Multi-process parameter optimization method

Abstract

The invention relates to the technical field of process quality control, in particular to a multi-process parameter optimization method, which comprises the following steps: acquiring raw material quality attributes, process parameters and finished product quality attributes, and screening from the finished product quality attributes to obtain key finished product quality attributes; obtaining an original data set based on the key finished product quality attributes, and constructing a causal relationship graph network between the key finished product quality attributes and corresponding raw material quality attributes and process parameters based on the original data; training and constructing a corresponding nonlinear correlation model based on the causal relationship graph; and obtaining optimized process parameters including raw material quality attributes and process parameters based on the nonlinear correlation model. Through causal relationship analysis and a nonlinear correlation model, the limitations of redundancy, unknown direction and the like caused by dependence on correlation only are solved, variable interaction is more accurately captured in the multi-process production process, and therefore the overall optimization capability is remarkably improved.

Description

Technical Field

[0001] The invention belongs to the technical field of process quality control, and in particular relates to a multi-process parameter optimization method. Background Art

[0002] The production process of de-walled Ganoderma lucidum spore powder involves multiple unit processes, the input of each process is raw materials, and the output is the intermediate or final product of the next process. During the production process, the fluctuations of raw material quality attributes (input quality, IQ) and process parameters (process parameter, PP) will gradually accumulate along the production process and be transmitted to the quality attributes of the finished product (output quality, OQ).

[0003] Adjusting the control range of process parameters during the production process can effectively reduce the fluctuation of finished product quality attributes. Traditional production parameter optimization methods rely on experience-driven statistical methods, such as the One-Factor-at-a-Time (OFAT), which gradually adjusts process parameters through multiple experiments. This method cannot consider the interaction between multiple variables, resulting in low optimization efficiency and easy to ignore the global optimal solution, making it difficult to adapt to complex and changing production processes. Therefore, for the process of multi-process production flow, the most practical improvement method is to collect historical batch production data and establish a mathematical model to describe the interaction between variables.

[0004] For chemical drugs, the internal material transformation mechanism of the production process is clear, and physical property equations such as mass, energy, momentum and phase equilibrium can be used to independently develop mechanism models for each unit process. By simply connecting these mechanism models in series, the dynamic operation of the entire production process can be simulated. However, for biological drugs such as de-walled Ganoderma lucidum spore powder, since the occurrence and development mechanism of the process is not well understood, the established process model is often a data-driven model. It is impossible to first establish a model for a single process separately, and then connect these independent models in series to establish a comprehensive model.

[0005] In view of this, the present invention is proposed. Summary of the invention

[0006] In order to solve the above-mentioned technical problems existing in the prior art, the present invention provides a multi-process parameter optimization method, which solves the problem that independent models cannot be connected in series to establish a comprehensive model, and can more accurately capture variable interactions in a multi-process production process, thereby significantly improving the overall optimization capability.

[0007] To achieve the above object, the technical solution of the present invention is as follows:

[0008] A multi-process parameter optimization method, characterized by comprising:

[0009] Obtaining an original data set, wherein the original data set includes key finished product quality attributes, raw material quality attributes, and process parameters; the key finished product quality attributes are obtained by screening based on the finished product quality attributes;

[0010] Constructing a cause-and-effect relationship diagram network between the key finished product quality attributes and the raw material quality attributes and process parameters based on the original data;

[0011] Using the causal relationship graph network and the raw material quality attributes and process parameters in the original data set as input parameters, and the key finished product quality attributes as output parameters, training and constructing a corresponding nonlinear association model;

[0012] Optimized process parameters including raw material quality attributes and process parameters are obtained based on the nonlinear correlation model.

[0013] Furthermore, the screening is achieved by combining a Shewhart control chart and a process capability index standard. The Shewhart control chart includes two control limits: a mean ±3 times standard deviation control limit of a finished product quality attribute and a control limit defined by a quality control standard. The screening of the key finished product quality attributes is performed based on the calculation results of the control limits and the process capability index standard.

[0014] Furthermore, the calculation formula of the process capability index standard is:

[0015]

[0016] Among them, T U is the upper tolerance limit, T L is the lower tolerance limit, s is the sample population standard deviation, is the mean of the measured values, 3s is three times the standard deviation, and []min is the minimum value in brackets.

[0017] Furthermore, before constructing the causal relationship graph network, the original data set is further subjected to sample expansion processing, and the expansion processing is implemented by the Bootstrap sampling algorithm. The specific calculation process includes:

[0018] Set the number of batch records sampled each time by the Bootstrap algorithm to N;

[0019] Random sampling is performed on the original data set, and after each sampling is completed, the N batch records are converted into a statistical vector by calculating the statistics of the batch records;

[0020] The statistics of the process parameter variables are recorded as the mean value, denoted as pp_mean; the statistics of the raw material quality attribute variables are denoted as P pk , denoted as iq_P pk; The statistics of the finished product quality attribute variables are recorded as oq_P pk .

[0021] Furthermore, the causal relationship graph network is constructed using a Bayesian network structure, and the causal relationship graph network is structurally optimized;

[0022] The structural optimization includes removing redundant nodes and edges in combination with a network structure optimization algorithm, wherein the optimization algorithm uses training set data and a Bayesian network structure optimization method.

[0023] Furthermore, the graph attention network algorithm is used to calculate the weights of the edges in the causal graph network. The specific weight calculation process includes:

[0024] Assume that the Bayesian network contains V nodes, and the feature vector of each node is denoted as h i ∈R D , where i = 1, 2, ... V, D is the dimension; using a shared weight matrix W∈R D*D′ For node h i and adjacent nodes h j Perform a linear transformation to obtain a new eigenvector h′ i and h′ j , where h′ i =Wh i ;

[0025] The feature vector h′ i and h′ j After concatenation, input the feedforward neural network Map it to a real number and activate it using the LeakyReLU function to get the attention correlation coefficient e ij , and its calculation formula is as follows:

[0026]

[0027] Among them, || represents vector concatenation, is a trainable parameter vector;

[0028] The attention correlation coefficient e ij Normalized by the softmax function, calculate the attention weight a between two nodes ij , and its calculation formula is:

[0029]

[0030] Among them, N i Represents all neighbor nodes of node i, according to the attention weight a ij Update the edge strength between node i and node j in the Bayesian network to quantify the strength of the causal relationship between variables.

[0031] Furthermore, the nonlinear association model is constructed using a graph convolutional neural network model, and the specific construction process includes:

[0032] The causal relationship graph network and the statistics of raw material quality attributes and process parameters are used as input. The graph convolutional neural network model includes: input layer, convolution, Flatten layer and full connection layer; the input layer node feature matrix is ​​H (0) , the node feature matrix H after convolutional layer mapping (1) , the propagation calculation formula from the input layer to the convolutional layer is:

[0033]

[0034] in, I is the identity matrix, A∈R n×n is the adjacency matrix of the graph, for The degree matrix of W (0) is the weight matrix to be trained, σ() is the activation function used by the convolutional layer, H (0) is the input layer node feature matrix;

[0035] The node feature matrix H (1) At the same time, the full connection layer is input to predict N key finished product quality attributes; the propagation rule from the convolutional layer to the full connection layer is:

[0036] H (i+i) =σ(H (i) W (i) )

[0037] Among them, H (i) is the node feature matrix of the i-th layer neural network, σ() is the activation function used by the convolutional layer, W (i) The weight matrix to be trained;

[0038] The Adam optimizer is used and the initial learning rate is set. The learning rate is halved every fixed training cycle. At the same time, the random inactivation probability of dropout is set, and the training is terminated early when the validation set loss does not decrease for a predetermined number of consecutive cycles. When the above conditions are met, the model training is considered complete.

[0039] Furthermore, the process parameters are obtained from a nonlinear association model using a genetic algorithm, and the mathematical model is:

[0040] SGA=(C,E,P0,M,φ,Γ,ψ,T)

[0041] Among them, C represents the individual encoding scheme, E represents the individual fitness evaluation function, P0 represents the initial population, M represents the population size, φ, Γ and ψ represent the selection operator, crossover operator and mutation operator respectively, and T represents the termination evolutionary generation.

[0042] Furthermore, the genetic algorithm specifically comprises the following steps:

[0043] Set P0, T, P Γ and P ψ , set the gene as the upper and lower control limits of the variable, the number of genes is twice the number of variables, and set the number of iterations t = 1;

[0044] The initial population is generated completely randomly, and then the chromosomes of the population are encoded using binary code to generate an initial population consisting of N chromosomes:

[0045] pop i (t), t=1, i=1, 2, …, N; when t=1, N=P0

[0046] Calculate the fitness evaluation function, calculate the fitness of each individual, the fitness of the i-th individual is f(i); let f(s) = minf(i), the s-th individual is the optimal individual of the population; calculate the selection probability based on the individual fitness value:

[0047]

[0048] in, is the sum of f(i), f(i) is the fitness of the ith one, and P i The probability distribution of randomly selecting some chromosomes from the current generation population to be inherited to the next generation population;

[0049] With probability P Γ Make the chromosomes cross over, and then with probability P ψ Make the chromosome gene mutate and form a new offspring pop i (t), and at the same time serves as the parent of the next genetic operation, iterating t+1 times and repeating;

[0050] When the number of iterations t reaches T, the algorithm terminates and outputs the optimal population.

[0051] Furthermore, the original data set also includes an original test set, and the optimized process parameters are verified based on the original test set.

[0052] Compared with the prior art, the multi-process parameter optimization method provided by the present invention overcomes the limitation that simple correlation analysis is difficult to grasp the real influencing factors by combining causal relationship modeling with graph attention network. At the same time, a complex nonlinear correlation model is established by using graph convolutional neural network, which can more accurately characterize the multi-variable coupling effect and cooperate with genetic algorithm to achieve global optimization, thereby significantly improving the quality consistency and controllability of multi-process production process, and meeting the needs of pharmaceutical and related industries for high-standard quality control. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A schematic diagram of a multi-process parameter optimization method provided by an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of the attention mechanism of the graph attention network provided by an embodiment of the present invention;

[0055] Figure 3 A graph convolutional neural network model architecture diagram provided for an embodiment of the present invention;

[0056] Figure 4 A Shewhart control chart of quality attributes provided by an embodiment of the present invention;

[0057] Figure 5 A schematic diagram of a Bayesian network after structural optimization provided by an embodiment of the present invention;

[0058] Figure 6 A schematic diagram of a weighted Bayesian network provided for an embodiment of the present invention;

[0059] Figure 7 A schematic diagram of the data distribution and optimal control range of process parameter statistics on a test set provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The technical solution of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention, and all other embodiments obtained by ordinary technicians in the field without making creative work are within the protection scope of the present invention.

[0061] It should be noted that, unless otherwise specifically stated, the relative arrangements of components and steps, and numerical expressions set forth in these embodiments should not be construed as limiting the scope of the present invention.

[0062] The following description of the exemplary embodiments is merely illustrative and is not intended to limit the present invention and its application or use in any sense. Techniques, methods and devices known to ordinary technicians in the relevant field may not be discussed in detail here, but where applicable, these techniques, methods and devices should be considered as part of this specification.

[0063] Explanation of professional terms:

[0064] iq: full name is input quality attribute;

[0065] oq: the full name is output quality;

[0066] pp: full name is process parameter;

[0067] GNN: The full name is Graph Neural Networks;

[0068] GA: full name is Genetic Algorithm;

[0069] Shewhart control chart: The full name is Shewhart (WA Shewhart) control chart;

[0070] GAT algorithm: the full name is Graph Attention Network algorithm;

[0071] GCN model: the full name is Graph Convolutional Neural Network model;

[0072] SGA algorithm: the full name is Simple Genetic Algorithms;

[0073] Bootstrap algorithm: the full name is the bootstrap method;

[0074] Embodiment 1

[0075] See also Figure 1 The present invention provides a multi-process parameter optimization method, which specifically relates to a production process of wall-removed Ganoderma lucidum spore powder. The control interval of process parameters can be adjusted during the production process to effectively reduce the fluctuation of the quality attributes of the finished product. The specific implementation process may include the following steps:

[0076] A1. Data collection and variable definition

[0077] In order to achieve comprehensive monitoring and optimization of multi-process production processes, it is necessary to systematically collect data on raw material quality attributes, process parameters and finished product quality attributes;

[0078] Among them, the quality attributes of raw materials are obtained through laboratory physical and chemical analysis, ensuring the accurate grasp of the initial quality of raw materials, such as: raw material moisture, total ash, peroxide value, triolein and crude polysaccharide content, etc.; the process parameters include: extraction temperature, extraction temperature, centrifugal filtration pressure, concentration temperature, concentration vacuum, drying temperature, drying vacuum, drying rapid moisture and crushing sieve mesh number, etc.; the finished product quality attributes include: moisture, fineness, crude polysaccharides (measured in glucose) and total triterpenes (measured in oleanolic acid), etc.

[0079] A2. Screening of key finished product quality attributes

[0080] The key finished product quality attributes related to the finished product quality are screened out from the finished product quality attributes; the finished product quality attributes include multiple physical and chemical indicators (such as moisture content, fineness, crude polysaccharide content and total triterpene content). In order to screen out the key finished product quality attributes that are highly related to the finished product quality, Shewhart control chart and process capability index (P pk ) criteria, and the screening results were combined for subsequent analysis.

[0081] The Shewhart control chart is a quality management tool proposed by Shewhart to determine whether the production process is in a controlled state. The Shewhart control chart used in the present invention sets two control limits:

[0082] Mean ± 3 times standard deviation control limit: that is, the mean of the quality attribute of the finished product plus or minus 3 times the standard deviation;

[0083] Control limits defined by factory quality control standards: upper and lower limits set according to the quality control standards established within the factory;

[0084] If a finished product quality attribute exceeds any of the control limits in the batch direction, it is considered that the finished product quality attribute is not well controlled and needs to be defined as a critical finished product quality attribute.

[0085] Process capability index P pk It comprehensively considers the degree of dispersion of data distribution and the degree of deviation of the mean of measured values ​​from the tolerance center, and is used to describe the process capability of the unit process. The calculation formula is as follows:

[0086]

[0087] Among them, T U is the upper tolerance limit, T L is the lower tolerance limit, s is the sample population standard deviation, is the mean of the measured values;

[0088] Refer to the evaluation criteria in Table 1. In many industries, P pk =1.33 is considered the minimum acceptable standard, P pk =1.67 is a higher quality requirement. For the pharmaceutical industry, quality control tends to be conservative. pk The standard of <1.67 helps to better control process fluctuations, thereby improving the quality consistency of the final product. pk <1.67, it is considered that the finished product quality attribute has a large quality fluctuation and needs to be defined as a key finished product quality attribute.

[0089] Table 1P pk Evaluation criteria

[0090]

[0091] A3. Sample expansion and statistical calculation

[0092] Based on the key finished product quality attributes, an original data set is obtained, wherein the original data set includes the key finished product quality attributes and its corresponding raw material quality attributes and process parameters; in order to fully explore the statistical characteristics of multiple batches of data and obtain a larger sample space that can be used for model training, a sampling algorithm is used to randomly sample the original data set, wherein the original data set includes an original training set and an original test set; and multiple batches of data of the raw material quality attributes, process parameters and key finished product quality attributes are processed to obtain mean statistics and statistics that can be used for model training, and describe the statistical characteristics of the parameters in multiple batches; the algorithm adopts the Bootstrap sampling algorithm, and the specific calculation process includes:

[0093] A31. Set the number of batch records sampled each time by the Bootstrap algorithm to N;

[0094] A32. After each sampling is completed, the N batch records are converted into a statistic vector by calculating the statistics of the batch records;

[0095] A33. The statistics of the process parameter variables are recorded as the mean value, denoted as pp_mean; the statistics of the raw material quality attribute variables are denoted as P pk , denoted as iq_P pk ; The statistics of the finished product quality attribute variables are recorded as oq_P pk .

[0096] Through the above method, the original multiple batches of data are processed to form a corresponding statistical matrix. Each sampling can capture the distribution differences and fluctuation characteristics between different batches, thereby providing more representative, robust and repeatable training samples for subsequent model training and parameter optimization.

[0097] A4. Bayesian network modeling and optimization

[0098] Based on the original data, a causal relationship graph network between key finished product quality attributes and their corresponding raw material quality attributes and process parameters is constructed; according to the sequence or causal relationship of the production process, a causal relationship graph network reflecting the influence of raw material quality attributes and process parameters on key finished product quality attributes is constructed; the causal relationship graph network is constructed using a Bayesian network structure, and the causal relationship graph network is structurally optimized;

[0099] See also Figure 5 The Bayesian network, also known as a directed graph model, is a type of model that uses a graph structure to describe the dependency relationship between parameters.

[0100] The Bayesian network consists of nodes and edges. The nodes represent specific indicators of raw material quality attributes (such as raw material moisture, ash content, etc.), process parameters (such as extraction temperature, filtration pressure, etc.) and finished product quality attributes (such as moisture, fineness, etc.);

[0101] The edges represent the causal relationship between the above parameters, and the arrows of the edges point from "cause" to "effect". For example: raw material moisture → extraction temperature means that the raw material moisture affects the extraction temperature. The subsequent weight calculation is for these "edges". By assigning a corresponding strength value to each causal relationship, the key influencing factors of the multi-process production process can be quantified and applied to parameter optimization and quality prediction; when applied to industrial processes, the Bayesian network must be predefined to make it consistent with common sense cognition to prevent the emergence of obviously unreasonable causal relationships between nodes.

[0102] The structural optimization includes removing redundant nodes and edges in combination with a network structure optimization algorithm, and the optimization algorithm uses training set data and a Bayesian network structure optimization method. At present, the main structural optimization methods include score-based structure learning, constraint-based structure learning, and hybrid structure learning.

[0103] The present invention adopts a scoring search-based method, the scoring function is the Bayesian information criterion, and the search algorithm is the Hill Climbing method. After the structure is optimized, the raw material quality attribute statistics and process parameter statistics retained in the Bayesian network will be regarded as key process parameter statistics, providing important support for subsequent parameter optimization and quality prediction.

[0104] A5. Causal weight calculation

[0105] The graph neural network algorithm is used to calculate the edge weights between nodes, quantify the strength of the causal relationship between variables, and provide a reference for subsequent optimization; the graph neural network algorithm uses the graph attention network algorithm to assign differentiated attention scores to the neighbors of each node in the Bayesian network, thereby updating the network connection weights and identifying neighboring nodes that are more important to the treated nodes, accurately characterizing the degree of causal influence between variables; see Figure 2 , the specific process includes:

[0106] A51. Assume that the Bayesian network contains V nodes, and the feature vector of each node is denoted as h i ∈R D , where i = 1, 2, ... V, D is the dimension; using a shared weight matrix W∈R D*D′ For node h i and adjacent nodes h j Perform a linear transformation to obtain a new eigenvector h′ i and h′ j , dimensions are D′, h′ i, h′ j ∈R D′ , h′ i The calculation formula is h′ i =Wh i ;

[0107] A52, the eigenvector h′ i and h′ j After concatenation, input the feedforward neural network Map it to a real number and activate it using the LeakyReLU function to get the attention correlation coefficient e ij , and its calculation formula is as follows:

[0108]

[0109] Among them, || represents vector concatenation, is a trainable parameter vector;

[0110] A53, the attention correlation coefficient e ijNormalized by the softmax function, calculate the attention weight a between two nodes ij , and its calculation formula is:

[0111]

[0112] Among them, N i Represents all neighbor nodes of node i, according to the attention weight a ij Update the edge strength between node i and node j in the Bayesian network to quantify the strength of the causal relationship between variables.

[0113] A6. Graph Convolutional Neural Network Modeling

[0114] The causal relationship graph network and the raw material quality attributes and process parameters in the original data set are used as input parameters, and the key finished product quality attributes are used as output parameters to train and construct a corresponding nonlinear association model; the nonlinear association of the causal relationship graph network is modeled using a neural network model to obtain a neural network model; refer to Figure 3 The neural network model adopts a graph convolutional neural network model, takes the causal relationship graph network and the statistics of raw material quality attributes and process parameters as input, and the statistics of key finished product quality attributes as output, so as to achieve an accurate description of the nonlinear correlation between parameters of multi-process production cycles;

[0115] Causal graph data is a typical non-Euclidean structure data, and the number of neighbor nodes of different nodes is different. The graph convolutional neural network model maintains translation invariance and can extract effective spatial features from it for deep learning. The graph convolutional neural network model includes: input layer, convolution, flatten layer and fully connected layer; the specific process of modeling includes:

[0116] A61, the input layer node feature matrix is ​​H (0) , the node feature matrix H after convolutional layer mapping (1) , the propagation calculation formula from the input layer to the convolutional layer is:

[0117]

[0118] in, I is the identity matrix, A∈R n×n is the adjacency matrix of the graph, for The degree matrix of W (0) is the weight matrix to be trained, σ() is the activation function used by the convolutional layer, H (0) is the input layer node feature matrix;

[0119] A62, the node feature matrix H(1) At the same time, the full connection layer is input to predict N key finished product quality attributes; the propagation rule from the convolutional layer to the full connection layer is:

[0120] H (i+i) =σ(H (i) W (i) )

[0121] Among them, H (i) is the node feature matrix of the i-th layer neural network, σ() is the activation function used by the convolutional layer, W (i) The weight matrix to be trained;

[0122] A63. Combine the Adam optimizer, learning rate decay, dropout and early stopping strategies during training to obtain the optimal graph convolutional neural network model.

[0123] During the training process, the Adam optimizer was used, the initial learning rate was set to 0.0001, and the learning rate was halved after every 2000 epochs; the learning process was terminated by setting the dropout random inactivation probability to 0.1 and enabling early stopping (training was stopped when the validation set loss did not improve for 500 consecutive epochs). After the optimal nonlinear association model was trained, the RMSE and The model's fit was evaluated using RMSEP and To evaluate its performance in the prediction stage.

[0124] A7. Parameter optimization and control range optimization

[0125] Based on the nonlinear association model, the optimized process parameters including raw material quality attributes and process parameters are obtained; the process parameters are obtained from the nonlinear association model using a genetic algorithm, and the reasonable control range of key raw material quality attributes and process parameters is inferred through the nonlinear association model. The genetic algorithm regards the feasible solutions in the problem domain as independent individuals or chromosomes in the population, and after encoding the individuals in the form of strings, the population is repeatedly subjected to genetic operations (inheritance, crossover and mutation), and each individual is evaluated according to a preset fitness function, and individuals with lower fitness are eliminated to achieve iterative optimization of individuals in the population. The genetic evolution operation of the basic genetic algorithm is simple, and only includes three basic genetic operators: selection, crossover and mutation. Its mathematical model is:

[0126] SGA=(C,E,P0,M,φ,Γ,ψ,T)

[0127] Among them, C represents the encoding scheme of individuals, E represents the individual fitness evaluation function, P0 represents the initial population, M represents the population size, φ, Γ and ψ represent the selection operator, crossover operator and mutation operator respectively, and T represents the termination evolutionary generation;

[0128] The specific steps of the genetic algorithm include:

[0129] A71. Initialization parameters

[0130] Set P0, T, PΓ, and P ψ , set the gene as the upper and lower control limits of the variable, the number of genes is twice the number of variables, and set the number of iterations t = 1;

[0131] A72. Initialize population coding

[0132] The initial population code is generated by a completely random method, and then the chromosomes of the population are encoded using binary code to generate an initial population consisting of N chromosomes:

[0133] pop i (t), t=1, i=1, 2, …, N; when t=1, N=P0

[0134] A73. Calculate the fitness evaluation function and select

[0135] Calculate the fitness of each individual, the fitness of the i-th individual is f(i). Let f(s) = minf(i), the s-th individual is the best individual in the population. Calculate the selection probability based on the individual fitness value:

[0136]

[0137] Among them, f(i) is the fitness of the i-th one, expressed as P i The probability distribution of randomly selecting some chromosomes from the current generation population to be inherited to the next generation population;

[0138] A74. Crossover and mutation operations

[0139] With probability P Γ Make the chromosomes cross over, and then with probability P ψ Make the chromosome gene mutate and form a new offspring pop i (t), and it serves as the parent of the next genetic operation, returning to step S3, iterating t+1 times and repeating;

[0140] A75. Algorithm termination conditions

[0141] When the number of iterations t reaches T, the algorithm terminates and outputs the optimal population.

[0142] A8. Verification of control scope

[0143] After using genetic algorithms to infer the optimal control range of the optimal process parameter statistics from the nonlinear association model, the optimal control range is applied to the test set. The effectiveness and feasibility of the control range are verified based on the expanded test set data: the upper and lower limits of the optimal control range are applied to the key variables in the test set, and the changes in key finished product quality attributes are predicted through the trained GCN model;

[0144] If the statistical indicators of key finished product quality attributes in the test set are significantly improved compared with the original control range, it means that the derived control range is scientific and stable at the data level.

[0145] Subsequently, a comprehensive evaluation is conducted based on actual factors such as the adjustable range of production equipment, process safety, and economic feasibility. If there is no equipment overload or unreasonable cost increase, it means that the control range is operational and applicable in actual production, ultimately achieving global optimization.

[0146] Embodiment 2

[0147] All algorithms in this example are compiled using Python (v3.9.12; Python Software Foundation, 2022); the specific steps include:

[0148] B1. Data collection and variable definition:

[0149] Analyze the production process of de-walled Ganoderma lucidum spore powder, collect the quality attribute data of raw material batches of de-walled Ganoderma lucidum spore powder produced from January to October 2021, the process parameter data of multiple processes (collected from extraction, filtration, concentration, drying, crushing and screening and other process links) and the quality attribute data of the final finished product batches;

[0150] The quality attribute data is cleaned to remove data records containing missing values ​​and outliers, and then the raw material data, process parameter data, and finished product quality attribute data are aligned through the production batch record table, and a total of 110 complete records are obtained. The original data is a matrix of size 110×19;

[0151] Refer to Table 2, there are 6 iqs, 9 pps, and 4 oqs. Table 2 iq, pp and oq description table details the data items represented by iq, pp and oq.

[0152] Table 2 IQ, PP and OQ Description Table

[0153]

[0154]

[0155] B2. Screening of key finished product quality attributes

[0156] Using Shewhart control charts and P pk Key OQ variables in standard screening physicochemical indicators; see Figure 4 In the Shewhart control chart, the dotted line is the mean ± 3 standard deviation control limit, and the solid line is the upper or lower control limit. The oq2 of some batches exceeds the dotted line range;

[0157] Calculate P pk , find the P of oq1, oq2, oq3 and oq4 pk The P values ​​of oq3 and oq4 were 2.146, 3.731, 0.947, and 1.440, respectively. pk <1.67. Therefore, both the Shewhart control chart and the P pk Standard, OQ2, OQ3 and OQ4 should be considered as key finished product quality attributes (OQ).

[0158] B3. Sample expansion and statistical calculation

[0159] Sample space expansion and statistics calculation: Use the Bootstrap sampling method to expand the sample space. First, randomly divide the original data set into the original training set and the original test set in a ratio of 2:1;

[0160] Then, 12,000 bootstrap samples were performed on the original training set and 6,000 bootstrap samples were performed on the original test set, with 40 batches of records collected each time;

[0161] After each sampling, the statistics are calculated once on a batch of 40 records. After all sampling and statistics calculations are completed, the training set and test set are expanded to matrices of size 12000×19 and 6000×19, respectively.

[0162] B4. Bayesian network modeling and optimization

[0163] The Bayesian network is predefined manually to ensure that it conforms to common sense. The network nodes include all pp statistics, q statistics, and key oq statistics; in this predefined network, the process parameters of the upstream process are used as the "cause" and the process parameters of the downstream process are used as the "effect". The cause-effect relationship is transmitted from the upstream to the downstream until it affects the key oq statistics.

[0164] See Figure 5 After structural optimization, the Bayesian network retains 4 key PP statistics and 6 key IQ statistics; there is a direct or indirect connection between the key IQ and the key OQ statistics.

[0165] B5. Causal weight calculation

[0166] The Bayesian network has 13 nodes, and the number of node features is the sample size of the training set (D = 12000). GAT is used to calculate the Bayesian network connection weights to obtain the weighted GXN Bayesian network.

[0167] See also Figure 6 , it can be observed that the total ash content of raw materials (iq2_ppk) and the moisture content of raw materials (iq1_ppk) jointly affect the extraction temperature 1 (pp1_mean) first, and then affect the fineness of the wall-removed Ganoderma lucidum spore powder (oq2_ppk);

[0168] At the same time, the fineness of the wall-removed Ganoderma lucidum spore powder (oq2_ppk) is also directly affected by the raw material crude polysaccharide (iq6_ppk), and the raw material moisture (iq1_ppk) affects the extraction temperature 2 (pp2_mean);

[0169] Triolein (iq5_ppk) has a strong causal relationship with the rapid moisture content (pp8_mean) of the drying process;

[0170] There is a strong causal relationship between the vacuum degree of the drying process (pp7_mean) and the total triterpenoid content (oq4_ppk) of the wall-removed Ganoderma lucidum spore powder;

[0171] There is also a strong causal relationship between the extraction temperature 2 (pp2_mean) of the extraction process and the fineness of the wall-removed Ganoderma lucidum spore powder (oq2_ppk);

[0172] There is a moderate causal relationship between the moisture content of the raw material (iq1_ppk) and the extraction temperature 2 (pp2_mean) of the extraction process.

[0173] B6. Graph Convolutional Neural Network Modeling

[0174] The key pharmaceutical process parameter statistics (6 key IQ statistics, 4 key PP statistics) in the Bayesian network are used as the node feature matrix, the connection relationship between the key pharmaceutical process parameters is used as the adjacency matrix to input the GCN model, and then the key OQ parameters are used as the output of the GCN model. The quantitative results of the trained GCN model are shown in Table 3:

[0175] On the test set, oq2_ppk, oq3_ppk and oq4_mean They are 0.401, 0.514 and 0.541 respectively, indicating that the model has certain predictive ability, but the error between the predicted value and the true value is still large. The reason for the large prediction error may be incomplete information, and there are some factors that are not perceived, such as some raw material quality attributes, some equipment process parameters, personnel operation and environment-related parameters, etc.

[0176] Table 3 Quantitative results of graph convolutional neural network model

[0177]

[0178] B7. Parameter optimization and control range verification

[0179] Based on the GCN prediction model, a genetic algorithm was used to optimize the key pharmaceutical process parameter intervals, which improved oq2_ppk, oq3_ppk and oq4_ppk;

[0180] Specifically, the initialization parameters of the genetic algorithm are set as follows: the initial population size P0 is 100, the termination evolution generation T is 150, and the crossover probability P Γ is 0.7, the mutation probability P ψ is 0.016; the genes are set as the upper and lower control limits of the key iq and key pp, and the number of genes is twice the number of variables (n=36). The fitness function of the genetic algorithm is set as:

[0181]

[0182] x represents the current individual, gcn(oq i ) is the output of the i-th indicator neural network model, and penalty(x) is the penalty term. After genetic algorithm reasoning, the optimal pharmaceutical process parameter control interval is obtained;

[0183] See also Figure 7 As shown, Figure 7 The horizontal axis is the parameter value, and the vertical axis is the frequency of occurrence of the parameter value. The dashed black lines on the vertical axis are the upper and lower control limits. The lines on both sides of the dashed lines on the vertical axis indicate that the value is outside the optimal range, and the lines on the inside of the dashed lines on the vertical axis indicate that the value is within the optimal range.

[0184] The optimized control interval is applied to the test set and compared with the results of the current interval (Table 4). It is found that the key oq statistics are significantly improved. On the test set, oq2_ppk increases from 4.229 to 5.396, oq3_ppk increases from 1.016 to 1.1.063, and oq4_ppk increases from 1.643 to 1.77.

[0185] Table 4 Comparison results of test set before and after optimization

[0186]

[0187] In summary, the present invention has the following advantages:

[0188] (1) By combining causal relationship modeling with graph attention networks, the problem of ignoring causal relationships in traditional optimization methods is solved, and redundant adjustments and directional ambiguity caused by traditional reliance on correlation analysis are avoided. At the same time, the attention mechanism can be used to assign weights to each causal edge and quantify the actual impact of different parameters on the results, thereby more effectively focusing on key influencing factors and improving the accuracy and interpretability of the overall process optimization.

[0189] (2) By building a complex nonlinear correlation model through graph convolutional neural network, the limitation of traditional linear assumptions or empirical speculation that it is difficult to accurately deal with the coupling of multiple variables is overcome. It can more comprehensively explore the nonlinear influence and hierarchical correlation between various parameters in the production process, and provide more in-depth and accurate analysis and prediction for multi-process and multi-factor production scenarios.

[0190] The above specific implementation methods are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to examples, a person skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A multi-process parameter optimization method, characterized in that: include: Obtaining an original data set, wherein the original data set includes key finished product quality attributes, raw material quality attributes, and process parameters; the key finished product quality attributes are obtained by screening based on the finished product quality attributes; Constructing a cause-and-effect relationship diagram network between the key finished product quality attributes and the raw material quality attributes and process parameters based on the original data; Using the causal relationship graph network and the raw material quality attributes and process parameters in the original data set as input parameters, and the key finished product quality attributes as output parameters, training and constructing a corresponding nonlinear association model; Optimized process parameters including raw material quality attributes and process parameters are obtained based on the nonlinear correlation model.

2. The method according to claim 1, characterized in that The screening is achieved by combining the Shewhart control chart and the process capability index standard. The Shewhart control chart includes two control limits: the mean ± 3 times standard deviation control limit of the finished product quality attribute and the control limit defined by the quality control standard. The key finished product quality attributes are screened based on the calculation results of the control limits and the process capability index standard.

3. The method according to claim 2, characterized in that The calculation formula of the process capability index standard is: Among them, T U is the upper tolerance limit, T L is the lower tolerance limit, s is the sample population standard deviation, is the mean of the measured values, 3s is three times the standard deviation, and []min is the minimum value in brackets.

4. The method according to claim 1, characterized in that Before constructing the causal relationship graph network, the original data set is further subjected to sample expansion processing, and the expansion processing is implemented by the Bootstrap sampling algorithm. The specific calculation process includes: Set the number of batch records sampled each time by the Bootstrap algorithm to N; Random sampling is performed on the original data set, and after each sampling is completed, the N batch records are converted into a statistical vector by calculating the statistics of the batch records; The statistics of the process parameter variables are recorded as the mean value, denoted as pp_mean; the statistics of the raw material quality attribute variables are denoted as P pk , denoted as iq_P pk ; The statistics of the finished product quality attribute variables are recorded as oq_P pk .

5. The method according to claim 1, characterized in that The causal relationship graph network is constructed using a Bayesian network structure, and the causal relationship graph network is structurally optimized; The structural optimization includes removing redundant nodes and edges in combination with a network structure optimization algorithm, wherein the optimization algorithm uses training set data and a Bayesian network structure optimization method.

6. The method according to claim 5, characterized in that It also includes using the graph attention network algorithm to calculate the weights of the edges in the causal graph network. The specific process of calculating the weights includes: Assume that the Bayesian network contains V nodes, and the feature vector of each node is denoted as h i ∈R D , where i = 1, 2, ... V, D is the dimension; using a shared weight matrix W∈R D*D′ For node h i and adjacent nodes h j Perform a linear transformation to obtain a new eigenvector h′ i and h′ j , where h′ i =Wh i ; The feature vector h′ i and h′ j After concatenation, input the feedforward neural network Map it to a real number and activate it using the LeakyReLU function to get the attention correlation coefficient e ij , and its calculation formula is as follows: Among them, || represents vector concatenation, is a trainable parameter vector; The attention correlation coefficient e ij Normalized by the softmax function, calculate the attention weight a between two nodes ij , and its calculation formula is: Among them, N i Represents all neighbor nodes of node i, according to the attention weight a ij Update the edge strength between node i and node j in the Bayesian network to quantify the strength of the causal relationship between variables.

7. The method according to claim 1, characterized in that The nonlinear association model is constructed using a graph convolutional neural network model, and the specific construction process includes: The causal relationship graph network and the statistics of raw material quality attributes and process parameters are used as input. The graph convolutional neural network model includes: input layer, convolution, Flatten layer and full connection layer; the input layer node feature matrix is ​​H (0) , the node feature matrix H after convolutional layer mapping (1) , the propagation calculation formula from the input layer to the convolutional layer is: in, I is the identity matrix, A∈R n×n is the adjacency matrix of the graph, for The degree matrix of W (0) is the weight matrix to be trained, σ() is the activation function used by the convolutional layer, H (0) is the input layer node feature matrix; The node feature matrix H (1) At the same time, the full connection layer is input to predict N key finished product quality attributes; the propagation rule from the convolutional layer to the full connection layer is: H (i+i) =σ(H (i) W (i) ) Among them, H (i) is the node feature matrix of the i-th layer neural network, σ() is the activation function used by the convolutional layer, W (i) The weight matrix to be trained; The Adam optimizer is used and the initial learning rate is set. The learning rate is halved every fixed training cycle. At the same time, the random inactivation probability of dropout is set, and the training is terminated early when the validation set loss does not decrease for a predetermined number of consecutive cycles. When the above conditions are met, the model training is considered complete.

8. The method according to claim 1, characterized in that The process parameters are obtained from a nonlinear association model using a genetic algorithm, and the mathematical model is: SGA=(C,E,P0,M,φ,Γ,ψ,T) Among them, C represents the individual encoding scheme, E represents the individual fitness evaluation function, P0 represents the initial population, M represents the population size, φ, Γ and ψ represent the selection operator, crossover operator and mutation operator respectively, and T represents the termination evolutionary generation.

9. The method according to claim 8, characterized in that The specific steps of the genetic algorithm include: Set P0, T, PΓ, and P ψ , set the gene as the upper and lower control limits of the variable, the number of genes is twice the number of variables, and set the number of iterations t = 1; The initial population is generated completely randomly, and then the chromosomes of the population are encoded using binary to generate an initial population consisting of N chromosomes: pop i (t), where t = 1, i = 1, 2, …, N; when t = 1, N = P0 Calculate the fitness evaluation function, calculate the fitness of each individual, the fitness of the i-th individual is f(i); let f(s) = minf(i), the s-th individual is the optimal individual of the population; calculate the selection probability based on the individual fitness value: in, is the sum of f(i), f(i) is the fitness of the ith one, and P i The probability distribution of randomly selecting some chromosomes from the current generation population to be inherited to the next generation population; With probability P Γ Make the chromosomes cross over, and then with probability P ψ Make the chromosome gene mutate and form a new offspring pop i (t), and at the same time serves as the parent of the next genetic operation, iterating t+1 times and repeating; When the number of iterations t reaches T, the algorithm terminates and outputs the optimal population.

10. The method according to claim 1, characterized in that The original data set also includes an original test set, and the optimized process parameters are verified based on the original test set.

Images