Industrial key quality index prediction method based on multi-level space-time correlation characteristics

Abstract

The invention is suitable for the technical field of data processing, and provides an industrial key quality index prediction method based on a multi-level space-time correlation feature, and the method comprises the steps: determining a process variable related to a key quality index, and obtaining a training sample; taking the whole process and each sub-process as a target process, and executing the following steps for each target process: screening out a target process variable corresponding to the target process from all the determined process variables, and obtaining a graph adjacency matrix of the target process by utilizing an adjacency matrix model and target process variable data; utilizing a space-time dependent perception feature extraction model to obtain space-time feature representation, corresponding to the training sample, of the target process; optimizing the adjacency matrix model, the space-time dependent perception feature extraction model and the prediction model based on the obtained space-time feature representation; and predicting the industrial key quality index at the prediction moment by using the model after parameter optimization. According to the invention, the accuracy and reliability of industrial key quality index prediction can be improved.

Description

Technical Field

[0001] The present application belongs to the field of data processing technology, and in particular relates to an industrial key quality indicator prediction method based on multi-level spatiotemporal correlation features. Background Art

[0002] With the continuous advancement of modern industrial technology, industrial production processes are becoming increasingly large-scale and complex. These processes involve a series of complex material energy conversions and physical and chemical reactions, aiming to efficiently convert raw materials into products. In this context, real-time optimization and control of the entire production process are essential to improving production efficiency and reducing energy consumption. Generally, real-time detection of key quality indicators is the most effective reflection of the industrial production status and the basis for achieving real-time optimization control. As a key tool for modeling complex industrial processes and promoting the transformation of process industries to intelligent manufacturing, it plays an important role in industrial process control.

[0003] However, due to the limitations of measurement technology and industrial environment, most of these quality indicators cannot be measured in time and need to be determined through laboratory testing, which leads to serious time lags in subsequent industrial process optimization and control, greatly reducing the timeliness and effectiveness of industrial process control. Due to the continuous accumulation of industrial data in distributed control systems, data-driven soft sensor modeling methods have become more effective. By constructing mathematical models between key quality indicators that are difficult to measure and process variables that are easy to measure, such as temperature, liquid level, flow, and current, key quality indicator perception can be indirectly achieved. In recent years, due to its outstanding ability in high-dimensional data processing and complex function representation, deep learning has been widely used in data-driven soft sensor modeling of industrial processes.

[0004] Although the prediction of key quality indicators in industrial processes has been widely studied, there are still several key issues that need to be addressed: 1) The entire industrial production process is usually composed of sub-processes consisting of multiple groups of different equipment units. Existing data-driven models often only use the global information of the entire industrial production process to guide the feature extraction process, while ignoring the local information of each group of sub-processes. In fact, there is a mutual coupling relationship between the production equipment units of each group of sub-processes, and they are affected by uncertain factors such as changes in operating conditions and fluctuations in production demand, resulting in differences in the working conditions of each group of sub-processes, showing different local characteristics. 2) In space, there is an interaction between the process variables of the industrial process of material flow and energy flow, and due to the influence of process connections, there are complex nonlinear coupling relationships; in time, due to the complex physical and chemical reaction mechanisms, feedback control and changes in operating conditions, industrial processes have inherently complex nonlinear dynamic behaviors. It can be seen that the current prediction of industrial key quality indicators has the problem of poor accuracy. Summary of the invention

[0005] The embodiment of the present application provides an industrial key quality indicator prediction method based on multi-level spatiotemporal correlation features, which can solve the problem of poor prediction accuracy of industrial key quality indicators.

[0006] The present application embodiment provides an industrial key quality indicator prediction method based on multi-level spatiotemporal correlation features, including:

[0007] Determine process variables related to key quality indicators of industrial production and obtain multiple training samples; the training samples include process variable data at historical moments and corresponding key quality indicator values; the entire industrial production process includes multiple sub-processes;

[0008] The whole process and each sub-process are respectively regarded as a target process, and are executed for each target process respectively: for each training sample respectively, the target process variables corresponding to the target process are screened out from all the determined process variables, and the graph adjacency matrix of the target process corresponding to the training sample is obtained according to the adjacency matrix model and the target process variable data corresponding to the target process variables in the training sample; the spatiotemporal feature representation of the target process corresponding to the training sample is obtained by using the spatiotemporal dependency perception feature extraction model; the spatiotemporal dependency perception feature extraction model includes a skip connection module and L sequentially connected spatiotemporal dependency perception blocks, the output ends of the L spatiotemporal dependency perception blocks are connected to the input end of the skip connection module, and the spatiotemporal dependency perception block includes a spatial topology feature extraction module The output end of the spatial topology feature extraction module and the output end of the temporal dynamic feature extraction module are connected to the input end of the spatial time feature fusion module. The output end of the spatial time feature fusion module is the output end of the spatial time dependency perception block. The input end of the spatial topology feature extraction module and the input end of the temporal dynamic feature extraction module are both the input end of the spatial time dependency perception block. The input of the spatial topology feature extraction module of the first spatial time dependency perception block among the L spatial time dependency perception blocks is the graph adjacency matrix and target process variable data of the target process corresponding to the training samples. The input of the temporal dynamic feature extraction module of the first spatial time dependency perception block is the target process variable data of the target process corresponding to the training samples.

[0009] For each training sample, according to the prediction model and the spatiotemporal feature representation of all target processes corresponding to the training sample, obtain the key quality indicator prediction values ​​of all target processes corresponding to the training sample, and use the key quality indicator prediction values ​​and key quality indicator values ​​corresponding to all training samples to optimize the parameters in the adjacency matrix model, the spatiotemporal dependency perception feature extraction model and the prediction model;

[0010] The adjacency matrix model with optimized parameters, the spatiotemporal dependency-aware feature extraction model and the prediction model are used to predict the industrial key quality indicators at the prediction moment.

[0011] Optionally, the adjacency matrix model includes an adjacency weight formula and a matrix fusion formula; obtaining a graph adjacency matrix of the target process corresponding to the training sample according to the adjacency matrix model and the target process variable data corresponding to the target process variable in the training sample, including:

[0012] For each target process variable among all the target process variables, respectively: generating a node embedding based on the target process variable data corresponding to the target process variable; for each other target process variable except the target process variable among all the target process variables, respectively, generating a node embedding based on the target process variable data corresponding to the other target process variables, and calculating the adjacency weights between the target process variable and other target process variables according to the generated node embeddings and the adjacency weight formula;

[0013] An asymmetric adjacency matrix is ​​constructed based on all adjacency weights, and the asymmetric adjacency matrix is ​​sparsely processed to obtain a sparse asymmetric adjacency matrix; each element in the asymmetric adjacency matrix is ​​the adjacency weight of two target process variables among all target process variables, and each element corresponds to a different adjacency weight;

[0014] Target process variable data corresponding to the target process variable are screened out from multiple training samples, and the correlation between every two target process variables is calculated based on the target process variable data corresponding to every two target process variables among all the target process variables, and a correlation adjacency matrix is ​​constructed based on all the calculated correlations; each element in the correlation adjacency matrix is ​​the correlation between two target process variables among all the target process variables, and each element corresponds to a different correlation;

[0015] The sparse asymmetric adjacency matrix and the correlation adjacency matrix are fused through the matrix fusion formula to obtain the graph adjacency matrix of the target process corresponding to the training samples.

[0016] Optionally, the adjacency weight formula is:

[0017] M 1i =tanh(αE 1i Γ1)

[0018] M 2j =tanh(αE 2j Γ2)

[0019]

[0020] in, E 1i represents the node embedding corresponding to the i-th target process variable, E 2j represents the node embedding corresponding to the jth target process variable, M represents the number of target process variables, d represents the node embedding dimension, Γ1 and Γ2 are both learnable parameters, α is a hyperparameter that controls the saturation rate of the activation function, tanh(·) and ReLU(·) are both activation functions, and A learnij Represents the adjacency weight between the i-th target process variable and the j-th target process variable, i≠j.

[0021] Optionally, the matrix fusion formula is:

[0022] A=α′A corr +(1-α)A learn

[0023] Where A represents the graph adjacency matrix, α′ represents the learnable trade-off parameter, and A corr represents the correlation adjacency matrix, A learn represents the asymmetric adjacency matrix after sparsification.

[0024] Optionally, the spatial topological feature extraction module includes a spatial position attention sub-module and a hybrid skip graph convolution sub-module connected in sequence, the input end of the spatial position attention sub-module is the input end of the spatial topological feature extraction module, and the output end of the hybrid skip graph convolution sub-module is the output end of the spatial topological feature extraction module.

[0025] Optionally, the temporal dynamic feature extraction module includes a temporal position attention sub-module and a multi-scale dilated causal convolution sub-module connected in sequence, the input end of the temporal position attention sub-module is the input end of the temporal dynamic feature extraction module, and the output end of the multi-scale dilated causal convolution sub-module is the output end of the temporal dynamic feature extraction module.

[0026] Optionally, the spatiotemporal feature fusion module is used to fuse the spatial topological features output by the spatial topological feature extraction module and the temporal dynamic features output by the temporal dynamic feature extraction module through a feature fusion formula to obtain a spatiotemporal feature fusion result;

[0027] The feature fusion formula is: Z l represents the spatiotemporal feature fusion result output by the spatiotemporal feature fusion module of the lth spatiotemporal dependency perception block, s l represents the spatial attention vector of the lth spatiotemporal dependency perception block, represents the spatial topological features output by the spatial topological feature extraction module of the lth spatiotemporal dependency perception block, t l represents the temporal attention vector of the lth spatiotemporal dependency perception block, Represents the temporal dynamic features output by the temporal dynamic feature extraction module of the lth spatiotemporal dependency perception block.

[0028] Optional,

[0029]

[0030] Among them, α1, α2, β1, and β2 are all 1×1 convolutional layers, and σ represents the sigmoid activation function.

[0031] Optionally, based on the prediction model and the spatiotemporal feature representation of all target processes corresponding to the training samples, the key quality indicator prediction values ​​of all target processes corresponding to the training samples are obtained, including:

[0032] For the whole process, the spatiotemporal feature representations of all target processes corresponding to the training samples are fused, and the fused spatiotemporal feature representations are input into the prediction model for processing to obtain the predicted values ​​of the key quality indicators of the whole process corresponding to the training samples;

[0033] For each sub-process, the spatiotemporal feature representation of the sub-process corresponding to the training sample is input into the prediction model for processing to obtain the predicted value of the key quality indicator of the sub-process corresponding to the training sample.

[0034] Optionally, the parameters in the adjacency matrix model, the spatiotemporal dependency-aware feature extraction model, and the prediction model are optimized using the key quality indicator prediction values ​​and key quality indicator values ​​corresponding to all training samples, including:

[0035] Calculate the prediction loss value based on the loss function, and use the prediction loss value to optimize the parameters in the adjacency matrix model, the spatiotemporal dependency perception feature extraction model and the prediction model;

[0036] The loss function is:

[0037]

[0038] Among them, J(Θ) represents the prediction loss value, Θ represents the learnable parameters in the adjacency matrix model, spatiotemporal dependency-aware feature extraction model and prediction model, N represents the number of training samples, groupN is the number of subprocesses, represents the predicted value of the key quality indicator of the whole process corresponding to the i-th training sample, y i represents the key quality indicator value in the i-th training sample, It represents the predicted value of the key quality indicator of the j-th sub-process corresponding to the i-th training sample.

[0039] The above solution of the present application has the following beneficial effects:

[0040] In an embodiment of the present application, the entire industrial production process is divided into multiple sub-processes, and a graph adjacency matrix of the entire process and each sub-process is constructed. On the basis of each graph adjacency matrix, a global spatiotemporal feature representation of the entire process and a local spatiotemporal feature representation of each sub-process are extracted using a multi-level spatiotemporal dependency perception block. Then, the adjacency matrix model, the spatiotemporal dependency perception feature extraction model, and the prediction model are optimized in combination with the global spatiotemporal feature representation and the local spatiotemporal feature representation. The adjacency matrix model, the spatiotemporal dependency perception feature extraction model, and the prediction model can learn the underlying topological structure relationship between each process variable related to the key quality indicators of industrial production, and mine the spatial dependency between the process variables and the temporal dependency in the time series data. Therefore, when the optimized adjacency matrix model, the spatiotemporal dependency perception feature extraction model, and the prediction model are used to predict the key industrial quality indicators, the accuracy and reliability of the prediction of the key industrial quality indicators can be greatly improved.

[0041] Other beneficial effects of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1 A flowchart of an industrial key quality indicator prediction method based on multi-level spatiotemporal correlation features provided in an embodiment of the present application;

[0044] Figure 2 A schematic diagram of the structure of a spatiotemporal dependency perception feature extraction model provided in one embodiment of the present application;

[0045] Figure 3 A simplified flow chart of a potassium salt flotation process in the related art;

[0046] Figure 4 This is a comparison curve of the predicted value and the actual value of the potassium ion grade of the concentrate in the potassium salt flotation process by the LSTNet model in an example;

[0047] Figure 5 This is a comparison curve of the predicted value and the actual value of the potassium ion grade of the concentrate in the potassium salt flotation process using the VW-SAE model in an example;

[0048] Figure 6 This is a comparison curve of the predicted value and the actual value of the potassium ion grade of the concentrate in the potassium salt flotation process using the LogTrans model in an example;

[0049] Figure 7 The following is a comparison curve of the predicted value and the actual value of the potassium ion grade of the concentrate in the potassium salt flotation process according to the prediction method of the present application in an example. DETAILED DESCRIPTION

[0050] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0051] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0052] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0053] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0054] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0055] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0056] In response to the current problem of poor prediction accuracy of industrial key quality indicators, an embodiment of the present application provides an industrial key quality indicator prediction method based on multi-level spatiotemporal correlation features. The entire industrial production process is divided into multiple sub-processes, and a graph adjacency matrix of the entire process and each sub-process is constructed. On the basis of each graph adjacency matrix, a multi-level spatiotemporal dependency perception block is used to extract the global spatiotemporal feature representation of the entire process and the local spatiotemporal feature representation of each sub-process. Then, the adjacency matrix model, the spatiotemporal dependency perception feature extraction model and the prediction model are optimized in combination with the global spatiotemporal feature representation and the local spatiotemporal feature representation. The adjacency matrix model, the spatiotemporal dependency perception feature extraction model and the prediction model can learn the underlying topological structure relationship between each process variable related to the key quality indicators of industrial production, and mine the spatial dependency between the process variables and the temporal dependency in the time series data, so that when the optimized adjacency matrix model, the spatiotemporal dependency perception feature extraction model and the prediction model are used to predict the industrial key quality indicators, the accuracy and reliability of the industrial key quality indicator prediction can be greatly improved.

[0057] The following is an exemplary description of the industrial key quality indicator prediction method based on multi-level spatiotemporal correlation features provided in the embodiments of the present application in conjunction with specific embodiments.

[0058] like Figure 1 As shown, the industrial key quality indicator prediction method based on multi-level spatiotemporal correlation features provided in the embodiment of the present application includes the following steps:

[0059] Step 11, determine process variables related to key quality indicators of industrial production and obtain multiple training samples.

[0060] The above training samples include the process variable data and the corresponding key quality indicator values ​​at the historical moment. The whole process of industrial production includes multiple sub-processes. Specifically, the whole process can be divided into multiple sub-processes based on the equipment involved in industrial production as a unit, that is, the process flow corresponding to each (or multiple) equipment in industrial production is taken as a sub-process.

[0061] For example, taking the industrial scenario of potassium salt flotation as an example, the concentrate potassium ion grade can be determined as the key quality indicator of the potassium salt flotation process through the process flow mechanism, and multiple process variables (such as temperature, liquid level, flow rate, current, etc.) with high correlation with the concentrate potassium ion grade are selected from the potassium salt flotation process data as auxiliary variables for modeling. Correspondingly, the above training samples include process variable data (i.e., temperature value, liquid level value, flow rate value, current value, etc.) at historical moments and corresponding key quality indicator values ​​(i.e., concentrate potassium ion grade value), and the historical moments corresponding to each training sample are different. In the real-time case, 2117 time industrial process data sample points can be obtained, and each sample point includes 60 dimensional features.

[0062] To facilitate subsequent data processing, after obtaining the historical industrial process data (i.e., the process variable data at historical moments and the corresponding key quality indicator values), it can be preprocessed, the preprocessed data can be segmented by time window using sliding window technology, and the input convolutional layer can be used to map the industrial process data features to the specified dimension to obtain the training samples.

[0063] In the specific implementation, the historical industrial process data is first preprocessed by filling missing values, removing outliers, and filtering data. Then the collected key quality indicator data (i.e., key quality indicator values) and their modeling auxiliary variables (i.e., process variables) are normalized. is the time series data set after preprocessing, where X and Y represent the modeling auxiliary variables and key quality indicators, respectively, N and M represent the total number of samples and the number of auxiliary variables, respectively. The calculation formula for normalizing χ is as follows:

[0064] The modeling auxiliary variable X is normalized by the following formula:

[0065] X=(XX min,i ) / (X max,i -X min,i )

[0066] The key quality indicator Y is normalized by the following formula:

[0067] Y=(YY min ) / (Y max -Y min )

[0068] Where X min,i and X max,i Respectively represent the minimum and maximum values ​​of the i-th auxiliary variable, Y min and Y max They respectively represent the minimum and maximum values ​​of the key quality indicator (i.e., the potassium ion grade of the concentrate).

[0069] In order to facilitate the subsequent model to extract time dynamic features, the sliding window technology is further used to segment the industrial process data set according to the time window, and the time window data is obtained as the original input data. Among them, R is the lookback window length of the historical sample set required for the model to extract time dynamic features, P is the prediction window length of the key sample indicator, and the step size of the window sliding is set to 1. The obtained time window data can be expressed as the following formula:

[0070]

[0071] Among them, X w and Y w Respectively represent auxiliary variables and key quality indicators in time window data, Slidingwindow R+P (·) means splitting the data set into time windows using a window of length R+P and a step size of 1.

[0072] Next, in order to more efficiently extract and process key spatiotemporal features, a 1×1 input convolution layer is used to map the data features to a high-dimensional latent space to obtain the original input data features of the model:

[0073]

[0074] Startconv(·) represents the input convolution operation, is the original input data feature of the subsequent spatiotemporal dependency perception feature extraction module, and c is the data feature dimension.

[0075] The first 90% of the time window data is used as the training set, and the last 10% of the time window data is used as the test set for subsequent testing of the generalization performance of the trained model. Among them, the auxiliary variables and key quality indicators in each time window data are a training sample.

[0076] Step 12, take the whole process and each sub-process as a target process respectively, and execute for each target process respectively: for each training sample respectively, select the target process variables corresponding to the target process from all the determined process variables, and obtain the graph adjacency matrix of the target process corresponding to the training sample based on the adjacency matrix model and the target process variable data corresponding to the target process variables in the training sample; use the spatiotemporal dependency-aware feature extraction model to obtain the spatiotemporal feature representation of the target process corresponding to the training sample.

[0077] In some embodiments of the present application, in order to facilitate the training of the model and enable the model to learn the underlying topological structure relationship between each process variable, it is necessary to obtain the spatiotemporal feature representation of each training sample in the entire process and each sub-process. Since the data processing process is the same in each process, for the convenience of description, the entire process and each sub-process are regarded as a target process, and the target process is taken as an example to illustrate the process of obtaining the spatiotemporal feature representation.

[0078] In a target process, it is necessary to obtain the spatiotemporal feature representation of each training sample. Specifically, for each training sample, it is first necessary to screen out the target process variable corresponding to the current target process from all the process variables determined in step 11, and then determine the process variable data corresponding to the target process variable (i.e., the target process variable data) from the process variable data of the training sample, and then obtain the graph adjacency matrix of the target process corresponding to the training sample based on the adjacency matrix model and the target process variable data corresponding to the target process variable in the training sample.

[0079] The adjacency matrix model includes an adjacency weight formula and a matrix fusion formula. According to the adjacency matrix model and the target process variable data corresponding to the target process variables in the training samples, the specific implementation method of obtaining the graph adjacency matrix of the target process corresponding to the training samples includes the following steps:

[0080] Step 12.1 is performed for each target process variable among all target process variables respectively: generating a node embedding based on the corresponding target process variable data of the target process variable; generating a node embedding based on the corresponding target process variable data of the other target process variables for each other target process variable except the target process variable among all target process variables respectively, and calculating the adjacency weight between the target process variable and other target process variables according to the generated node embedding and the adjacency weight formula.

[0081] The adjacency weight formula is:

[0082] M 1i =tanh(αE 1i Γ1)

[0083] M 2j =tanh(αE 2j Γ2)

[0084]

[0085] in, E 1i represents the node embedding corresponding to the i-th target process variable, E 2j represents the node embedding corresponding to the jth target process variable, M represents the number of target process variables, d represents the node embedding dimension, Γ1 and Γ2 are parameters that can be learned during training, and α is a hyperparameter that controls the saturation rate of the activation function. The asymmetry of the adaptive adjacency matrix is ​​realized. Tanh(·) and ReLU(·) both represent activation functions. learnij represents the adjacency weight between the i-th target process variable and the j-th target process variable, i≠j, j=1,…,T, i=1,…,T, and T represents the number of target process variables. Exemplarily, when generating node embeddings, the target process variable data can be input into the embedding layer for processing to obtain the corresponding node embeddings.

[0086] Step 12.2, construct an asymmetric adjacency matrix based on all adjacency weights, and perform sparse processing on the asymmetric adjacency matrix to obtain a sparse asymmetric adjacency matrix. Each element in the asymmetric adjacency matrix is ​​the adjacency weight of two target process variables among all target process variables, and each element corresponds to a different adjacency weight, that is, each element corresponds to a different target process variable.

[0087] In some embodiments of the present application, a topk algorithm may be used to select the first k nearest nodes (i.e. other target process variables) of each node (i.e. the target process variable) as its neighbors, while retaining the weights of the connected nodes and setting the weights of the non-connected nodes to zero, as shown in the following expression:

[0088] for i=1,2,…,M,M+1

[0089] idx = argtopk(A learn [i,:])

[0090] A learn [i:-idx]=0

[0091] Among them, argtopk(·) represents the first k most relevant indexes of the returned vector, i represents the i-th node index, and M represents the number of target process variables. The above expression means that for a target process variable, the first k adjacent weights between it and other target process variables are retained in the order of roughly small adjacent weights, and the other adjacent weights are set to zero.

[0092] Step 12.3, select the target process variable data corresponding to the target process variable from the multiple training samples, and calculate the correlation between every two target process variables according to the target process variable data corresponding to every two target process variables among all the target process variables, and construct a correlation adjacency matrix based on all the calculated correlations. Each element in the correlation adjacency matrix is ​​the correlation between two target process variables among all the target process variables, and each element corresponds to a different correlation, that is, each element corresponds to a different target process variable.

[0093] In some embodiments of the present application, the Pearson correlation coefficient can be used to calculate the correlation between the target process variables, that is, the Pearson correlation coefficient between the target process variables is used as the correlation between the two. It should be noted that when calculating the correlation between the target process variables, the target process variable data in all training samples are used for calculation. Taking the two process variables of temperature and liquid level as examples, it is necessary to treat the temperature values ​​in all training samples as a sequence, and the liquid level values ​​in all training samples as a sequence, and then calculate the Pearson correlation coefficient of the two sequences.

[0094] The calculation formula of Pearson correlation coefficient is: Pearson(X i ,X j ) represents the sequence X i With sequence X j The Pearson correlation coefficient between them, cov(·) represents the sequence X i With sequence X j The covariance between them, var(·) represents the variance of the sequence, and the sequence X i represents the sequence corresponding to the i-th target process variable, the sequence X j Represents the sequence corresponding to the j-th target process variable.

[0095] After calculating the correlation, the correlation adjacency matrix can be obtained by using the threshold parameter k. Specifically, the element A in the correlation adjacency matrix can be determined by the following formula: corr,ij :

[0096]

[0097] A corr,ij represents the correlation weight between the i-th target process variable and the j-th target process variable, i≠j, j=1,…,T, i=1,…,T, and T represents the number of target process variables. It can be understood that after the correlation weights between the target process variables are determined, a correlation adjacency matrix is ​​constructed based on all the correlation weights, and each element in the correlation adjacency matrix is ​​the correlation (i.e., correlation weight) between two target process variables among all the target process variables, and each element corresponds to a different correlation.

[0098] Step 12.4, the sparse asymmetric adjacency matrix and the correlation adjacency matrix are fused through the matrix fusion formula to obtain the graph adjacency matrix of the target process corresponding to the training sample.

[0099] The above matrix fusion formula is:

[0100] A=α′A corr +(1-α)Alearn

[0101] Where A represents the graph adjacency matrix, and α′ represents a learnable trade-off parameter used to adjust the correlation adjacency matrix A. corr and the asymmetric adjacency matrix A learn In order to better capture the underlying complex relationships in industrial processes, A corr represents the correlation adjacency matrix, A learn represents the asymmetric adjacency matrix after sparsification.

[0102] In some embodiments of the present application, in a target process, it is necessary to obtain the spatiotemporal feature representation of each training sample. Specifically, the graph adjacency matrix of the training sample and the target process variable data in the training sample can be input into the spatiotemporal dependency perception feature extraction model for processing to obtain the spatiotemporal feature representation of the training sample.

[0103] like Figure 2 As shown, the above-mentioned spatiotemporal dependency perception feature extraction model includes a skip connection module (such as Figure 2 ) and L sequentially connected spatiotemporal dependency perception blocks (such as Figure 2 The output ends of the L spatiotemporal dependency perception blocks are connected to the input end of the jump connection module, and the output end of the jump connection module is the output end of the spatiotemporal dependency perception feature extraction model, which is used to output the spatiotemporal feature representation. It can be understood that when the data input to the spatiotemporal dependency perception feature extraction model is the data corresponding to the entire process, the jump connection module outputs the global spatiotemporal feature representation of industrial production, and when the data input to the spatiotemporal dependency perception feature extraction model is the data corresponding to the sub-process, the jump connection module outputs the local spatiotemporal feature representation of industrial production.

[0104] The above-mentioned spatiotemporal dependency perception block includes a spatial topological feature extraction module (such as Figure 2 The graph convolution spatial feature extraction module in Figure 2The output of the spatial topology feature extraction module and the output of the temporal dynamic feature extraction module are both connected to the input of the spatial-temporal feature fusion module, the output of the spatial-temporal feature fusion module is the output of the spatial-temporal dependency perception block, the input of the spatial topology feature extraction module and the input of the temporal dynamic feature extraction module are both the input of the spatial-temporal dependency perception block, the input of the spatial topology feature extraction module of the first spatial-temporal dependency perception block among the L spatial-temporal dependency perception blocks is the graph adjacency matrix and target process variable data of the target process corresponding to the training samples, and the input of the temporal dynamic feature extraction module of the first spatial-temporal dependency perception block is the target process variable data of the target process corresponding to the training samples; for other spatial-temporal dependency perception blocks except the first spatial-temporal dependency perception block, their input is the output of the previous spatial-temporal dependency perception block.

[0105] The spatial topological feature extraction module is used to extract the spatial topological features of its input data. Specifically, the spatial topological feature extraction module includes sequentially connected spatial position attention submodules (such as Figure 2 The spatial position attention in Figure 2 The input end of the spatial position attention submodule is the input end of the spatial topological feature extraction module, and the output end of the hybrid skip graph convolution submodule is the output end of the spatial topological feature extraction module.

[0106] The temporal dynamic feature extraction module is used to extract the temporal dynamic features of its input data. Specifically, the temporal dynamic feature extraction module includes sequentially connected temporal position attention submodules (such as Figure 2 The temporal position attention in Figure 2 The input end of the temporal position attention submodule is the input end of the temporal dynamic feature extraction module, and the output end of the multi-scale dilated causal convolution submodule is the output end of the temporal dynamic feature extraction module.

[0107] The spatiotemporal feature fusion module is used to deeply fuse the extracted spatial topological features and temporal dynamic features through the attention mechanism, and to mine the rich interactive spatiotemporal patterns in the input data. Specifically, the spatiotemporal feature fusion module is used to fuse the spatial topological features output by the spatial topological feature extraction module and the temporal dynamic features output by the temporal dynamic feature extraction module through the feature fusion formula to obtain the spatiotemporal feature fusion result, thereby achieving a more comprehensive and accurate description of the spatiotemporal dependency relationship of the input sequence.

[0108] The feature fusion formula is: Z l represents the spatiotemporal feature fusion result output by the spatiotemporal feature fusion module of the lth spatiotemporal dependency perception block, sl represents the spatial attention vector of the lth spatiotemporal dependency perception block, represents the spatial topological features output by the spatial topological feature extraction module of the lth spatiotemporal dependency perception block, t l represents the temporal attention vector of the lth spatiotemporal dependency perception block, represents the temporal dynamic features output by the temporal dynamic feature extraction module of the lth spatiotemporal dependency perception block, l = 1,…,L, L is the number of spatiotemporal dependency perception blocks.

[0109]

[0110] Among them, α1, α2, β1, and β2 are all 1×1 convolutional layers, and σ represents the sigmoid activation function.

[0111] The skip connection module is used to add the spatiotemporal feature fusion results output by each spatiotemporal dependency perception block to obtain a spatiotemporal feature representation. As an optional example, the skip connection module may adopt a commonly used skip connection module.

[0112] The following is an exemplary description of the data processing process of the spatial topological feature extraction module and the temporal dynamic feature extraction module.

[0113] Among them, the spatial topological feature extraction module mainly uses the spatial position attention mechanism to locate the key process variable nodes, and uses the hybrid skip graph convolution layer to fuse the feature information of each process variable and its neighboring process variables based on the constructed graph adjacency matrix to obtain the spatial topological features between the process variables. As an optional example, the spatial position attention submodule in the spatial topological feature extraction module can adopt the commonly used position attention module, and the hybrid skip graph convolution submodule can be a commonly used hybrid skip graph convolution network. Of course, it can also be obtained by executing the following steps to obtain spatial topological features.

[0114] Specifically, the spatial position attention submodule in the spatial topology feature extraction module is used to perform the following steps:

[0115] Step 12.5, use the one-dimensional pooling kernel (1, W) in the time direction to aggregate the features of each dimension along the direction of the variable (i.e., process variable), and obtain the feature vector of variable direction perception by encoding the spatial dependency between variables. The output of the hth variable of the cth channel is It can be expressed as follows:

[0116]

[0117] Where W represents the dimension of the input data feature Z (i.e., the target process variable data) of the spatial position attention submodule in the time direction, z c(h,i) represents the input data feature of the hth variable of the cth channel, and i represents the index of the data feature in the time direction.

[0118] Step 12.6, aggregate all eigenvectors according to the generated variable direction Get y S , passing it through two 1×1 convolution transformations in succession to encode the spatial dependencies between the input data variables and more accurately locate the feature representation of the key variables. The process can be expressed as follows:

[0119] f S =δ(F S1 (y S ))

[0120] g S =δ(F S2 (f S ))

[0121] Among them, F S1 and F S2 Represent two 1×1 convolution transformations, δ is a nonlinear activation function, y S Aggregate eigenvectors for the generated variable directions, f S The intermediate feature map that encodes spatial information for the variable direction, g S is the variable direction attention weight obtained by encoding.

[0122] Step 12.7: Encode the variable direction attention weight g S Combined with the original input data feature Z, we get the feature representation X that embeds the spatial dependency between variables. S , to achieve a more accurate representation of the key variable characteristics. The process is expressed as:

[0123] X S =g S ·Z

[0124] Among them, X S It is the input feature of the subsequent hybrid skip graph convolution submodule.

[0125] The hybrid skip graph convolution submodule in the spatial topology feature extraction module is used to perform the following steps:

[0126] Step 12.8, using the hybrid skip graph convolution layer, based on the constructed graph adjacency matrix, processes the information flow of each node (i.e., process variable) in the graph. By propagating information horizontally, the original state of some nodes is retained during the propagation process, and while retaining the local features of the original nodes, its deep neighborhood is explored. It can be specifically expressed by the following formula:

[0127]

[0128] Among them, H (k) represents the hidden state of the node after the kth propagation, H in represents the input hidden features of the previous layer, β is a hyperparameter that controls the proportion of original node features retained, is the normalized graph adjacency matrix, H (k-1) represents the hidden state of the node before the kth propagation, k represents the current number of propagation layers, is the diagonal matrix of the graph adjacency matrix A.

[0129] Step 12.9 introduces the information selection step to filter out the redundant information generated by each hop of graph convolution and retain the important information, which can be expressed as:

[0130]

[0131] Among them, H out represents the node feature representation finally selected after multi-step propagation, K is the propagation depth of the graph convolution process, and W (k) Used as a feature selector to select node feature information.

[0132] Step 12.10, the graph convolution module consists of two hybrid skip graph convolution layers, which process the inflow and outflow information of each node respectively, and add the outputs of the two hybrid skip graph convolution layers to get the net inflow information of each node:

[0133] Z S =MixhopGCN(X S ,A)+MixhopGCN(X S ,A T )

[0134] MixhopGCN(·) represents the process of extracting spatial topological features between variables by using mixed jump graph convolutional layers, Z S It is the spatial topological feature finally output by the spatial topological feature extraction module.

[0135] In the specific implementation, we first perceive the spatial information of variables through spatial position attention, and use convolution transformation to encode the spatial dependency between variables to obtain the attention weight matrix in the variable direction to emphasize the spatial feature representation of key variable nodes. Then, based on the constructed graph adjacency matrix, we use the hybrid skip graph convolution layer to propagate node information horizontally and select node information vertically. We use the graph convolution module to process the information flow of each node in the graph, and gradually extract the spatial topological features between variables.

[0136] The temporal dynamic feature extraction module mainly uses the temporal position attention mechanism to locate key time points, and uses multi-scale dilated causal convolution to efficiently capture the sequential patterns and dynamic changes of each variable at different scales, and obtain the temporal dynamic features between sequences. As an optional example, the temporal position attention submodule in the temporal dynamic feature extraction module can be the commonly used position attention, and the multi-scale dilated causal convolution submodule can be the commonly used multi-scale dilated causal convolution kernel. Of course, it can also be obtained by executing the following steps.

[0137] Among them, the time position attention submodule in the time dynamic feature extraction module is mainly used to perform the following steps:

[0138] Step 12.11, use the one-dimensional pooling kernel (H, 1) in the spatial direction to aggregate the features of each dimension along the time direction, and encode the long-range dependencies of different time steps to obtain the feature vector perceived in the time direction. The output of the c-th channel at the w-th time step is It can be expressed as follows:

[0139]

[0140] Among them, H represents the dimension of the input data feature Z (i.e., the target process variable data) of the time-position attention submodule in the spatial direction, z c (j,w) represents the input data features of the cth channel at the wth time step, and j represents the index of the data features in the spatial direction.

[0141] Step 12.12, aggregate all feature vectors according to the generated time direction Get y T , passing it through two 1×1 convolution transformations successively to encode the temporal dependency between different time steps of the input data and more accurately locate the feature representation of the key time step. The process can be expressed as follows:

[0142] f T =δ(F T1 (y T ))

[0143] g T =δ(F T2 (f T ))

[0144] Among them, F T1 and F T2 Represent two 1×1 convolution transformations, δ is a nonlinear activation function, y T Aggregate feature vector for the generated time direction, f T is the intermediate feature map that encodes temporal information in the temporal direction, g Tis the encoded temporal attention weight.

[0145] Step 12.13: Encode the temporal attention weight g T Combined with the original input feature Z, we get the feature representation X that embeds the temporal dependencies between different time steps T , to achieve a more accurate representation of the key time step features. The process is expressed as:

[0146] X T =g T ·Z

[0147] Among them, X T It is the input feature of the subsequent multi-scale expanded causal convolution submodule.

[0148] The multi-scale dilated causal convolution submodule in the temporal dynamic feature extraction module is mainly used to perform the following steps:

[0149] In step 12.14, a multi-scale dilated causal convolution is designed to capture the local temporal pattern of each variable (i.e., process variable) in the industrial time series data. The dilated convolutions at different layers are used to expand the receptive field and increase the distance of information propagation. The multi-scale information extraction makes the model more sensitive to the local dependencies in the sequence data. The specific expression of the multi-scale dilated causal convolution process is:

[0150]

[0151] in, represents a two-dimensional convolution filter, * d represents a convolution operation with a dilation factor, d is the dilation factor, k i Represents the size of the i-th convolution filter.

[0152] As the convolutional layers are stacked, the receptive field is continuously expanded through the expansion factor to extract long-term information. Assuming that the initial expansion factor is 1 and the convolution kernel size is c, the receptive field R of the m-layer temporal convolutional network is:

[0153]

[0154] Among them, q is the exponential growth rate of the dilation factor of each temporal convolutional layer.

[0155] Step 12.15, two multi-scale dilated causal convolution modules are used to realize deep extraction of temporal dynamic features. One of them is followed by a tanh activation function as a filter to capture key information, and the other is followed by a sigmoid activation function as a gate to control the outflow of information to obtain the output temporal dynamic feature Z T As shown below:

[0156] Z T=tanh(X T *F1)×σ(X T *F2)

[0157] Where σ(·) represents the sigmoid activation function, F1 represents the convolution filter used in conjunction with the tanh activation function, F2 represents the convolution filter used in conjunction with the sigmoid activation function, and Z T It is the temporal dynamic feature finally output by the temporal dynamic feature extraction module.

[0158] In the specific implementation, we first perceive the time information through the time position attention, and use the convolution transformation to encode the dynamic dependency between different time steps of the input data to obtain the attention weight matrix in the time direction, so as to highlight the dynamic feature representation of the key time step. Then, we use the multi-scale expansion causal convolution and gating mechanism to achieve the deep extraction of the dynamic features of the time series data.

[0159] It can be seen that by taking the global information of the entire industrial production process and the local information of each group of equipment unit sub-processes as input, and passing through multiple spatiotemporal dependency perception blocks composed of a graph convolution spatial feature extraction module, a temporal convolution temporal feature extraction module and a spatiotemporal feature fusion module, an effective spatiotemporal feature representation can be obtained.

[0160] That is, for the actual industrial production process, considering that the entire industrial production process is usually composed of sub-processes composed of multiple groups of different equipment units, in order to simultaneously extract the global spatiotemporal features of the entire industrial production process and the local spatiotemporal features of each group of equipment unit sub-processes, the global information of the entire industrial production process and the local information of each group of equipment unit sub-processes are used as the input of the spatiotemporal dependency perception feature extraction model, and processed by multiple stacked spatiotemporal dependency perception blocks, so as to more fully extract the deep spatiotemporal features of the entire industrial production process and each group of sub-processes. Furthermore, the output of each spatiotemporal dependency perception block is connected by skip connection to obtain the final global and local spatiotemporal feature representation of industrial production.

[0161] Step 13, for each training sample respectively, according to the prediction model and the spatiotemporal feature representation of all target processes corresponding to the training sample, obtain the key quality indicator prediction values ​​of all target processes corresponding to the training sample, and use the key quality indicator prediction values ​​and key quality indicator values ​​corresponding to all training samples to optimize the parameters in the adjacency matrix model, the spatiotemporal dependency perception feature extraction model and the prediction model.

[0162] In some embodiments of the present application, for each training sample, the key quality indicator prediction value of each target process corresponding to the training sample (that is, the key quality indicator prediction value of the training sample under each target process) can be obtained in the following manner:

[0163] Step 13.1, for the whole process, the spatiotemporal feature representations of all target processes corresponding to the training samples are fused, and the fused spatiotemporal feature representations are input into the prediction model for processing to obtain the predicted values ​​of the key quality indicators of the whole process corresponding to the training samples.

[0164] That is, for the whole process, when predicting the key quality indicators, the data input to the prediction model is: the fusion result of the spatiotemporal feature representation corresponding to the training sample in all target processes (i.e., the whole process and each sub-process), and the fusion result can be obtained by splicing the spatiotemporal feature representation corresponding to the training sample in all target processes. The fused spatiotemporal feature representation Z all for:

[0165] Z all =concat(Z,Z group1 ,...,Z groupN )

[0166] Among them, Z represents the spatiotemporal feature representation corresponding to the training sample in the whole process, Z group1 Indicates the spatiotemporal feature representation corresponding to the training sample in the first subprocess, Z groupN It represents the spatiotemporal feature representation corresponding to the training sample in the Nth subprocess, and groupN is the number of subprocesses.

[0167] Specifically, the fused spatiotemporal feature representation Z all Input to the 1×1 output convolution layer (the output convolution layer is a way to implement the prediction model) for processing to obtain the predicted values ​​of the key quality indicators corresponding to the training samples in the whole process

[0168]

[0169] Among them, Endconv(·) represents the output convolution operation.

[0170] Step 13.2, for each sub-process, the spatiotemporal feature representation of the sub-process corresponding to the training sample is input into the prediction model for processing to obtain the key quality indicator prediction value of the sub-process corresponding to the training sample.

[0171] That is, for a sub-process, when predicting the key quality indicators, the data input into the prediction model is: the spatiotemporal feature representation corresponding to the training sample in the sub-process. Specifically, the spatiotemporal feature representation corresponding to the training sample in the sub-process is input into the 1×1 output convolution layer for processing, and the key quality indicator prediction value corresponding to the training sample of the sub-process can be obtained.

[0172] In the specific implementation, in order to fully consider the global information of the entire industrial production process and the local information of each group of equipment unit sub-processes, the deep spatiotemporal features extracted from the entire industrial production process and each group of sub-processes are fused by using multiple stacked spatiotemporal dependency perception blocks to obtain a more effective multi-level spatiotemporal feature representation to improve the prediction performance of the model. Then, the multi-level spatiotemporal features obtained are used to predict key quality indicators through the output convolution layer, and the predicted values ​​of key quality indicators are output. In this way, a comprehensive analysis and modeling of complex spatiotemporal dependencies in the industrial production process is achieved.

[0173] After obtaining the predicted values ​​of key quality indicators corresponding to all training samples in all target processes, the predicted loss values ​​can be calculated based on the loss function, and the predicted loss values ​​can be used to optimize the parameters in the adjacency matrix model, spatiotemporal dependency-aware feature extraction model, and prediction model.

[0174] The above loss function is:

[0175]

[0176] Among them, J(Θ) represents the prediction loss value, Θ represents the learnable parameters in the adjacency matrix model, spatiotemporal dependency-aware feature extraction model and prediction model, N represents the number of training samples, groupN is the number of subprocesses, represents the predicted value of the key quality indicator of the whole process corresponding to the i-th training sample, y i represents the key quality indicator value in the i-th training sample, It represents the predicted value of the key quality indicator of the j-th sub-process corresponding to the i-th training sample.

[0177] During the training process, a suitable optimizer, learning rate, and number of iterations can be selected, and then the weight parameters (i.e., the learnable parameters) of the entire key indicator prediction model (i.e., the adjacency matrix model, the spatiotemporal dependency perception feature extraction model, and the prediction model) are updated in a supervised reverse manner through the obtained loss function value (i.e., the above-mentioned prediction loss value) until the model converges to obtain the trained key indicator prediction model, i.e., the optimized adjacency matrix model, the spatiotemporal dependency perception feature extraction model, and the prediction model.

[0178] It is understandable that in some embodiments of the present application, in order to improve the accuracy of the prediction, before using the trained key indicator prediction model to perform real-time prediction of key quality indicators, the prediction accuracy and generalization performance of the trained key indicator prediction model are first verified using the collected test set samples. After the verification is qualified, the modeling auxiliary variables collected in real time are used to predict the future key quality indicator values, thereby realizing the real-time online prediction of key quality indicators.

[0179] Step 14, using the adjacency matrix model after parameter optimization, the spatiotemporal dependency perception feature extraction model and the prediction model to predict the industrial key quality indicators at the prediction time.

[0180] In some embodiments of the present application, when real-time prediction is performed using the adjacency matrix model, spatiotemporal dependency perception feature extraction model and prediction model after parameter optimization, the current process variable data can be collected and then input into the adjacency matrix model, spatiotemporal dependency perception feature extraction model and prediction model after parameter optimization for processing to obtain the predicted value of the industrial key quality indicators at the prediction time, thereby achieving real-time prediction of the industrial key quality indicators.

[0181] It can be understood that in the real-time prediction process, the current process variable data includes the data of each current process variable (such as the current temperature value, liquid level value, flow value, current value, etc.), and then the graph adjacency matrix and spatiotemporal feature representation of the current process variable data in the whole process and each sub-process are obtained according to the processing process of the aforementioned training samples, and the spatiotemporal feature representation corresponding to the whole process and all sub-processes are fused. Finally, the fused spatiotemporal feature representation is input into the prediction model for processing, and the predicted value of the industrial key quality indicators at the prediction time (that is, the current industrial key quality indicator prediction) can be obtained to realize the real-time prediction of industrial key quality indicators.

[0182] The method will be further described below in conjunction with a specific embodiment. Figure 3 As shown, the following takes the potassium salt flotation process as an example to further illustrate the industrial key quality indicator prediction method based on the above-mentioned multi-level spatiotemporal correlation characteristics and spatiotemporal dependence perception.

[0183] Using the coefficient of determination (R 2 ), root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) to quantitatively evaluate the prediction performance of the obtained key indicator prediction model. The LSTNet model, VW-SAE model and LogTrans model are used to compare with the prediction method of this application to verify the effectiveness of the proposed method. The prediction results are as follows: Figure 4 , 5 , 6, and 7. The comparison results of the prediction performance indicators of various methods are shown in Table 1.

[0184] Model <![CDATA[R 2 ]]> RMSE MAE MAPE LSTNet 0.7426 0.0877 0.0689 0.1908 VW-SAE 0.6722 0.0990 0.0800 0.2298 LogTrans 0.7629 0.0842 0.0643 0.1783 Prediction method of this application 0.7875 0.0797 0.0601 0.1652

[0185] Table 1

[0186] It can be seen that the industrial key quality indicator prediction method based on multi-level spatiotemporal correlation features provided in this application has achieved the best prediction effect in each evaluation indicator compared with traditional deep learning prediction methods such as LSTNet model, VW-SAE model and LogTrans model, which also verifies the effectiveness of the prediction method provided in this application.

[0187] In summary, the industrial key quality indicator prediction method provided in the embodiment of the present application utilizes graph adjacency matrix, spatial topological features and time dynamic feature extraction, as well as multi-level full-process global features and each group of sub-process local feature fusion and other technologies to accurately obtain the underlying topological structural relationship between industrial process variables, and on this basis fully explore and utilize the spatial dependency between industrial process variables and the temporal dependency in time series data, while taking into account the global information of the entire industrial production process and the local information of each group of sub-processes to obtain a more effective multi-level spatiotemporal feature representation, thereby improving the accuracy and reliability of modeling prediction, and thus when using the constructed model to predict industrial key quality indicators, the accuracy and reliability of industrial key quality indicator prediction can be greatly improved.

[0188] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A prediction method for industrial key quality indicators based on multi-level spatiotemporal correlation features, It is characterized in that include: Identify process variables related to key quality indicators of industrial production and obtain multiple training samples; The training samples include process variable data at historical moments and corresponding key quality indicator values; the industrial production process includes multiple sub-processes; The whole process and each of the sub-processes are respectively regarded as a target process, and are respectively executed for each target process: for each training sample, the target process variable corresponding to the target process is screened out from all the determined process variables, and the graph adjacency matrix of the target process corresponding to the training sample is obtained according to the adjacency matrix model and the target process variable data corresponding to the target process variable in the training sample; the spatiotemporal feature representation of the target process corresponding to the training sample is obtained by using the spatiotemporal dependency perception feature extraction model; the spatiotemporal dependency perception feature extraction model includes a jump connection module and L spatiotemporal dependency perception blocks connected in sequence, the output ends of the L spatiotemporal dependency perception blocks are all connected to the input end of the jump connection module, and the spatiotemporal dependency perception block includes a spatial topology feature extraction module Block, time dynamic feature extraction module and space-time feature fusion module, the output end of the spatial topology feature extraction module and the output end of the time dynamic feature extraction module are both connected to the input end of the space-time feature fusion module, the output end of the space-time feature fusion module is the output end of the space-time dependency perception block, the input end of the spatial topology feature extraction module and the input end of the time dynamic feature extraction module are both the input end of the space-time dependency perception block, the input of the spatial topology feature extraction module of the first space-time dependency perception block among the L space-time dependency perception blocks is the graph adjacency matrix and target process variable data of the target process corresponding to the training sample, and the input of the time dynamic feature extraction module of the first space-time dependency perception block is the target process variable data of the target process corresponding to the training sample; For each training sample, according to the prediction model and the spatiotemporal feature representation of all target processes corresponding to the training sample, obtain the key quality indicator prediction values ​​of all target processes corresponding to the training sample, and optimize the parameters in the adjacency matrix model, the spatiotemporal dependency perception feature extraction model and the prediction model by using the key quality indicator prediction values ​​and key quality indicator values ​​corresponding to all training samples; The adjacency matrix model with optimized parameters, the spatiotemporal dependency-aware feature extraction model and the prediction model are used to predict the industrial key quality indicators at the prediction moment.

2. The prediction method according to claim 1, It is characterized in that The adjacency matrix model includes an adjacency weight formula and a matrix fusion formula; the step of obtaining a graph adjacency matrix of the target process corresponding to the training sample according to the adjacency matrix model and the target process variable data corresponding to the target process variable in the training sample includes: For each target process variable among all the target process variables, respectively: generating a node embedding based on the target process variable data corresponding to the target process variable; for each other target process variable among all the target process variables except the target process variable, respectively generating a node embedding based on the target process variable data corresponding to the other target process variables, and calculating the adjacency weights between the target process variable and other target process variables according to the generated node embedding and an adjacency weight formula; An asymmetric adjacency matrix is ​​constructed based on all the adjacency weights, and the asymmetric adjacency matrix is ​​sparsely processed to obtain a sparsely processed asymmetric adjacency matrix; each element in the asymmetric adjacency matrix is ​​an adjacency weight of two target process variables among all the target process variables, and each element corresponds to a different adjacency weight; Filtering target process variable data corresponding to the target process variable from the multiple training samples, and calculating the correlation between every two target process variables according to the target process variable data corresponding to every two target process variables among all the target process variables, and constructing a correlation adjacency matrix based on all the calculated correlations; each element in the correlation adjacency matrix is ​​the correlation between two target process variables among all the target process variables, and each element corresponds to a different correlation; The sparse asymmetric adjacency matrix and the correlation adjacency matrix are fused through a matrix fusion formula to obtain a graph adjacency matrix of the target process corresponding to the training sample.

3. The prediction method according to claim 2, It is characterized in that The adjacency weight formula is: M 1i =tanh(αE 1i C 1 ) M 2j =tanh(αE 2j C 2 ) in, E 1i represents the node embedding corresponding to the i-th target process variable, E 2j represents the node embedding corresponding to the jth target process variable, M represents the number of target process variables, d represents the node embedding dimension, Γ 1 and Γ 2 are all learnable parameters, α is a hyperparameter that controls the saturation rate of the activation function, tanh(·) and ReLU(·) both represent activation functions, and A learnij Represents the adjacency weight between the i-th target process variable and the j-th target process variable, i≠j.

4. The prediction method according to claim 2, It is characterized in that The matrix fusion formula is: A=α′A corr +(1-a)A learn Where A represents the graph adjacency matrix, α′ represents the learnable trade-off parameter, and A corr represents the correlation adjacency matrix, A learn represents the asymmetric adjacency matrix after sparsification.

5. The prediction method according to claim 1, It is characterized in that The spatial topological feature extraction module includes a spatial position attention submodule and a hybrid skip graph convolution submodule connected in sequence, the input end of the spatial position attention submodule is the input end of the spatial topological feature extraction module, and the output end of the hybrid skip graph convolution submodule is the output end of the spatial topological feature extraction module.

6. The prediction method according to claim 1, It is characterized in that The temporal dynamic feature extraction module includes a temporal position attention submodule and a multi-scale dilated causal convolution submodule connected in sequence, the input end of the temporal position attention submodule is the input end of the temporal dynamic feature extraction module, and the output end of the multi-scale dilated causal convolution submodule is the output end of the temporal dynamic feature extraction module.

7. The prediction method according to claim 1, It is characterized in that The spatiotemporal feature fusion module is used to fuse the spatial topological features output by the spatial topological feature extraction module and the temporal dynamic features output by the temporal dynamic feature extraction module through a feature fusion formula to obtain a spatiotemporal feature fusion result; The feature fusion formula is: Z l represents the spatiotemporal feature fusion result output by the spatiotemporal feature fusion module of the lth spatiotemporal dependency perception block, s l represents the spatial attention vector of the lth spatiotemporal dependency perception block, represents the spatial topological features output by the spatial topological feature extraction module of the lth spatiotemporal dependency perception block, t l represents the temporal attention vector of the lth spatiotemporal dependency perception block, Represents the temporal dynamic features output by the temporal dynamic feature extraction module of the lth spatiotemporal dependency perception block.

8. The prediction method according to claim 7, It is characterized in that Among them, α 1 , α 2 , β 1 , β 2 All are 1×1 convolutional layers, and σ represents the sigmoid activation function.

9. The prediction method according to claim 1, It is characterized in that According to the prediction model and the spatiotemporal feature representation of all target processes corresponding to the training samples, the key quality indicator prediction values ​​of all target processes corresponding to the training samples are obtained, including: For the whole process, the spatiotemporal feature representations of all target processes corresponding to the training samples are fused, and the fused spatiotemporal feature representations are input into the prediction model for processing to obtain the predicted values ​​of the key quality indicators of the whole process corresponding to the training samples; For each sub-process respectively, the spatiotemporal feature representation of the sub-process corresponding to the training sample is input into the prediction model for processing to obtain the key quality indicator prediction value of the sub-process corresponding to the training sample.

10. The prediction method according to claim 1, It is characterized in that The method uses the key quality indicator prediction values ​​and key quality indicator values ​​corresponding to all training samples to optimize the parameters in the adjacency matrix model, the spatiotemporal dependency perception feature extraction model and the prediction model, including: Calculating a prediction loss value based on a loss function, and optimizing parameters in an adjacency matrix model, a spatiotemporal dependency perception feature extraction model, and a prediction model using the prediction loss value; The loss function is: Among them, J(Θ) represents the prediction loss value, Θ represents the learnable parameters in the adjacency matrix model, spatiotemporal dependency-aware feature extraction model and prediction model, N represents the number of training samples, groupN is the number of subprocesses, represents the predicted value of the key quality indicator of the whole process corresponding to the i-th training sample, y i represents the key quality indicator value in the i-th training sample, It represents the predicted value of the key quality indicator of the j-th sub-process corresponding to the i-th training sample.

Images