Soft Sensor Method for Key Process Indicators Based on Parameter Evolution of Dynamic Mechanism Model

Abstract

The present invention relates to the technical field of monitoring key indicators in industrial processes, and specifically discloses a soft measurement method for key process indicators based on the evolution of parameters of a dynamic mechanism model. The present invention establishes an embedded hybrid model based on a mechanism model and a GraphTrans model, estimates the parameters of the mechanism model through the GraphTrans model, ensuring the accuracy of the embedded hybrid model. At the same time, the mechanism model provides specific physical meanings for the dynamic mechanism model parameters output by the GraphTrans model, and completes the prediction of key indicators in industrial processes through the embedded hybrid model, solving the problem that in the traditional soft measurement method introducing a comprehensive state space, the conflict between static parameters and dynamic characteristics in the calculation process of weighted sum will restrict the application effect of the integrated mechanism model.

Description

Technical Field

[0001] This application relates to the technical field of monitoring key indicators in industrial processes, and specifically discloses a soft measurement method for key process indicators based on the evolution of dynamic mechanism model parameters. Background Art

[0002] In the process of modern industry achieving high-level manufacturing, key process indicators have become an important foothold for regulating the production process, and various optimization and control methods are carried out around the key process indicators of different industrial objects. In the process of obtaining key process indicators in some industrial processes, the use of soft measurement technology to obtain key process indicators can make up for the problem of low detection frequency of key process indicators.

[0003] Traditional soft measurement methods establish a mechanism model and use knowledge in fields such as thermodynamics, electrochemistry, and fluid mechanics to summarize and generalize system phenomena, and describe the reaction details of the modeling object in the form of partial differential equations. However, the relationship between industrial process parameters and mechanism model parameters is complex. At the same time, in complex industrial modeling tasks, there is a conflict between the dynamic industrial reaction characteristics and the static parameters of the mechanism model, which limits their accuracy.

[0004] Therefore, in the prior art, a comprehensive state space is introduced to describe the highly complex dynamics of industrial processes and the additional requirements imposed by intelligent and optimal manufacturing systems. In the application process, aiming at the multi-modal and time-varying characteristics of the reaction process dynamics, the comprehensive state space is divided into sub-spaces indicating different operating conditions, and a mechanism model is trained for each different operating sub-space to obtain the corresponding mechanism model parameters. Then, the system dynamics is characterized as a weighted sum of sub-models, where the weights are the probabilities that the current operating point belongs to different operating conditions, and the weights will be updated as the operating point moves in the comprehensive state space. This integrated modeling method describes the dynamic reaction characteristics of industrial objects in the form of a weighted sum of mechanism models, and to a certain extent overcomes the accuracy defects of mechanism models, and has good performance in the prediction task of key process indicators in the purification process of hydrometallurgical zinc smelting.

[0005] However, the introduced comprehensive state space describes the dynamic reaction characteristics of industrial objects in the form of a weighted sum of multiple mechanism models. The characterization ability of a single mechanism model is limited, and the contradiction between its static mechanism parameters and the dynamic reaction characteristics of the modeling object will lead to insufficient modeling accuracy of the mechanism model, restricting its application effect. The weighted sum of mechanism models does not fundamentally overcome the contradiction between the static parameters of the mechanism model and the dynamic reaction characteristics of the modeling object. A single static mechanism model is difficult to describe the dynamic reaction characteristics of a class of sub-conditions, and a weighted combination of a series of static mechanism models is also difficult to describe the complete reaction process. The conflict between static parameters and dynamic characteristics will restrict the application effect of the integrated mechanism model.

[0006] The present invention provides a soft sensing method for key process indicators based on the evolution of dynamic mechanism model parameters to solve the above problems. Summary of the Invention

[0007] The object of the present invention is to solve the problem that in the traditional soft sensing method introducing a comprehensive state space, the conflict between static parameters and dynamic characteristics in the calculation process of weighted sum will restrict the application effect of the integrated mechanism model.

[0008] To achieve the above object, the basic solution of the present invention provides a soft sensing method for key process indicators based on the evolution of dynamic mechanism model parameters, including constructing an embedded hybrid model, and the embedded hybrid model includes:

[0009] A mechanism model, which is constructed based on the knowledge of the process field.

[0010] The GraphTrans model embedded in the mechanism model, and the GraphTrans model includes an encoding block with S layers stacked in sequence and a fully connected layer connected to the encoding block of the S-th layer. Each encoding block contains a graph convolutional neural network module and a Transformer module.

[0011] According to the established embedded hybrid model, soft sensing of key process indicators is carried out, including the following steps:

[0012] Extraction of data: Extract the process parameters of the process being carried out, and establish an association network between the process parameters according to the knowledge of the process field to obtain the graph structure data between the process parameters.

[0013] Input of data: The mechanism model synchronously receives the extracted process parameters and the dynamic mechanism model parameters, and the dynamic mechanism model parameters are calculated by the GraphTrans model receiving the graph structure data established according to the association network.

[0014] Processing of data: The mechanism model receives the process parameters and the dynamic mechanism model parameters to perform real-time estimation of key process indicators, and completes the soft sensing of key process indicators.

[0015] Further, the graph structure data is the adjacency matrix of the graph, and when establishing the association network between the process parameters according to the knowledge of the process field, a graph mask matrix between the process parameters is also obtained.

[0016] Further, after obtaining the graph mask matrix, the graph mask matrix is supplemented into each encoding block in the GraphTrans model to supplement the graph mask matrix into the Transformer module.

[0017] Further, the graph mask matrix is expressed as follows:

[0018] ;

[0019] ;

[0020] In the formula, M represents the graph mask matrix, m ij is an element of the graph mask matrix, M is a square matrix of size N×N, and N is the number of variable nodes; when node i can affect the change of node j, m ij = 1, and when node i cannot affect the change of node j, m ij = 0.

[0021] Furthermore, the mechanism model is constructed based on the first-order reaction kinetics equation, the material conservation equation, the Arrhenius equation, and the continuous stirred tank reactor equation.

[0022] Furthermore, after synchronously receiving the extracted process parameters and the dynamic mechanism model parameters, the calculation formula of the prediction process performed by the mechanism model is as follows:

[0023] Y = f mech (X; θ);

[0024] In the formula, f mech (•) represents the mechanism model, θ is the dynamic parameter output by the GraphTrans model, X is the process parameter, and Y is the predicted key process index.

[0025] Furthermore, the steps for the GraphTrans model to obtain the dynamic mechanism model parameters are as follows:

[0026] The encoding block of the first layer receives all the initial node features, and aggregates the features from neighboring nodes through the graph convolutional neural network module in the encoding block of the first layer;

[0027] Input the aggregated features of the neighboring nodes into the Transformer module in the same layer encoding block. Through the Transformer module with a graph mask matrix, combined with the multi-head self-attention mechanism carried inside the Transformer module, extract different variable-related patterns, and extract the attention features between associated variables;

[0028] Input the obtained attention features into the encoding block of the next layer and implement layer-by-layer feature extraction. Finally, the result is input from the encoding block of the last layer into the fully connected layer to obtain the dynamic mechanism model parameters.

[0029] Furthermore, the process of aggregating the features from neighboring nodes by the graph convolutional neural network module in the encoding block of the first layer is as follows:

[0030] ;

[0031] In the formula, H l is the node feature matrix of the l-th layer, which is the adjacency matrix added with self-connection, and initially, H 0 = X;

[0032] In the formula, A represents the adjacency matrix, , I is the identity matrix, is the degree matrix of, and , W l is the weight matrix of the l-th layer, and , σ(•) is the non-linear activation function;

[0033] In the formula, R is a real number, N is the number of variables, which is also the number of nodes in the graph, and F is the feature length.

[0034] Furthermore, the process of extracting the attention features between associated variables by using the Transformer module with a graph mask matrix is as follows:

[0035] Calculate the attention weights:

[0036] ;

[0037] In the formula, Q, K, and V are the query matrix, key matrix, and value matrix respectively, and Q = XW q , K = XW k , V = XW v , W q , W k , W v all represent learnable weight matrices, and , is the scaling factor for stabilizing the gradient, and T is the transpose symbol of the matrix;

[0038] Calculate the output A h of each head in the multi-head self-attention mechanism:

[0039] ;

[0040] In the formula, Q h , K h , V h are the query matrix, key matrix, and value matrix of the h-th head respectively, and T is the transpose symbol of the matrix;

[0041] The outputs of all heads are concatenated and linearly transformed to generate the final output, and during the calculation process, the calculation formula is as follows:

[0042] ;

[0043] In the formula, ⊙ represents element-wise multiplication, and M represents the graph mask matrix;

[0044] The Replace function ensures zero attention weights by setting the mask value to a negative number, and the expression is as follows:

[0045] Replace(i)={i,i≠0;-1e 9 ,i=0}.

[0046] Furthermore, the process of calculating the dynamic mechanism model parameters by the fully connected layer is as follows:

[0047] θ = W out H S ;

[0048] In the formula, H S is the attention feature output by the last layer of the encoding block, and W out is the learnable weight matrix, representing the learned mechanism model parameters, and , where P is the dimension of the mechanism parameters.

[0049] The principle and effect of this basic solution are as follows:

[0050] 1. Compared with the prior art, the present invention establishes an embedded hybrid model based on the mechanism model and the GraphTrans model, estimates the parameters of the mechanism model through the GraphTrans model, ensures the accuracy of the embedded hybrid model. At the same time, the mechanism model provides specific physical meanings for the dynamic mechanism model parameters output by the GraphTrans model, enhancing the interpretability of the embedded hybrid model. The embedded hybrid model is used to complete the prediction of the key indicators of the industrial process, and solves the problem that in the calculation process of the weighted sum of the traditional soft measurement method introducing the comprehensive state space, the conflict between static parameters and dynamic characteristics will restrict the application effect of the integrated mechanism model.

[0051] 2. Compared with the prior art, the present invention introduces a graph mask matrix into the Transformer module and applies it in the calculation process of the self-attention mechanism, introduces the matrix variable correlation information into the Transformer module, clears the influence of irrelevant variables on the attention weights to prevent the generation of useless information, so as to better model the industrial process with complex variable relationships. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0053] Figure 1 It shows a schematic diagram of the purification process of zinc hydrometallurgy proposed in the embodiments of the present application;

[0054] Figure 2 It shows a flowchart of the soft measurement method for key process indicators based on the evolution of dynamic mechanism model parameters proposed in the embodiments of the present application;

[0055] Figure 3 It shows a schematic diagram of the difference between the soft measurement prediction result and the true value of the key process indicators of the soft measurement method for key process indicators based on the evolution of dynamic mechanism model parameters proposed in the embodiments of the present application applied to the cobalt removal reactor No. 1;

[0056] Figure 4 It shows a change curve diagram of the parameters output by the GraphTrans model proposed in the embodiments of the present application. Among them, (a) is the change curve diagram of parameter A β of the change curve diagram, (b) is the change curve diagram of parameter E e of the change curve diagram, (c) is the change curve diagram of parameter α, and (d) is the change curve diagram of parameter e eq of the change curve diagram. Detailed implementation manners

[0057] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following will, in conjunction with the accompanying drawings and preferred embodiments, describe in detail the specific implementation manners, structures, features and their effects of the present invention as follows.

[0058] The soft measurement method for key process indicators based on the evolution of dynamic mechanism model parameters is as follows in the embodiments. This is an example of applying the soft measurement method for key process indicators based on the evolution of dynamic mechanism model parameters provided by the present invention to the soft measurement of cobalt ion concentration in the cobalt removal reactor No. 1 in the purification process of a zinc hydrometallurgy process.

[0059] In this purification process of zinc hydrometallurgy, as Figure 1As shown, there are two copper removal processes and five cobalt removal processes in sequence. In this embodiment, only the cobalt removal process is shown. In the cobalt removal process, five reactors in series are built in sequence, namely the first cobalt removal reactor to the fifth cobalt removal reactor. The solution overflowing from copper removal is received by the first reactor, and through cascade reaction, the cobalt removal process is completed. In the purification process of zinc hydrometallurgy, the cobalt ion concentration is used as the key process index. In this embodiment, specifically, the soft measurement is carried out for the cobalt ion concentration at the outlet of the first cobalt removal reactor as the key process index. The specific implementation is as follows:

[0060] First, an embedded hybrid model is constructed. The constructed embedded hybrid model includes:

[0061] A mechanism model, which is constructed based on the knowledge in the field of the process. Specifically, a mechanism model is established according to the field knowledge and process characteristics of the cobalt removal process in the zinc hydrometallurgy field.

[0062] The mechanism model is an important source of the interpretability of the embedded hybrid model, providing specific physical meanings for the parameters finally output by the GraphTrans model. The mechanism model describes the internal physical and chemical reaction laws through its internal structure, thus providing structural interpretability for the model. The parameters output by the GraphTrans model are directly used as the mechanism parameters of the mechanism model. Although these parameters are initially dimensionless parameters, in the prediction framework of the mechanism model, through the training process of backpropagation of the prediction error, the parameters output by the GraphTrans model gradually approach the real mechanism parameters. In other words, these parameters play the due role of the real mechanism parameters within the prediction framework of the mechanism model, can accurately describe the production phenomenon, and thus can be equivalently regarded as the real mechanism parameters.

[0063] In practical applications, the mechanism model f of the first cobalt removal reactor mech(•) is established based on the first-order reaction kinetics equation, material conservation equation, Arrhenius equation and continuous stirred tank reactor equation. The partial differential equation of the obtained mechanism model is as follows:

[0064] ;

[0065] Among them, the information of each process parameter in the established mechanism model is shown in Table 1 below:

[0066] Table 1

[0067]

[0068] And, in this embodiment, the parameters that need to be identified in the mechanism model include the reaction coefficient A β , the activation energy of ion precipitation E e , the cathode transfer coefficient α and the cathode reaction equilibrium potential eeq 。

[0069] The embedded hybrid model further includes a GraphTrans model embedded in the mechanism model. The GraphTrans model includes an encoding block with S layers stacked in sequence and a fully connected layer connected to the encoding block of the S-th layer. Each encoding block contains a graph convolutional neural network (GCN) module and a Transformer module, and the output of the GraphTrans model is used as the input of the mechanism model.

[0070] The GraphTrans model is a data model integrating a graph convolutional neural network (GCN) and a Transformer model. As Figure 2 shown, Figure 2 the components of the GraphTrans model are shown.

[0071] The mechanism model in the embedded hybrid model is used as a prediction model to estimate key process indicators, completing the soft measurement of key process indicators.

[0072] Specifically, according to the established embedded hybrid model, the soft measurement of key process indicators is carried out, including the following steps:

[0073] Data extraction: Extract the process parameters of the process being carried out, and establish an association network between the process parameters according to the knowledge in the field of the process, obtaining the graph structure data between the process parameters;

[0074] Data input: The mechanism model synchronously receives the extracted process parameters and dynamic mechanism model parameters, and the dynamic mechanism model parameters are calculated by the GraphTrans model receiving the graph structure data;

[0075] Data processing: After the mechanism model receives the process parameters and dynamic mechanism model parameters, it performs real-time estimation of key process indicators, completing the soft measurement of key process indicators.

[0076] Among them, the established graph structure data includes the adjacency matrix of the graph, which is expressed as follows:

[0077] ;

[0078] In the formula, A represents the adjacency matrix, a ij is an element of the adjacency matrix, A is a square matrix of size N×N, N is the number of variable nodes, v ij is a directed edge connecting node i and node j, E is the set of all directed edges. When v ij ∈E, it means that there is a directed edge from node i to node j. At this time, a ij = 1, otherwise a ij= 0. In the association graph obtained in this embodiment, there are 10 nodes in total, N = 10.

[0079] The adjacency matrix A represents the information transfer direction between node variables and can guide the information aggregation process of the GCN module.

[0080] According to the association network between process parameters, a graph mask matrix is also established, and the graph mask matrix is supplemented into each encoding block in the GraphTrans model. Specifically, the graph mask matrix is supplemented into the Transformer module.

[0081] The established graph mask matrix is as follows:

[0082] ;

[0083] ;

[0084] In the formula, M represents the graph mask matrix, m ij is an element of the graph mask matrix, M is a square matrix of size N×N, N is the number of variable nodes, M can represent the causal relationship between nodes. When node i can affect the change of node j, m ij = 1. When node i cannot affect the change of node j, m ij = 0.

[0085] In this embodiment, introducing the graph mask matrix into the calculation mechanism of the Transformer model can enable the Transformer to have the ability to process graph-structured data, thereby more efficiently extracting the variable association features in the process parameters.

[0086] In the process of data input, after the adjacency matrix of the graph is input into the GraphTrans model, the steps for the GraphTrans model to obtain dynamic mechanism model parameters are as follows:

[0087] The first-layer encoding block receives all initial node features, and the GCN module in the first-layer encoding block aggregates the features from neighbor nodes to capture local structure information. The specific calculation process is as follows:

[0088] ;

[0089] In the formula, H l is the node feature matrix of the l-th layer, is the adjacency matrix with self-connections added, and , initially H 0 = X, X is the standardized variable matrix, and , where \(R\) is a real number, \(N\) is the number of variables, which is also the number of nodes in the graph, and \(F\) is the feature length. In the original data, it is the time step length. One sampling value is obtained in one time step, and the sampling values for a period of time are represented by \(F\) time steps. And , \(I\) is the identity matrix, is the degree matrix of , \(W\) l is the weight matrix of the \(l\)-th layer, and , \(\sigma(\cdot)\) is a non-linear activation function, such as ReLU.

[0090] Input the features of the aggregated neighbor nodes into the Transformer module in the same layer encoding block. Through the Transformer module with a graph mask matrix, combined with the multi-head self-attention mechanism carried inside the Transformer module, extract different variable-related patterns and obtain the attention features between associated variables. The specific calculation is as follows:

[0091] Calculate the attention weights:

[0092] ;

[0093] In the formula, \(Q\), \(K\), and \(V\) are the query matrix, key matrix, and value matrix respectively, and \(Q = XW\) q , \(K = XW\) k , \(V = XW\) v , \(W\) q , \(W\) k , \(W\) v all represent learnable weight matrices, and , is the scaling factor for stabilizing the gradient, and \(T\) is the transpose symbol of the matrix.

[0094] Calculate the output \(A\) of each head in the multi-head self-attention mechanism h :

[0095] ;

[0096] In the formula, \(Q\) h , \(K\) h , \(V\) h are the query matrix, key matrix, and value matrix of the \(h\)-th head respectively, and \(T\) is the transpose symbol of the matrix.

[0097] The outputs of all heads are concatenated and linearly transformed to generate the final output. During the calculation process, by applying the mask matrix to the attention mechanism, the attention weights of irrelevant node pairs are eliminated. The calculation formula is as follows:

[0098] ;

[0099] In the formula, ⊙ represents element-wise multiplication, M represents the graph mask matrix, T is the transpose symbol of the matrix, and the Replace function ensures zero attention weights by setting the mask value to a sufficiently small negative number:

[0100] Replace(i)={i,i≠0;-1e 9 ,i=0}.

[0101] By applying the masked matrix, the attention mechanism focuses only on causally related variable pairs;

[0102] The obtained attention features are input into the encoding block of the next layer to achieve layer-by-layer feature extraction. Finally, the features H S are input into the fully connected layer to obtain the dynamic mechanism model parameters:

[0103] θ = W out H S ;

[0104] In the formula, H S is the attention feature output by the last encoding block, W out is the learnable weight matrix, representing the learned mechanism model parameters, and , P is the dimension of the mechanism parameters; each element in θ will be used as the parameter inside the mechanism model. Therefore, the output dimension needs to be specifically set to match the number of parameters required by the mechanism model.

[0105] After that, data processing is performed. The mechanism model receives the process parameters and the dynamic mechanism model parameters, and performs real-time estimation of the key process indicators through the partial differential equations of the established mechanism model, completing the soft measurement of the key process indicators. The difference between its prediction result and the true value is as Figure 3 shown.

[0106] The prediction process of the mechanism model is as follows:

[0107] Y = f mech (X;θ);

[0108] In the formula, f mech (•) represents the mechanism model, expressed as a set of partial differential equations derived from domain knowledge, θ is the dynamic parameter output by the GraphTrans model, X is the process parameter, and Y is the predicted key process indicator, such as ion concentration, etc.

[0109] When the mechanism model receives the dynamic mechanism model parameters, the mechanism model assigns specific physical meanings to the parameter values output by the GraphTrans model. Its parameters: A β , E e , α and eeq The change curves are successively as follows Figure 4 shown in Figures (a), (b), (c), and (d) of Figure 4 . It can be seen that each parameter variable fluctuates around a specific value, but the fluctuation amplitudes are different. It is consistent with the actual production conditions. After the process design is completed and production starts, the internal reaction state of the reactor should be kept as stable as possible to ensure the production of qualified products. The fluctuations in the parameter values of the mechanism model directly reflect the parameter mismatch phenomenon.

[0110] In this embodiment, five other models are also introduced to predict the cobalt ion concentration at the outlet of the first cobalt removal reactor. Combined with the embedded hybrid model, there are a total of six models as shown in Table 2 below:

[0111] Table 2

[0112]

[0113] The present invention uses the mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and R 2 to measure the accuracy of the model. The specific calculation formulas are as follows:

[0114] ;

[0115] ;

[0116] ;

[0117] ;

[0118] where N is the number of samples in the test data set, R n is the actual value of the nth sample, and P n is the predicted value of the nth sample, is the mean of the true values of the samples, is the mean of the predicted values. The performance of the six models shown in Table 2 in the four indicators is shown in Table 3:

[0119] Table 3

[0120]

[0121] It can be seen that the GraphTrans model proposed by the present invention is superior to a series of classical models such as GCN and Transformer, verifying the excellent dynamic feature extraction ability of the model, and being able to detect subtle differences in the reactor operation mode from process parameters. At the same time, it can be found that GraphTrans-CM is significantly superior to GraphTrans, proving the effectiveness of the embedded hybrid model proposed by the present invention. The effective combination of the data model and the mechanism model can further improve the prediction accuracy of process indicators.

[0122] The present invention establishes an embedded hybrid model based on the mechanism model and the GraphTrans model. By using the GraphTrans model to estimate the parameters of the mechanism model in real time, the accuracy of the embedded hybrid model is guaranteed. At the same time, the mechanism model provides specific physical meanings for the dynamic mechanism model parameters output by the GraphTrans model, enhancing the interpretability of the embedded hybrid model. The prediction of key indicators of industrial processes is completed through the embedded hybrid model, solving the problem that in the traditional soft measurement method introducing a comprehensive state space, the conflict between static parameters and dynamic characteristics in the calculation process of weighted sum will restrict the application effect of the integrated mechanism model.

[0123] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to obtain equivalent embodiments with equivalent changes. However, as long as it does not depart from the technical content of the present invention, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A soft-sensing method for key process indicators based on the evolution of dynamic mechanism model parameters, characterized in that: The method comprises building an embedded hybrid model, wherein the embedded hybrid model comprises: A mechanism model, wherein the mechanism model is constructed based on the knowledge of the field to which the process belongs; A GraphTrans model embedded in the mechanism model, wherein the GraphTrans model includes S layers of coding blocks stacked in sequence and a fully connected layer connected to the coding block of the Sth layer, each coding block including a graph convolutional neural network module and a Transformer module; According to the established embedded hybrid model, soft measurement of key process indicators is performed, including the following steps: Data extraction: extract the process parameters of the process being carried out, and establish a correlation network between the process parameters based on the knowledge of the field to which the process belongs, and obtain the graph structure data between the process parameters; The graph structure data is an adjacency matrix of the graph, and an association network between process parameters is established according to the domain knowledge to which the process belongs, and a graph mask matrix between the process parameters is also obtained; The input of data, the mechanism model synchronously receives the extracted process parameters and dynamic mechanism model parameters, and the dynamic mechanism model parameters are calculated by the GraphTrans model based on the graph structure data established according to the association network; The steps of obtaining dynamic mechanism model parameters of the GraphTrans model are as follows: The encoding block of the first layer receives all the initial node features, and aggregates the features from neighboring nodes through the graph convolutional neural network module in the encoding block of the first layer; The features of the aggregated neighbor nodes are input into the Transformer module in the encoding block of the same layer. By using the Transformer module with a graph mask matrix and combining the multi-head self-attention mechanism inside the Transformer module, different variable correlation patterns are extracted to obtain the attention features between the associated variables. The obtained attention features are input into the encoding block of the next layer and feature extraction is realized layer by layer. Finally, the encoding block of the last layer inputs the results into the fully connected layer to obtain the dynamic mechanism model parameters. The process of aggregating features from neighboring nodes by the graph convolutional neural network module in the encoding block of the first layer is as follows: ； In the formula, H l is the node feature matrix of the lth layer, is the adjacency matrix with self-connection added, and , initially H 0 =X, X is the standardized variable matrix, and ; In the formula, A represents the adjacency matrix, , I is the identity matrix, yes The degree matrix of , W l is the weight matrix of the lth layer, and , σ(•) is a nonlinear activation function; In the formula, R is a real number, N is the number of variables, which is also the number of nodes in the graph, and F is the characteristic length; For data processing, the mechanism model receives process parameters and dynamic mechanism model parameters, makes real-time estimation of key process indicators, and completes soft measurement of key process indicators.

2. The key process indicator soft-sensing method based on the dynamic mechanism model parameter evolution according to claim 1 is characterized in that: After obtaining the graph mask matrix, each encoding block in the GraphTrans model is supplemented with the graph mask matrix to supplement the graph mask matrix into the Transformer module.

3. The key process indicator soft-sensing method based on dynamic mechanism model parameter evolution according to claim 1 is characterized in that: The graph mask matrix is ​​represented as follows: ； ； Where M represents the image mask matrix, m ij is the element of the graph mask matrix, M is a square matrix of size N×N, where N is the number of variable nodes; When node i can affect the change of node j, m ij =1, when node i cannot affect the change of node j, m ij =0.

4. The key process indicator soft-sensing method based on dynamic mechanism model parameter evolution according to claim 1 is characterized in that: The mechanism model is constructed based on the first-order reaction kinetics equation, the material conservation equation, the Arrhenius equation and the continuous stirred tank reactor equation.

5. The key process indicator soft-sensing method based on dynamic mechanism model parameter evolution according to claim 1 is characterized in that: After the mechanism model synchronously receives the extracted process parameters and the dynamic mechanism model parameters, the calculation formula of the prediction process is as follows: Y=f mech (X;θ); In the formula, f mech (•) represents the mechanism model, θ is the dynamic parameter output by the GraphTrans model, X is the process parameter, and Y is the predicted key process indicator.

6. The key process indicator soft-sensing method based on dynamic mechanism model parameter evolution according to claim 1 is characterized in that: The process of extracting the attention features between associated variables by using the Transformer module with a graph mask matrix is ​​as follows: Calculate the attention weights: ； Where Q, K and V are the query matrix, key matrix and value matrix respectively, and Q = XW q , K = XW k , V = XW v , W q ,W k ,W v denotes a learnable weight matrix, and , is the scaling factor used to stabilize the gradient, and T is the transposed sign of the matrix; Calculate the output A of each head in the multi-head self-attention mechanism h : ； In the formula, Q h ,K h ,V h are the query matrix, key matrix and value matrix of the h-th head respectively, and T is the transpose symbol of the matrix; The outputs of all heads are concatenated and transformed linearly to generate the final output, which is summarized in the calculation process. The calculation formula is as follows: ； Where ⊙ represents element-by-element multiplication, and M represents the graph mask matrix; The Replace function ensures zero attention weight by setting the mask value to a negative number. The expression is as follows: Replace(i)={i,i≠0;-1e 9 ,i=0}。 7. The key process indicator soft-sensing method based on dynamic mechanism model parameter evolution according to claim 1 is characterized in that: The process of calculating the dynamic mechanism model parameters by the fully connected layer is as follows: θ=W out H S ; In the formula, H S is the attention feature output by the last layer encoding block, W out is a learnable weight matrix, representing the learned mechanism model parameters, and , P is the mechanism parameter dimension.

Images