculate the degree to which each rule in a fuzzy system is triggered for a given real-time input. Assume there are two input variables: temperature... and pressure The current input value is [90, 1.4]. Consider the first rule in the rule base, where the antecedent is if the temperature... yes And pressure yes .in and It is a fuzzy set defined by the previously constructed asymmetric Gaussian membership function.

[0087] Processing input variable temperature Assuming fuzzy sets The center value is 85, the left standard deviation is 3, and the right standard deviation is 4.5. Since the current input value of 90 is greater than the center value of 85, a right standard deviation of 4.5 is chosen to calculate the membership degree. The temperature is calculated using the Gaussian function formula. For fuzzy sets The membership degree is approximately 0.59. (This refers to handling the input variable pressure.) Assuming fuzzy sets The center is 1.5, the left standard deviation is 0.2, and the right standard deviation is 0.3. Since the current input value of 1.4 is less than the center 1.5, a left standard deviation of 0.2 is used to calculate the membership degree. The pressure is then calculated. For fuzzy sets The membership degree is approximately 0.78. Since the antecedent of the rule is connected by a logical AND, the activation strength of the rule is obtained by algebraically multiplying the membership degrees of all input variables. In this example, the activation strength of the first rule is approximately 0.46.

[0088] In an optional embodiment, adjusting the weights of each fuzzy rule in the weighted averaging process based on the information entropy of the activation strengths of all current fuzzy rules includes:

[0089] The activation intensities of all current fuzzy rules are normalized to obtain the normalized activation intensities. Calculate the information entropy of the current activation strength. Introduce an exponential factor that depends on entropy. ,in For the total number of rules, The preset positive adjustment coefficient is used; the formula is employed. Calculate the intermediate weights, and then... Normalization is performed to obtain the weights used for the weighted average. .

[0090] For example, for a system with 4 rules, if the activation strength is normalized... for The information entropy was calculated. ;Pick ,but ; Calculate the intermediate weights The final weights are obtained after normalization. This weight distribution is more balanced than the original distribution, achieving the intention of balancing the contributions of each rule under high uncertainty; conversely, when the entropy is small, This will cause the weight distribution to concentrate on high Regarding the rules, strengthen the contribution of the dominant rules. Ultimately, use... The control quantity is obtained by weighted averaging of the consequent outputs of each rule and then defuzzifying it.

[0091] In a second embodiment, the present invention provides an intelligent proportioning control system for the synthesis process of antistatic agents, comprising the following modules:

[0092] The acquisition module is used to acquire real-time and historical process parameters of the antistatic agent synthesis process; based on the skewness and kurtosis of the data distribution of the historical process parameters, an asymmetric membership function is constructed for the input variables, and the real-time process parameters are fuzzified using the asymmetric membership function;

[0093] The module is used to map data samples in the historical process parameters to graph theory nodes, build associations based on the Euclidean distance of the data samples in the feature space, and use a community detection algorithm to divide the graph theory nodes into multiple communities; extract the core features of each community to generate a fuzzy rule, and set the initial weight of the fuzzy rule according to the modularity of the community, thereby constructing an initial fuzzy rule library;

[0094] The adjustment module is used to identify consequent parameters for the fuzzy rules in the initial fuzzy rule base using the recursive least squares method, and to merge redundant fuzzy rules based on the cosine similarity between consequent parameter vectors; and to optimize the antecedent parameters using the particle swarm optimization algorithm.

[0095] The calculation module is used to input the real-time process parameters after the above fuzzification process into the fuzzy neural network obtained by the above optimization, calculate the activation intensity of each fuzzy rule; adjust the weight of each fuzzy rule in the weighted averaging process based on the information entropy of the activation intensity of all current fuzzy rules; and defuzzify the result after weighted averaging to obtain the proportion control amount of each component of the antistatic agent.

[0096] In this specification, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise limited, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the associated listed items.

[0097] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0098] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent proportioning control in the synthesis process of an antistatic agent, characterized in that, Includes the following steps: The real-time and historical process parameters of the antistatic agent synthesis process are obtained; based on the skewness and kurtosis of the data distribution of the historical process parameters, an asymmetric membership function is constructed for the input variables, and the real-time process parameters are fuzzified using the asymmetric membership function; The data samples in the historical process parameters are mapped to graph theory nodes, and the associations are constructed based on the Euclidean distance of the data samples in the feature space. The community detection algorithm is then used to divide the graph theory nodes into multiple communities. The core features of each community are extracted to generate a fuzzy rule, and the initial weight of the fuzzy rule is set according to the modularity of the community, thereby constructing an initial fuzzy rule library; For the fuzzy rules in the initial fuzzy rule base, the recursive least squares method is used to identify the consequent parameters, and redundant fuzzy rules are merged based on the cosine similarity between the consequent parameter vectors; the particle swarm optimization algorithm is used to optimize the antecedent parameters. The real-time process parameters after the above fuzzification process are input into the fuzzy neural network obtained by the above optimization, and the activation intensity of each fuzzy rule is calculated. Based on the information entropy of the activation intensity of all current fuzzy rules, the weights of each fuzzy rule in the weighted averaging process are adjusted; the result after weighted averaging is defuzzified to obtain the proportion control amount of each component of the antistatic agent.

2. The method according to claim 1, characterized in that, The skewness and kurtosis of the data distribution based on the historical process parameters are used to construct asymmetric membership functions for the input variables, including: An asymmetric Gaussian membership function is used, with the center of the function set as the mean of historical process parameters. The relative magnitudes of the left and right standard deviations are determined based on the skewness S, and the overall scale of the standard deviation is adjusted based on the kurtosis K and the reference kurtosis value.

3. The method according to claim 1, characterized in that, The process of extracting the core features of each community to generate a fuzzy rule includes: The Louvain algorithm is used to divide the graph theory nodes into communities. For each community, the arithmetic mean of all data samples in the community is calculated on each feature dimension. The resulting mean vector is used as the core feature of the community and as the center of the membership function of the newly generated fuzzy rule antecedent.

4. The method according to claim 1, characterized in that, The step of setting the initial weights of the fuzzy rules based on the modularity of the community includes: Calculate the contribution of each partitioned community i to the overall network modularity. and use the formula Assign values ​​to the initial weights, where Let be the initial weight of the i-th fuzzy rule. The sum of contributions to all community modules.

5. The method according to claim 1, characterized in that, The fuzzy rules for merging redundancy based on cosine similarity between consequent parameter vectors include: Calculate the cosine similarity between the consequent parameter vectors of any two rules in the initial fuzzy rule base. When the cosine similarity is greater than a preset similarity threshold, merge the two rules. The consequent parameter of the merged new rule is the arithmetic mean of the consequent parameters of the original two rules, and the antecedent parameter is the weighted average of the initial weights of the original two rules.

6. The method according to claim 1, characterized in that, The inertia weights and acceleration constants of the particle swarm optimization algorithm are adjusted based on the information entropy of the activation intensity of all fuzzy rules in the previous control cycle, including: Calculate the normalized activation intensity of all fuzzy rules in the previous control cycle. Calculate the information entropy H, and use the following formula to calculate the inertia weight for the current cycle. and acceleration constant , : , , ; Where a, b, c, d, and e are preset parameters.

7. The method according to claim 1, characterized in that, The calculation of the activation strength of each fuzzy rule includes: For each input variable, the asymmetric membership function is used to calculate the membership degree between the variable and the membership function center of each fuzzy rule antecedent, and the membership degrees of all input variables in a single rule are algebraically multiplied to obtain the activation strength of the fuzzy rule.

8. The method according to claim 1, characterized in that, The adjustment of the weights of each fuzzy rule in the weighted averaging process based on the information entropy of the activation strengths of all current fuzzy rules includes: The activation intensities of all current fuzzy rules are normalized to obtain the normalized activation intensities. Calculate the information entropy H1 of the current activation strength, and introduce an exponential factor that depends on entropy. Where N is the total number of rules. The preset positive adjustment coefficient is used; the formula is employed. Calculate the intermediate weights, and then... Normalization is performed to obtain the weights used for the weighted average. .

9. An intelligent proportioning control system for the synthesis process of an antistatic agent, characterized in that, Includes the following modules: The acquisition module is used to acquire real-time and historical process parameters of the antistatic agent synthesis process; based on the skewness and kurtosis of the data distribution of the historical process parameters, an asymmetric membership function is constructed for the input variables, and the real-time process parameters are fuzzified using the asymmetric membership function; The construction module is used to map data samples in the historical process parameters to graph theory nodes, construct associations based on the Euclidean distance of the data samples in the feature space, and use a community detection algorithm to divide the graph theory nodes into multiple communities. The core features of each community are extracted to generate a fuzzy rule, and the initial weight of the fuzzy rule is set according to the modularity of the community, thereby constructing an initial fuzzy rule library; The adjustment module is used to identify consequent parameters for the fuzzy rules in the initial fuzzy rule base using the recursive least squares method, and to merge redundant fuzzy rules based on the cosine similarity between consequent parameter vectors; and to optimize the antecedent parameters using the particle swarm optimization algorithm. The calculation module is used to input the real-time process parameters after the above fuzzification process into the fuzzy neural network obtained by the above optimization, and calculate the activation intensity of each fuzzy rule. Based on the information entropy of the activation intensity of all current fuzzy rules, the weights of each fuzzy rule in the weighted averaging process are adjusted; the result after weighted averaging is defuzzified to obtain the proportion control amount of each component of the antistatic agent.

10. The system according to claim 9, characterized in that, The skewness and kurtosis of the data distribution based on the historical process parameters are used to construct asymmetric membership functions for the input variables, including: An asymmetric Gaussian membership function is used, with the center of the function set as the mean of historical process parameters. The relative magnitudes of the left and right standard deviations are determined based on the skewness S, and the overall scale of the standard deviation is adjusted based on the kurtosis K and the reference kurtosis value.

Images