An industrial quality prediction method based on prior knowledge constraint graph convolution

Abstract

The application relates to an industrial quality prediction method based on prior knowledge constraint graph convolution. The method collects multivariate time series data containing quality variables and process variables, deeply mines the Granger causality relationship between the variables based on the multivariate time series data, preliminarily obtains a directed information transmission matrix, and dynamically adjusts and refines the information transmission relationship between the variables in combination with a prior sub-process knowledge mask matrix to generate a dynamic adjacency matrix. A multi-head spatiotemporal graph convolution long short-term memory network is designed based on the dynamic adjacency matrix to learn long short-term spatiotemporal features. The method uses the Granger causality relationship based on the constraint prior knowledge to effectively mine the correlation between the quality variables and the process variables, and uses the multi-head spatiotemporal graph convolution long short-term memory network to capture long short-term dependency relationships, which helps to improve the accuracy of industrial system quality prediction.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of industrial quality prediction, in particular to an industrial quality prediction method based on prior knowledge constraint graph convolution. BACKGROUND

[0002] In modern industrial processes, real-time accurate measurement and monitoring of key product quality is the key to effectively tracking production efficiency and maintaining safe production standards. Existing soft sensor modeling methods using shallow learning are used for industrial quality prediction, but the structure configuration is relatively shallow, which limits the expression and generalization ability of the model.

[0003] Due to the complex coupling relationship in the large amount of data recorded by sensors and professional equipment, accurate mining of data correlation to extract similar features is the key to successful modeling for deep learning modeling. However, the existing multivariate correlation coefficient graph analysis method mostly analyzes the relationship between variables at the same time, does not consider the time lag, and does not comprehensively consider the actual industrial process reaction mechanism, lacking a certain credibility. In addition, in the aspect of feature dependence mining, mainstream deep learning methods such as Transformer, recurrent neural network, graph neural network strategy and its variants rarely consider simultaneously utilizing long and short term dependencies within industrial data variables and potential spatial dependencies between variables to learn more effective features. SUMMARY

[0004] Therefore, it is necessary to provide an industrial quality prediction method based on prior knowledge constraint graph convolution, comprising:

[0005] S1: Collecting multivariate time series data containing quality variables and process variables in the process industrial system;

[0006] S2: Mining the Granger causality relationship between each variable in the multivariate time series data to obtain the information transmission relationship between each variable, and constructing a directed information transmission matrix based on the information transmission relationship between each variable; determining whether each variable belongs to the same sub-process based on prior knowledge, and constructing a prior sub-process knowledge mask matrix according to the determination result; fusing the directed information transmission matrix, the prior sub-process knowledge mask matrix and the designed adaptive weight matrix to obtain a dynamic adjacency matrix;

[0007] S3: Designing a multi-head spatio-temporal graph convolution long short-term memory network based on the dynamic adjacency matrix, inputting the past time process variables and the past time quality variables into the multi-head spatio-temporal graph convolution long short-term memory network, obtaining a plurality of preliminary prediction results of the current time quality variables, and linking each preliminary prediction result; decoding the linked result to obtain the prediction value of the current time quality variable.

[0008] Preferably, the process industrial system includes a hydrocracking industrial system; the quality variable includes the total C5 of light naphtha in the hydrocracking process; the process variables include the top bed temperature, bottom bed temperature, reactor differential pressure, tower top pressure, and tower top reflux flow rate of the hydrocracking industrial system during the hydrocracking process.

[0009] Preferably, in S2, the process of obtaining the information transmission relationship between variables includes:

[0010] For any variable A and any other variable B in a multivariate time series dataset, construct a univariate autoregressive model of variable A and a multivariate regression model of variable A with respect to variable B.

[0011] When variable A is a quality variable, the univariate autoregressive model of variable A is expressed as:

[0012] ;

[0013] in, express The quality variable at any given time; This indicates the 0th first parameter; This indicates the first parameter; This indicates the second first parameter; Indicates the first The first parameter; express The quality variable at any given time; express The quality variable at any given time; express The quality variable at any given time; Indicates the optimal lag order; express The first residual at time;

[0014] When variable B is the first process variable, the multivariate regression model of variable A with respect to variable B is expressed as:

[0015] ;

[0016] in, This indicates the 0th second parameter; Indicates the first second parameter; This indicates the second parameter; Indicates the first A second parameter; This indicates the 0th third parameter; This indicates the first third parameter; This indicates the second third parameter; Indicates the first a third parameter; denotes a first process variable at a time instant; denotes a first process variable at a time instant; denotes a first process variable at a time instant; denotes a second residual at a time instant;

[0017] Based on the univariate autoregressive model and the multivariate regression model, a joint hypothesis statistic is calculated, and the calculation formula is:

[0018]

[0019] wherein, denotes the joint hypothesis statistic; denotes the residual sum of squares of the univariate autoregressive model; denotes the residual sum of squares of the multivariate regression model; denotes the sample length of the quality variable;

[0020] When the joint hypothesis statistic is greater than a threshold value, it is determined that the variable B is a cause of the variable A, that is, the variable B has an information transmission relationship with the variable A.

[0021] Preferably, in S2, constructing the directed information transmission matrix based on the information transmission relationship between the variables comprises:

[0022] When there is an information transmission relationship between the variables, the value is 1, otherwise it is 0; and when the variable is a loop, the value is 1;

[0023] Taking the information transmission relationship between the quality variable and itself and each process variable as the first row of the directed information transmission matrix;

[0024] Iterating through all the process variables in order, for the first process variable, taking the information transmission relationship between the variable and itself and other variables as the first row of the directed information transmission matrix, , denotes the number of process variables;

[0025] Substituting the corresponding values into each position of the directed information transmission matrix, the directed information transmission matrix is constructed, and the diagonal elements of the directed information transmission matrix are all 1.

[0026] Preferably, in S2, according to prior knowledge, it is determined whether each variable belongs to the same sub-process, and according to the determination result, a prior sub-process knowledge mask matrix is constructed, comprising:

[0027] ​The process industry in the process industry system is a multi-process production mode, different processes are different sub-processes, and determining whether any two variables belong to the same sub-process is to determine whether any two variables belong to the same process according to prior knowledge;

[0028] When the two variables belong to the same sub-process, the value is 1, otherwise it is 0.5; and when the variable is self-loop, the value is 1;

[0029] Whether the quality variable and itself and each process variable belong to the same sub-process is taken as the first row of the prior sub-process knowledge mask matrix;

[0030] All process variables are traversed in order, and for the first process variable, whether it and itself and other variables belong to the same sub-process is taken as the first row of the prior sub-process knowledge mask matrix, , represents the number of process variables;

[0031] The corresponding value is substituted into each position of the prior sub-process knowledge mask matrix to construct the prior sub-process knowledge mask matrix, and the diagonal elements of the prior sub-process knowledge mask matrix are all 1.

[0032] Preferably, in S2, the Hadamard product of the directed information transmission matrix, the prior sub-process knowledge mask matrix and the designed adaptive weight matrix is calculated to obtain a dynamic adjacency matrix; the adaptive weight matrix is a random weight matrix, and the weight values in the random weight matrix are automatically updated with the model parameters updated in the training process of the multi-head spatio-temporal graph convolution long short-term memory network.

[0033] Preferably, in S3, each spatio-temporal graph convolution long short-term memory network includes an input gate, a forget gate, a cell state, and an output gate, and the input gate, the forget gate, and the output gate all perform directed graph convolution on the input object through the dynamic adjacency matrix; the directed graph convolution expression is:

[0034] ;

[0035] Among them, represents the adaptive weight; represents the directed graph convolution operation; represents the object of the directed graph convolution; represents the gelu activation function; represents the variable in-degree matrix; represents the dynamic adjacency matrix; represents the adaptive bias;

[0036] The variable in-degree matrix is ​​a diagonal matrix, and the values ​​on its diagonal are the sum of the element values ​​of the corresponding column in the directed information transfer matrix.

[0037] Preferably, to predict the value of the quality variable at the current moment by sampling the process variables and quality variables of q consecutive past moments through a sliding window, it is necessary to connect q spatiotemporal graph convolutional long short-term memory network units into a spatiotemporal graph convolutional long short-term memory network. Any spatiotemporal graph convolutional long short-term memory network unit includes:

[0038] Input Gate:

[0039] ;

[0040] ;

[0041] in, express The input gate's output at a given time; This represents the GeLU activation function; Representation layer normalization; Indicates the weights of the input gates; This represents a directed graph convolution operation; express The hidden state at any given moment; express Process variables at any given time; Indicates a link operation; Indicates the bias of the input gate; express The state of candidate cells at any given time; Represents the hyperbolic tangent function; The weights corresponding to the states of candidate cells; This indicates the bias corresponding to the candidate cell state.

[0042] Forgotten Gate:

[0043] ;

[0044] in, express The output of the forget gate at any given moment; Indicates the weight of the forget gate; Indicates the bias of the forget gate;

[0045] Cell state:

[0046] ;

[0047] in, express Cellular state at any given moment; express cell state at time t; represents Hadamard product;

[0048] output gate:

[0049] ;

[0050] ;

[0051] wherein, represents output of output gate at time t; represents weight of output gate; represents bias of output gate; represents hidden state at time t;

[0052] hidden state output by qth spatio-temporal graph convolutional long short-term memory network unit preliminary prediction result output by corresponding spatio-temporal graph convolutional long short-term memory network, the preliminary prediction result comprising multi-channel predicted values of process variables and multi-channel predicted values of quality variables.

[0053] Preferably, in S3, the decoding link result comprises:

[0054] ;

[0055] wherein, represents predicted value of quality variable at time t; represents multi-layer perception; represents flattening operation; represents linking operation; represents multi-channel predicted values of quality variables output by 1st spatio-temporal graph convolutional long short-term memory network; represents multi-channel predicted values of quality variables output by 2nd spatio-temporal graph convolutional long short-term memory network; represents multi-channel predicted values of quality variables output by qth spatio-temporal graph convolutional long short-term memory network; represents multi-channel predicted values of quality variables output by qth spatio-temporal graph convolutional long short-term memory network; represents number of spatio-temporal graph convolutional long short-term memory networks.

[0056] Preferably, after the quality variables and the process variables at past time are normalized to eliminate dimensional differences, the sliding window method is used for sampling.

[0057] Beneficial effects: the method collects multivariate time series data containing quality variables and process variables, deeply mines the Granger causality relationship between variables based on the multivariate time series data, preliminarily obtains a directed information transmission matrix, and dynamically adjusts and refines the information transmission relationship between variables in combination with the prior process knowledge mask matrix to generate a dynamic adjacency matrix; a multi-head spatiotemporal graph convolution long short-term memory network is designed based on the dynamic adjacency matrix to learn long short-term spatiotemporal features. The method uses the Granger causality relationship based on the constraint prior knowledge to effectively mine the correlation between the quality variables and the process variables, and uses the multi-head spatiotemporal graph convolution long short-term memory network to capture the long short-term dependency relationship, which helps to improve the accuracy of industrial system quality prediction. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creating any inventive labor.

[0059] Figure 1 The flowchart of the industrial quality prediction method based on prior knowledge constraint graph convolution in the embodiments of the present application.

[0060] Figure 2 The schematic diagram of the hydrocracking industrial system in the embodiments of the present application.

[0061] Figure 3 The comparison diagram of the predicted value and the true value of the quality variable in the embodiments of the present application. DETAILED DESCRIPTION

[0062] In order to make the above purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below. In the following description, many specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the spirit of the present application, so the present application is not limited to the specific embodiments disclosed below.

[0063] In addition, the terms "first", "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "multiple" is at least two, such as two, three, etc., unless otherwise specifically limited.

[0064] AsFigure 1 As shown, the embodiment provides an industrial quality prediction method based on prior knowledge constraint graph convolution, comprising:

[0065] S1: Collecting multivariate time series data containing quality variables and process variables in the process industrial system.

[0066] In the embodiment, the process industrial system includes a hydrocracking industrial system, and a structural schematic diagram of the hydrocracking industrial system is as shown in Figure 2 .

[0067] The data used in the embodiment is derived from the light naphtha c5 total data set in the hydrocracking process. Hydrocracking is one of the petroleum refining processes, which is a process of converting heavy oil into gas, gasoline, jet fuel, diesel, etc. under the conditions of heating, high hydrogen pressure and catalyst. In the hydrocracking process, multivariate time series data in the reactor are collected in real time by industrial sensors. These data cover key indicators in the hydrocracking process, which can fully reflect the running state of the hydrocracking process and provide rich information for subsequent data analysis and model construction. These key indicators are process variables; and the quality variable corresponds to the light naphtha c5 total in the hydrocracking process. Each variable is continuously collected at an interval of 4 hours for 36 months, a total of 2606 samples. After normalization to eliminate dimensional differences, the sliding window method (window size of 8) is used for sampling, and the data are divided into training set and test set according to the ratio of 8:2. The past 8 time process variables and quality variable data are used to predict the predicted value of the current time quality variable. The types of process variables are shown in Table 1;

[0068] Table 1 Type table of process variables

[0069] P101A outlet total flow Regulated control temperature of recycle hydrogen to furnace F101 flow R101 one bed top temperature A R101 one bed bottom temperature A R101 two bed top temperature A R101 three bed middle temperature A R101 three bed bottom temperature A R101 reactor pressure differential R102 two bed top temperature A R102 three bed top temperature A R102 four bed top temperature A PI10602 column top pressure Hot lube oil to T201 flow Cold lube oil to E107 flow regulation C102 new hydrogen outlet to cracking flow FI20102 column top reflux flow Stripper column bottom liquid flow To E215 debutanizer flow PI20301 column top pressure FI20302 column top reflux flow Fractionator T202 stripping steam inlet flow regulation TE20315 column bottom temperature indication Mid section return temperature From E205 total recycle tail oil flow regulation To E216 heavy naphtha flow From T202 aviation fuel draw temperature Aviation fuel out of unit From T202 diesel fuel draw temperature TE20402 column top temperature PI20701 column top pressure TI20801 column top reflux flow Heavy naphtha out of unit —— —— ——

[0070] Let the quality variable be ; and the process variable be , wherein, denotes the sample length of the quality variable, denotes the number of process variables.

[0071] S2: Mining the Granger causality between each variable in the multivariate time series data to obtain the information transmission relationship between each variable, and constructing a directed information transmission matrix based on the information transmission relationship between each variable; judging whether each variable belongs to the same sub-process according to the prior knowledge, and constructing a prior sub-process knowledge mask matrix according to the judgment result; fusing the directed information transmission matrix, the prior sub-process knowledge mask matrix and the designed adaptive weight matrix to obtain a dynamic adjacency matrix.

[0072] Specifically, the process of obtaining the information transmission relationship between variables includes:

[0073] For any variable A and any other variable B in a multivariate time series dataset (which can be a quality variable and any process variable, or any process variable and any other variable), construct a univariate autoregressive model of variable A and a multivariate regression model of variable A with respect to variable B.

[0074] When variable A is a quality variable, the univariate autoregressive model of variable A is expressed as:

[0075] ;

[0076] in, express The quality variable at any given time; This indicates the 0th first parameter; This indicates the first parameter; This indicates the second first parameter; Indicates the first The first parameter; express The quality variable at any given time; express The quality variable at any given time; express The quality variable at any given time; Indicates the optimal lag order; express The first residual at time step (the difference between the value calculated by the univariate autoregressive model and the actual value);

[0077] When variable B is the first process variable, the multivariate regression model of variable A with respect to variable B is expressed as:

[0078] ;

[0079] in, This indicates the 0th second parameter; Indicates the first second parameter; This indicates the second parameter; Indicates the first A second parameter; This indicates the 0th third parameter; This indicates the first third parameter; This indicates the second third parameter; Indicates the first A third parameter; express The first process variable at time t; express the first process variable at the time point; denotes the first process variable at the time point; denotes the second residual (the difference between the value calculated by the multivariate regression model and the true value) at the time point; denotes the optimal lag order determined by Akaike information criterion (AIC) of the multivariate regression model; the first parameter, the second parameter and the third parameter are parameters estimated by the least square method;

[0080] Based on the univariate autoregressive model and the multivariate regression model, a joint hypothesis statistic is calculated, and whether the value of the variable B at the past time point can significantly improve the prediction ability of the variable A is determined by joint hypothesis testing, and the calculation formula is:

[0081] ;

[0082] wherein, denotes the joint hypothesis statistic; denotes the residual sum of squares of the univariate autoregressive model; denotes the residual sum of squares of the multivariate regression model; denotes the sample length of the quality variable;

[0083] When the joint hypothesis statistic is greater than a threshold value (0.05 is adopted in the embodiment), it is determined that the variable B is the cause of the variable A, that is, the variable B has an information transmission relationship with the variable A.

[0084] Further, the directed information transmission matrix is constructed based on the information transmission relationship between the variables, which includes:

[0085] When there is an information transmission relationship between the variables, the value is 1, otherwise it is 0; in order to ensure that the information of itself is not lost in the graph convolution process, the information transmission relationship of the variable to itself is determined to exist, that is, when the variable is a variable and the variable itself (i.e. two variables), the value is 1;

[0086] The information transmission relationship between the quality variable and itself and each process variable is taken as the first row of the directed information transmission matrix;

[0087] All the process variables are traversed in order, and for the first process variable, the information transmission relationship between the variable and itself and other variables is taken as the first row of the directed information transmission matrix, , denotes the number of process variables;

[0088] The corresponding values are substituted into each position of the directed information transmission matrix to construct the directed information transmission matrix, denoted as ; and the diagonal elements of the directed information transmission matrix are all 1.

[0089] Further, according to prior knowledge, it is judged whether each variable belongs to the same sub-process, and a prior sub-process knowledge mask matrix is constructed according to the judgment result, including:

[0090] The process industry in the process industry system is a multi-process production mode, and different processes are different sub-processes. Whether any two variables belong to the same sub-process is to judge whether any two variables belong to the same process according to prior knowledge;

[0091] When the two variables belong to the same sub-process, the value is 1 (indicating that the two variables have strong correlation), otherwise it is 0.5 (indicating that the two variables have weak correlation); and when the variable is a loop, the value is 1;

[0092] Whether the quality variable and itself and each process variable belong to the same sub-process is taken as the first row of the prior sub-process knowledge mask matrix;

[0093] All process variables are traversed in order. For the first process variable, whether it and itself and other variables belong to the same sub-process is taken as the first row of the prior sub-process knowledge mask matrix, , wherein n represents the number of process variables;

[0094] The corresponding values are substituted into each position of the prior sub-process knowledge mask matrix to construct the prior sub-process knowledge mask matrix, denoted as ; and the diagonal elements of the prior sub-process knowledge mask matrix are all 1.

[0095] In this embodiment, the Hadamard product between the directed information transmission matrix, the prior sub-process knowledge mask matrix and the designed adaptive weight matrix is calculated to obtain a dynamic adjacency matrix; the adaptive weight matrix is a random weight matrix, and the weight values in the random weight matrix are automatically updated with the model parameters updated in the training process of the multi-head spatio-temporal graph convolution long short-term memory network.

[0096] The multi-head spatio-temporal graph convolution long short-term memory network is trained and optimized on the training set to minimize the error between the predicted value and the true value of the quality variable, and the model parameters of the multi-head spatio-temporal graph convolution long short-term memory network are constantly updated. The weight values in the random weight matrix are automatically updated with the model parameters updated each time.

[0097] S3: a multi-head spatio-temporal graph convolutional long short-term memory network is designed based on a dynamic adjacency matrix, process variables at past time points and quality variables at past time points are input into the multi-head spatio-temporal graph convolutional long short-term memory network, a plurality of preliminary prediction results of the quality variables at the current time point are obtained, and the preliminary prediction results are linked; the linked results are decoded to obtain the predicted values of the quality variables at the current time point.

[0098] In the embodiment, in order to fully exert the role of the subsequent spatio-temporal graph convolutional long short-term memory network, the channel dimensions of the quality variables and the process variables in the multivariate time series data are respectively embedded into Q dimensions, i.e., the dimensions of the multivariate time series data are embedded into . .

[0099] In the embodiment, the multi-head spatio-temporal graph convolutional long short-term memory network is formed by spatio-temporal graph convolutional long short-term memory networks in parallel, each spatio-temporal graph convolutional long short-term memory network is formed by q=8 spatio-temporal graph convolutional long short-term memory network units in series, and each spatio-temporal graph convolutional long short-term memory network unit includes an input gate, a forget gate, a cell state, and an output gate, and the input gate, the forget gate, and the output gate perform directed graph convolution on the input objects through a dynamic adjacency matrix; the expression of the directed graph convolution is:

[0100] ;

[0101] wherein, represents an adaptive weight; represents a directed graph convolution operation; represents an object of the directed graph convolution; represents a gelu activation function; represents a variable in-degree matrix; represents a dynamic adjacency matrix; represents an adaptive bias;

[0102] The variable in-degree matrix is a diagonal matrix, and each value on the diagonal line is the sum of the element values of the corresponding column in the directed information transmission matrix.

[0103] Specifically, the process variables and the quality variables at the past q consecutive time points are sampled by a sliding window to predict the value of the quality variable at the current time point, and any spatio-temporal graph convolutional long short-term memory network unit includes:

[0104] Input gate:

[0105] ;

[0106] ;

[0107] wherein, denotes the output of the input gate at time step t; denotes the gelu activation function; denotes layer normalization; denotes the weights of the input gate; denotes the directed graph convolution operation; denotes the hidden state at time step t; denotes the process variable at time step t; denotes the link operation; denotes the bias of the input gate; denotes the candidate cell state at time step t; denotes the hyperbolic tangent function; denotes the weights corresponding to the candidate cell state; denotes the bias corresponding to the candidate cell state;

[0108] the forget gate:

[0109] ;

[0110] wherein, denotes the output of the forget gate at time step t; denotes the weights of the forget gate; denotes the bias of the forget gate;

[0111] the cell state:

[0112] ;

[0113] wherein, denotes the cell state at time step t; denotes the cell state at time step t; denotes the Hadamard product;

[0114] the output gate:

[0115] ;

[0116] ;

[0117] wherein, denotes the output of the output gate at time step t; denotes the weights of the output gate; denotes the bias of the output gate; denotes the hidden state at time step t;

[0118] In any one spatio-temporal graph convolutional long short-term memory network unit, the dimensions of parameters 、 、 、 、 、 、 、 、 are all ;

[0119] The hidden state output by the q=8th spatio-temporal graph convolutional long short-term memory network unit is a preliminary prediction result output by the spatio-temporal graph convolutional long short-term memory network, and the preliminary prediction result includes multi-channel predicted values of process variables and multi-channel predicted values of quality variables .

[0120] Further, the decoding link result includes:

[0121] ;

[0122] wherein denotes the predicted value of the quality variable at the t moment; denotes a multi-layer perception; denotes a flattening operation; denotes a linking operation; denotes the multi-channel predicted values of the quality variable output by the 1st spatio-temporal graph convolutional long short-term memory network; denotes the multi-channel predicted values of the quality variable output by the 2nd spatio-temporal graph convolutional long short-term memory network; denotes the multi-channel predicted values of the quality variable output by the qth spatio-temporal graph convolutional long short-term memory network; denotes the number of spatio-temporal graph convolutional long short-term memory networks. The comparison between the predicted value and the true value of the quality variable is shown as indicated by . Figure 3 The industrial quality prediction method based on prior knowledge constraint graph convolution provided by the embodiment has the following beneficial effects:

[0123]

[0124] ​​The method collects multivariate time series data containing quality variables and process variables, deeply mines the Granger causality relationship between variables based on the multivariate time series data, preliminarily obtains a directed information transmission matrix, and dynamically adjusts and refines the information transmission relationship between variables in combination with a priori process knowledge mask matrix to generate a dynamic adjacency matrix; the directed graph convolution operation is performed through the dynamic adjacency matrix, which replaces the traditional information bidirectional transmission and fixed transmission between variables in a complex multivariate industrial system, and a multi-head spatiotemporal graph convolution long short-term memory network is used to learn long short-term spatiotemporal features. The method uses the Granger causality relationship based on the constraint prior knowledge to effectively mine the correlation between the quality variables and the process variables, and uses the multi-head spatiotemporal graph convolution long short-term memory network to capture the long short-term dependency relationship, which helps to improve the accuracy of the industrial system quality prediction.

[0125] The technical features of the above-described embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described, but it should be considered that any combination of the technical features is within the scope of the present disclosure as long as the combination does not result in contradictions.

[0126] The above-described embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the patent scope of the application. It should be noted that, for those skilled in the art, some modifications and improvements can be made without departing from the concept of the present application, and these are all within the protection scope of the present application. Therefore, the patent protection scope of the present application should be subject to the appended claims.

Claims

1. An industrial quality prediction method based on priori knowledge constraint graph convolution, characterized in that, Comprise: S1: collect multivariate time series data containing quality variables and process variables in a process industry system; S2: mine Granger causality between variables in the multivariate time series data to obtain information transmission relationship between variables, and construct a directed information transmission matrix based on the information transmission relationship between variables; determine whether each variable belongs to the same sub-process based on prior knowledge, and construct a prior sub-process knowledge mask matrix according to the determination result; fuse the directed information transmission matrix, the prior sub-process knowledge mask matrix and the designed adaptive weight matrix to obtain a dynamic adjacency matrix; S3: design a multi-head spatio-temporal graph convolution long short-term memory network based on the dynamic adjacency matrix, input the process variables at past time and the quality variables at past time into the multi-head spatio-temporal graph convolution long short-term memory network to obtain a plurality of preliminary prediction results of the quality variables at current time, and link the preliminary prediction results; decode the linked results to obtain the predicted values of the quality variables at current time; The multi-head spatiotemporal graph convolution long short-term memory network is formed by parallel connection of q spatiotemporal graph convolution long short-term memory networks, each of which is formed by series connection of q spatiotemporal graph convolution long short-term memory network units, each of which comprises an input gate, a forget gate, a cell state and an output gate, and the input gate, the forget gate and the output gate perform directed graph convolution on an input object through a dynamic adjacency matrix; a directed graph convolution expression is: ； wherein, denotes an adaptive weight; denotes a directed graph convolution operation; denotes an object of directed graph convolution; denotes a gelu activation function; denotes a variable in-degree matrix; denotes a dynamic adjacency matrix; denotes an adaptive bias; The variable in-degree matrix is a diagonal matrix, and each value on the diagonal is the sum of the element values of the corresponding column in the directed information transmission matrix.

2. The industrial quality prediction method based on priori knowledge constrained graph convolution according to claim 1, characterized in that, The process industry system comprises a hydrocracking industrial system; the quality variables comprise light naphtha c5 total in the hydrocracking process; and the process variables comprise bed top temperature, bed bottom temperature, reactor pressure difference, tower top pressure and tower top reflux amount of the hydrocracking industrial system in the hydrocracking process.

3. The method of claim 1, wherein the method comprises: In S2, the process of obtaining the information transmission relationship between variables comprises: For any variable A and any other variable B in the multivariate time series data, a univariate autoregressive model of variable A and a multivariate regression model of variable A with respect to variable B are constructed; When variable A is a quality variable, the univariate autoregressive model of variable A is expressed as: ； wherein represents the quality variable at the time point t; represents the 0th first parameter; represents the 1st first parameter; represents the 2nd first parameter; represents the th first parameter; represents the quality variable at the time point t; represents the quality variable at the time point t; represents the quality variable at the time point t; represents the optimal lag order; represents the first residual at the time point t; When variable B is the first process variable, the multivariate regression model of variable A with respect to variable B is expressed as: ； wherein, represents the 0th second parameter; represents the 1st second parameter; represents the 2nd second parameter; represents the th second parameter; represents the 0th third parameter; represents the 1st third parameter; represents the 2nd third parameter; represents the th third parameter; represents the first process variable at the time instant t; represents the first process variable at the time instant t; represents the first process variable at the time instant t; represents the second residual at the time instant t; Based on the univariate autoregressive model and the multivariate regression model, a joint hypothesis statistic is calculated, and the calculation formula is: ； wherein, denotes the joint hypothesis statistic; denotes the residual sum of squares of a univariate autoregressive model; denotes the residual sum of squares of a multivariate regression model; denotes the sample length of the quality variable; When the joint hypothesis statistic is greater than a threshold value, it is determined that variable B is the cause of variable A, that is, variable B has an information transmission relationship with variable A.

4. The industrial quality prediction method based on priori knowledge constrained graph convolution according to claim 3, characterized in that, In S2, the directed information transmission matrix is constructed based on the information transmission relationship between variables, comprising: When there is an information transmission relationship between variables, its value is 1, otherwise it is 0; and when the variable is a loop, its value is 1; The information transmission relationship between the quality variables and themselves and the process variables is taken as the first row of the directed information transmission matrix; All process variables are traversed in order, and for the first process variable, the information passing relations between it and itself and other variables are taken as the first row, , of the directed information passing matrix, where N represents the number of process variables. The corresponding values are substituted into each position of the directed information transmission matrix to construct the directed information transmission matrix, and the diagonal elements of the directed information transmission matrix are all 1.

5. The method of claim 3, wherein the method is characterized by, In S2, whether each variable belongs to the same sub-process is determined based on prior knowledge, and a prior sub-process knowledge mask matrix is constructed according to the determination result, comprising: The process industry in the process industry system is a multi-process production mode, and different processes are different sub-processes. Whether any two variables belong to the same sub-process is to determine whether any two variables belong to the same process according to prior knowledge. When two variables belong to the same sub-process, the value is 1, otherwise 0.5; and when the variable is self-loop, the value is 1; Taking whether the quality variable and each process variable belong to the same sub-process as the first row of the prior sub-process knowledge mask matrix; All process variables are traversed in order, and for the first process variable, it is checked whether it belongs to the same sub-process as itself and the other variables as a priori sub-process knowledge mask matrix of the first row, , denotes the number of process variables; The corresponding values are substituted into each position of the prior sub-process knowledge mask matrix to construct the prior sub-process knowledge mask matrix, and the diagonal elements of the prior sub-process knowledge mask matrix are all 1.

6. The industrial quality prediction method based on priori knowledge constrained graph convolution according to claim 3, characterized in that, In S2, the Hadamard product of the directed information transmission matrix, the prior sub-process knowledge mask matrix and the designed adaptive weight matrix is calculated to obtain a dynamic adjacency matrix; The adaptive weight matrix is a random weight matrix, and the weight values in the random weight matrix are automatically updated with the model parameters updated in the training process of the multi-head spatio-temporal graph convolution long short-term memory network.

7. The method of claim 1, wherein, The past q consecutive time process variables and quality variables are sampled by the sliding window to predict the value of the quality variable at the current time, and any one spatio-temporal graph convolution long short-term memory network unit comprises: Input gate: ； ； wherein, represents the output of the input gate at time step t; represents the gelu activation function; represents layer normalization; represents the weight of the input gate; represents the directed graph convolution operation; represents the hidden state at time step t; represents the process variable at time step t; represents the link operation; represents the bias of the input gate; represents the candidate cell state at time step t; represents the hyperbolic tangent function; represents the weight corresponding to the candidate cell state; represents the bias corresponding to the candidate cell state; Forget gate: ； wherein, represents the output of the forget gate at the time instant; represents the weight of the forget gate; represents the bias of the forget gate; Cell state: ； wherein denotes cell state at time t; denotes cell state at time t; denotes Hadamard product; Output gate: ； ； wherein, denotes the output of the output gate at time step t; denotes the output of the output gate at time step t; denotes the weight of the output gate; denotes the bias of the output gate; denotes the output of the output gate at time step t; denotes the hidden state at time step t; a hidden state output by the qth spatio-temporal graph convolutional long short-term memory network unit a preliminary prediction result corresponding to an output of the spatio-temporal graph convolutional long short-term memory network, the preliminary prediction result including a multi-channel predicted value of the process variable and a multi-channel predicted value of the quality variable.

8. The industrial quality prediction method based on priori knowledge constrained graph convolution according to claim 7, characterized in that, In S3, the decoding link result comprises: ； wherein, denotes a predicted value of the quality variable at the time instant; denotes a multi-layer perceptron; denotes a flattening operation; denotes a concatenation operation; denotes a multi-channel predicted value of the quality variable output by the 1st spatio-temporal graph convolutional long short-term memory network; denotes a multi-channel predicted value of the quality variable output by the 2nd spatio-temporal graph convolutional long short-term memory network; denotes a multi-channel predicted value of the quality variable output by the i-th spatio-temporal graph convolutional long short-term memory network; denotes a multi-channel predicted value of the quality variable output by the i-th spatio-temporal graph convolutional long short-term memory network; denotes the number of spatio-temporal graph convolutional long short-term memory networks. 9.The industrial quality prediction method based on priori knowledge constrained graph convolution according to claim 1, wherein, The past time quality variables and process variables are sampled by the sliding window method after normalization to eliminate the dimensional difference.

Images