A data-driven industrial process operating condition prediction method

Abstract

The application discloses a data-driven industrial process operation condition prediction method, and belongs to the technical field of condition prediction. The method comprises the following steps: regarding each production batch as a node, obtaining the embedding features of the node, and performing node feature embedding; wherein the production batch comprises a historical batch with known conditions and a future batch with unknown conditions; the node corresponding to the historical batch is a known label node, and the node corresponding to the future batch is an unknown label node; a dynamic graph comprising a plurality of known label nodes and an unknown label node is constructed; the propagation and update of the node features are performed in the dynamic graph, and the missing features of the unknown label node are obtained; and the missing features and the embedding features of the unknown label node are combined to realize industrial process operation condition prediction. The data-driven industrial process operation condition prediction method provided by the application solves the feature extraction and condition prediction problems under the condition of multi-source heterogeneous data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of working condition prediction, in particular to a data-driven industrial process running working condition prediction method. BACKGROUND

[0002] Industrial processes (such as chemical, power, steel, pharmaceutical, etc.) usually involve complex physical and chemical reactions, multivariable coupling control and dynamic environmental changes. The running working condition determines the production efficiency, product quality and equipment life. Running working condition prediction through data-driven modeling, early prediction of future state, becomes the core demand of industrial intelligence. For industrial processes, the evaluation of running working condition is beneficial to the real-time monitoring and grasp of the current production process. Accurate prediction of running working condition is beneficial to guide production operation adjustment and process performance improvement, and is more helpful to improve the utilization rate of resources, reduce the running cost of complex industrial manufacturing process, and improve the economic benefit of enterprises. When performing running working condition prediction, the running working condition to be predicted has not actually occurred, only production plan related information is available, so running working condition prediction is a more difficult problem.

[0003] The prediction of running working condition is mainly to predict the key variables or indicators reflecting the running working condition in the future time to evaluate the future running state. Many data-driven methods are applied to running working condition prediction, such as Gaussian process regression, recursive fuzzy neural network, long short-term memory network, etc. Some researches combine different methods to achieve better prediction effect. For example, an ultra-short-term working condition prediction method based on convolutional neural network (CNN) and long short-term memory network (LSTM), a running state prediction method of convolutional neural network-bidirectional gated recurrent unit (CNN-BiGRU) fusing attention mechanism. Some researches combine data-driven methods and expert knowledge methods to realize prediction, an uncertainty life planning probability prediction method combining a probability prediction model with a Bayesian filtering method based on expert knowledge.

[0004] The prediction of running conditions depends on the implementation of the prediction algorithm. For the study of prediction methods, many scholars have proposed effective solutions. For battery state of health prediction, an encoder-decoder framework based on attention mechanism is used for prediction, a battery state of health prediction framework based on GraphSAGE. For the field of traffic, Multi-task Hypergraph Convolutional Neural Network (MT-HGCN) can be used for traffic flow prediction, three prediction models based on Transformer, Generative Adversarial Network (GAN) and Conditional Variational Auto-encoder (CVAE) are used to predict future trajectories. For hydrological forecasting, a Spatiotemporal Hetero Graph-based LSTM (SHG-LSTM) model is used for multi-step flood forecasting in advance. Dynamic Spatio-temporal Directed Graph Attention Network (DSTDGAT) is used for short-term wind speed prediction.

[0005] As can be seen from the above research, graph neural networks are a powerful method for advance prediction. Graph neural networks generally have vertex-level, edge-level and graph-level prediction. Node classification and graph classification are two graph learning problems that predict the class labels of nodes and graphs, respectively.

[0006] For node classification, GCN with Multi-similarity Attribute Matrices Fusion (MSF-GCN) can be used to implement node classification, Discriminable Squeeze and Excitation GCN (D-SEGCN) based on feature attention mechanism can be used for semi-supervised node classification, and Hierarchical GAT (HGAT) can be used for semi-supervised node classification. However, due to the presence of noise in many graph structures or the absence of graph structure in the data itself, the graph structure needs to be learned. A unified structure learning framework for graph attention networks is needed, which can be used for semi-supervised node classification. An end-to-end graph learning framework, Iterative Deep Graph Learning (IDGL), can be used to jointly and iteratively learn graph structure and graph embedding.

[0007] In summary, the prediction of operating conditions can be achieved using node classification based on graph neural networks. However, in process industries, how to construct graph structures and node embeddings based on data, and how to update the features of unknown label nodes based on the features of known label nodes for classification prediction need to be further explored. Currently, how to achieve feature extraction and operating condition prediction under multi-source heterogeneous data is still a difficult problem to be solved. SUMMARY

[0008] The present application provides a data-driven industrial process operating condition prediction method to solve the technical problem of feature extraction and operating condition prediction under multi-source heterogeneous data.

[0009] To solve the above technical problems, the present application provides the following technical solutions:

[0010] On the one hand, the present application provides a data-driven industrial process operating condition prediction method, comprising:

[0011] Each production batch is regarded as a node, and the embedding features of the node are obtained for node feature embedding; wherein the production batch includes historical batches with known operating conditions and future batches with unknown operating conditions, the node corresponding to the historical batch is a known label node, and the node corresponding to the future batch is an unknown label node;

[0012] A dynamic graph containing a plurality of known label nodes and an unknown label node is constructed;

[0013] The propagation and update of node features in the dynamic graph are performed to obtain the missing features of the unknown label node;

[0014] The missing features and embedding features of the unknown label node are combined to realize the prediction of the operating conditions of the industrial process.

[0015] Further, for the unknown label node, the process of obtaining the embedding features of the node includes:

[0016] The product model and process setting value corresponding to the unknown label node are collected;

[0017] The vector representation of the product model is obtained, then the vector representation of the product model is spliced with the process setting value, and the spliced result is input into a fully connected layer for feature embedding to obtain the embedding features of the unknown label node.

[0018] Further, for the known label node, the process of obtaining the embedding features of the node includes:

[0019] The product model, process setting value, preset index variable and operating condition label corresponding to the known label node are collected;

[0020] Feature extraction is performed on the preset index variable to obtain feature data of the preset index variable;

[0021] The vector representation of the product model is obtained, the vector representation of the product model is spliced with the process setting value, and the splicing result is input into a fully connected layer for feature embedding to obtain production plan information embedding features;

[0022] The production plan information embedding features corresponding to the known label node, the feature data of the preset index variable, and the working condition label are spliced to obtain embedding features corresponding to the known label node.

[0023] Further, the feature data of the preset index variable is extracted by a conditional variational autoencoder.

[0024] Further, in the dynamic graph, the known label node has a directed edge, and the determination method is:

[0025] Calculate the Euclidean distance between the embedding features of the current known label node and each known label node, and select the known label node corresponding to the Euclidean distance less than the preset threshold as the candidate node;

[0026] Determine whether the index of each candidate node is greater than the index of the current known label node; if the index of the candidate node is greater than the index of the current known label node, there is a directed edge from the current known label node to the corresponding candidate node, otherwise, there is no directed edge from the current known label node to the corresponding candidate node.

[0027] Further, in the dynamic graph, the unknown label node has a directed edge, and the determination method is:

[0028] For the unknown label node, the nodes corresponding to the first two batches of adjacent production point to it;

[0029] Calculate the Euclidean distance between the embedding features of the unknown label node and the production plan information embedding features of each known label node, and point the known label node corresponding to the minimum Euclidean distance to the unknown label node.

[0030] Further, the node in the dynamic graph has K out-degree neighbor nodes; K is a parameter selected and set by itself.

[0031] Further, a dynamic graph containing a plurality of known label nodes and an unknown label node is constructed, comprising:

[0032] For the current time slice, initialize an adjacency matrix A∈R (n+1)×(n+1) ; Wherein, n is the number of known label nodes in the current graph, and the node with the latest time occurrence is the unknown label node;

[0033] calculating the Euclidean distance between the embedding features of the first n nodes, sorting the Euclidean distance between the embedding features of the current node and other nodes from small to large, starting from the node with the smallest distance, judging whether the time sequence of the node is later than that of the current node; when the time sequence of the node is later than that of the current node, the element at the corresponding position in the adjacency matrix is assigned a value of 1; judging whether the out-degree neighbor node number of the current node reaches the set K value, if yes, performing the same operation on the next node, if not, continuing to determine the out-degree neighbor of the current node until the out-degree neighbor node number reaches the set K value; retaining the self-loop of each node, i.e. the diagonal element of the adjacency matrix is 1;

[0034] calculating the Euclidean distance between the embedding features of the unknown label node and the production plan information embedding features of the known label node; selecting the node corresponding to the minimum Euclidean distance and the first two nodes adjacent to the unknown label node in time sequence as the in-degree neighbor nodes of the unknown label node, and recording the elements at the corresponding positions of the adjacency matrix A as 1.

[0035] Further, the propagation and update of node features in the dynamic graph are performed to obtain the missing features of the unknown label node, comprising:

[0036] inputting the dynamic graph into the LGASNN module, and utilizing the LGASNN module to perform the propagation and update of node features in the dynamic graph to obtain the missing features of the unknown label node;

[0037] wherein, the LGASNN module comprises a GAT branch and a GraphSAGE branch; the process of utilizing the LGASNN module to perform the propagation and update of node features in the dynamic graph comprises:

[0038] in the GAT branch, the embedding features of the nodes are respectively input into LSTM and GAT, the dynamic change features captured by LSTM are aggregated by GAT; the dynamic change features captured by LSTM are superimposed with the aggregation results of GAT after passing through a linear mapping layer, and the superimposed results are input into a layer normalization layer; after the output results of the layer normalization layer pass through a Relu mechanism, the first features are obtained; the first features are respectively input into LSTM and GAT, the dynamic change features captured by LSTM are aggregated by GAT; the dynamic change features captured by LSTM are superimposed with the aggregation results of GAT after passing through a linear mapping layer, and the superimposed results are input into a layer normalization layer; after the output results of the layer normalization layer pass through a Relu mechanism, the updated node features are obtained, denoted as GAT branch features;

[0039] In the GraphSAGE branch, the embedding features of the nodes are respectively input into LSTM and GraphSAGE, the dynamic change features are captured through LSTM, and the mean aggregation is performed through GraphSAGE; the dynamic change features captured by LSTM are superimposed with the results of the mean aggregation performed through GraphSAGE after passing through a linear mapping layer, the superimposed results are input into a layer normalization layer, and the output results of the layer normalization layer are obtained after passing through a Relu mechanism, and the second features are obtained; the second features are respectively input into LSTM and GraphSAGE, the dynamic change features are captured through LSTM, and the mean aggregation is performed through GraphSAGE; the dynamic change features captured by LSTM are superimposed with the results of the mean aggregation performed through GraphSAGE after passing through a linear mapping layer, the superimposed results are input into a layer normalization layer, and the output results of the layer normalization layer are obtained after passing through a Relu mechanism, and the updated node features are obtained, which are denoted as GraphSAGE branch features.

[0040] The GAT branch features and the GraphSAGE branch features jointly constitute the missing features of the unknown label nodes.

[0041] Further, the missing features and the embedding features of the unknown label nodes are combined to realize the industrial process running condition prediction, including:

[0042] The missing features and the embedding features of the unknown label nodes are spliced, and the spliced features are input into a full connection layer, and the prediction results of the industrial process running condition are obtained through the full connection layer and a Softmax layer.

[0043] In another aspect, the present application also provides an electronic device, which comprises a processor and a memory; wherein the memory stores at least one instruction, which is loaded and executed by the processor to realize the above method.

[0044] In another aspect, the present application also provides a computer readable storage medium, which stores at least one instruction, which is loaded and executed by the processor to realize the above method.

[0045] The technical scheme provided by the present application has at least the following beneficial effects:

[0046] The application firstly carries out feature embedding on the labeled nodes and unlabeled nodes, adopts a feature extraction method based on CVAE for key variables, adopts a character embedding and position embedding method for string input, and carries out feature embedding on the vector converted from the string and the production plan setting value related variable. Secondly, based on the constructed dynamic directed graph and the constructed batch label, the missing features of the unknown label nodes are obtained through the proposed LGASNN model for feature propagation and updating. Finally, the embedding features of the unknown label nodes and the features obtained through propagation and updating are fused to obtain the prediction level of the operating condition. Thus, the feature extraction and operating condition prediction problems under the condition of multi-source heterogeneous data are solved. The actual hot strip rolling process production data is used for application verification. The experimental results verify the effectiveness of the operating condition prediction method of the application. Compared with the existing static GCN and label propagation algorithm based on dynamic graph, the accuracy of operating condition prediction is increased by an average of 7%. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0048] Figure 1 is a dynamic LGASNN model framework provided by the embodiment of the application;

[0049] Figure 2 is a flowchart of a data-driven industrial process operating condition prediction method provided by the embodiment of the application;

[0050] Figure 3 is a schematic diagram of a node feature embedding module of a dynamic graph provided by the embodiment of the application;

[0051] Figure 4 is a processing flowchart of a steel grade provided by the embodiment of the application;

[0052] Figure 5 is a schematic diagram of a directed edge determination of a graph provided by the embodiment of the application;

[0053] Figure 6 is a schematic diagram of a dynamic graph structure provided by the embodiment of the application;

[0054] Figure 7 is a structural block diagram of an LGASNN module provided by the embodiment of the application;

[0055] Figure 8 is a multi-head attention schematic diagram;

[0056] Figure 9 is a structure of an LSTM unit;

[0057] Figure 10 is a schematic diagram of a training process of a dynamic LGASNN model provided by an embodiment of the present application;

[0058] Figure 11 is a performance comparison schematic diagram when determining parameters of a LGASNN model provided by an embodiment of the present application; wherein (a) is a model prediction performance comparison of different η g ; (b) is a model prediction performance comparison of different N bg ; (c) is a model prediction performance comparison of different η c ; (d) is a model prediction performance comparison of different N bc ; (e) is a model prediction performance comparison of different d v ; (f) is a model prediction performance comparison of different d H ; (g) is a model prediction performance comparison of different d M ; (h) is a model prediction performance comparison of different d o ;

[0059] Figure 12 is a loss function curve of a dynamic LGASNN model training process provided by an embodiment of the present application;

[0060] Figure 13 is a running condition prediction result of a dynamic LGASNN model provided by an embodiment of the present application;

[0061] Figure 14 is a running condition prediction result of a comparative experiment provided by an embodiment of the present application; wherein (a) is a running condition prediction result of a GCN model based on a static full connection graph; (b) is a running condition prediction result of a label propagation algorithm based on a dynamic graph;

[0062] Figure 15 is a system block diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical scheme and advantages of the present application clearer, the following will further describe the embodiments of the present application in combination with the drawings.

[0064] First, it should be pointed out that in the embodiments of the present application, the words "exemplary", "for example", and the like are used to indicate examples, instances, or illustrations. Any embodiment or design presented as "exemplary" in the present application should not be construed as preferable or advantageous over other embodiments or design schemes. Rather, the word "exemplary" is used to present concepts in a concrete manner. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or can be either one of the two.

[0065] First embodiment

[0066] To solve the problem of feature extraction and working condition prediction under multi-source heterogeneous data, the present embodiment provides a data-driven industrial process running working condition prediction method (dynamic LGASNN model). First, for the historical batch whose running working condition level is known, i.e. the node whose label is known and the future batch whose running working condition level is unknown, i.e. the node without working condition label, feature embedding is performed. The feature extraction method based on conditional variational autoencoder CVAE is used for key variables, and the extracted features and labels are used as part of the embedding features of the known label node. The steel grade and set value of the known label node and the unknown label node are known, the character embedding and position embedding method is used for the string form input, the vector representation obtained by converting the string input is obtained, and the vector obtained by converting the string and the production plan set value variable are spliced, the full connection layer is used for feature embedding, the embedding features of the label unknown node are obtained, and the obtained features and the key variables are used as the embedding features of the known label node. Secondly, based on the extracted node features and mechanism knowledge, a dynamic graph facing the running working condition is constructed, and the batch working condition label is constructed according to the multiple quality indexes. Then, the dynamic LGASNN model is proposed, the propagation and update of node features are carried out through the proposed dynamic graph neural network model based on graph attention, GraphSAGE and long short-term memory network, the combination of GAT, GraphSAGE and LSTM is realized in the LGASNN module, the aggregation of node features and the capture of dynamic change characteristics of nodes are realized, and the missing features of the unknown label node are obtained by propagation and update. Finally, the embedding features of the unknown label node and the features obtained by propagation and update are fused, and the running working condition prediction result of a batch to be produced in the future is obtained based on this.

[0067] The method is applied to verify the actual production data of the hot strip rolling process, compared with the static graph convolutional network (Graph Convolutional Networks, GCN) and the label propagation algorithm based on dynamic graph, the accuracy of working condition prediction is improved by an average of 7%.

[0068] The method can be realized by an electronic device, and is used in a hot strip rolling process. The running conditions of future batches can be predicted according to the information of the produced batches and the production plan related information of unproduced batches, so as to facilitate timely intervention and prevention of faults, and guide production operation adjustment and process performance improvement.

[0069] To implement the method, the following settings are first made: the steel grade in the form of a string in the production plan information is denoted as s, and the set value related information such as the thickness set value, the width set value, etc. is denoted as The key variables such as the thickness deviation and the width deviation are denoted as The label of the running condition is denoted as Y. Among them, the information of n historical batches and the production plan related information of a batch to be predicted running condition are known information, and the running condition of a batch to be produced in the future is the prediction target. That is, the information of the historical batches is known, and only part of the information of a batch to be produced in the future is known. The method of predicting the running condition evaluation level of the next production batch through the known information will be introduced later.

[0070] As shown in Figure 1 , the method mainly includes four parts: node feature embedding, dynamic graph and batch label construction, LGASNN module for node feature propagation and update, and feature fusion node label prediction.

[0071] Specifically, the execution flow of the method is as shown in Figure 2 , including the following steps:

[0072] S1, each production batch is regarded as a node, the embedded features of the node are obtained, and node feature embedding is performed; wherein the production batch includes historical batches with known running conditions and future batches with unknown running conditions, the nodes corresponding to the historical batches are known label nodes, and the nodes corresponding to the future batches are unknown label nodes;

[0073] The node feature embedding process is as shown in Figure 3 .

[0074] This embodiment converts the prediction of the running condition of a batch to be produced in the future into the prediction of the unknown label node in the dynamic graph, that is, the node classification problem. To perform node prediction, node feature embedding is needed to obtain the initial features of the node, so as to perform subsequent node feature propagation and update.

[0075] The embedded features of the unlabeled node of the running condition to be predicted are obtained by feature embedding of the set information and the steel grade. For the i-th slab, that is, the i-th node, the steel grade is s i , and the set value variable of the production plan is The record form of the steel grade is a string, which contains letters, numbers, special symbols, etc., so character embedding is performed to convert it into a numerical vector, which facilitates subsequent feature processing. The processing process of the steel grade is shown in Figure 4 .

[0076] The advantage of using character embedding is that it can handle new steel types that appear in the test set. The characters in the string are converted to the index value of the pre-set character sequence, and then the embedding layer f idx is used to convert the string into a numerical index embedded as a vector. e

[0077] C i = f e (f idx (s i )) (1)

[0078] In the formula, the matrix where the length of the string s i is 12, and d s is the dimension of the character embedding vector of the string. Thereafter, the embedding matrix C i is position embedded, and the position embedded matrix is denoted as O i . The position embedded matrix is obtained by training.

[0079]

[0080] In the formula, is the string embedding matrix after adding position embedding. After adding position embedding, the same characters in different positions can be distinguished. In order to simplify the calculation, the mean value of each column of is obtained to obtain the vector

[0081] Thereafter, the set value of the slab and the vector obtained by string embedding are spliced, and the slab setting information and the embedding feature V i p of the steel grade are obtained through two fully connected layers, i.e.:

[0082]

[0083] In the formula, [;] represents the splicing of two matrices.

[0084] For historical batches, not only the production plan related information is known, but also the information of some key variables in the production process is known. Therefore, the embedding feature of the labeled node contains not only the embedding feature V i p ​The embedding feature of the key variable is also included. The label of the historical slab is known, so the CVAE module is used to extract the feature of the key variable to obtain the feature of the key variable related to the evaluation grade. For the i-th slab, the label is The label is in the form of one-hot encoding. The grade of the running condition prediction in the embodiment is divided into four categories, i.e. excellent, good, medium and poor. The key variable of the node and the real label are spliced together as the input of the CVAE. First, the multi-layer fully connected layer is used to obtain:

[0085]

[0086] In the formula, f e represents the mapping function of the encoder multi-layer fully connected layer, The average matrix μ and the covariance matrix Σ are obtained after the fully connected layer and satisfy the following conditions:

[0087] lnΣ i =f Σ (H i ) (5)

[0088] μ i =f μ (H i ) (6)

[0089] In the formula, f Σ and f μ represent the fully connected layer. The reparameterization trick is used to obtain the latent variable z, which satisfies:

[0090] z i =μ i +exp(Σ i / 2)*∈ i (7)

[0091] In the formula, ∈ i ~N(0,I), After that, the latent variable and the real label y i The key variable H is reconstructed by the decoder, which satisfies:

[0092]

[0093] In the formula, f d represents the mapping function of the multi-layer fully connected layer of the decoder. Therefore, the embedding feature of the known label node is composed of the embedding feature V i of the production plan related information, p the embedding feature H i of the key variable, and the evaluation label y i of the slab indicated by the current node.

[0094] S2, constructing a dynamic graph comprising a plurality of known label nodes and one unknown label node;

[0095] Specifically, in the present embodiment, the above S2 comprises:

[0096] S21, batch label construction

[0097] The indicators considered for each batch are the quality indicators of the finish rolling outlet, including thickness, width, and crown. The mean value of the thickness deviation, the mean value of the width deviation, and the mean value of the crown of each batch are calculated. The values obtained by the above calculation are segmented and scored, and the resulting scores are weighted and fused using the CRITIC method to obtain the final comprehensive score Score.

[0098] Suppose that according to different indicators, a score matrix is obtained where N is the number of samples, and num is the number of indicators considered.

[0099] On the basis of the score matrix, the standard deviation of different indicators is calculated to obtain the standard deviation The Spearman correlation coefficient of each indicator with other indicators is calculated to obtain the Spearman correlation coefficient matrix The more inconsistent the scores of indicator j and indicator k are, the lower r j,k . The value of the jth indicator and other indicators can represent a measure of the amount of information transmitted by the jth indicator. The amount of information transmitted by the jth indicator can be represented as:

[0100]

[0101] The higher the value of the corresponding indicator, the greater the amount of information transmitted by the corresponding indicator, and the higher the relative importance of the total indicator. Therefore, the weight of the indicator can be represented as:

[0102]

[0103] The comprehensive score Score obtained for each batch is:

[0104]

[0105] where is the row vector of the score matrix φ, is the score of each indicator, and ω i is the weight of each indicator.

[0106] For each batch of Score, according to its score segment, it can be divided into good, good, medium or poor evaluation level. Generally, if the Score is in the range of [0, 0.6], it is considered to be poor; in the range of (0.6, 0.7], it is considered to be medium; in the range of (0.7, 0.85], it is considered to be good; in the range of (0.85, 1], it is considered to be excellent.

[0107] S22, dynamic graph construction

[0108] In order to construct the operating condition prediction model based on the graph neural network, it is necessary to construct the dynamic graph with batches as nodes. Most of the existing methods convert the graph structure learning problem into a distance or feature similarity based metric learning problem. The embodiment mainly determines the graph structure based on the Euclidean distance of the operating condition oriented extracted features, constructs the K nearest neighbor graph of the node, and the difference is that the K nearest neighbor here limits the number of out-degree neighbor nodes to K.

[0109] Each graph in the dynamic graph can be regarded as a static graph. The historical batches are the produced batches, and the operating condition evaluation level is known, which is the labeled known node in the graph. A batch to be produced in the future has not been produced, and the operating condition evaluation level is unknown, which is a labeled unknown node. A batch corresponds to a node, and the production plan information of all nodes is known. It is assumed that the node embedding feature of the production plan information is V i p , and the node embedding feature of the key variable of each batch is H i , i represents the batch index. The known label node uses the production plan information embedding feature and the embedding feature of the key variable, and the unknown label node uses the production plan information embedding feature to construct the graph structure.

[0110] Take the construction of the dynamic graph using 10 historical batches and one batch to be produced in the future as an example to introduce the graph construction method. First, calculate the Euclidean distance between the features [V i p ; H i ] of the known label nodes in the current graph, and sequentially determine if the index of the node with smaller distance is greater than the index of the current node. If so, there is a directed edge from the current node to the node with smaller distance. Here, the index constraint is to limit the directed edge to always point from the current node to the node with later production time than the current node. As shown in Figure 5 , using the symbol B to represent the batch node, and B i-7 , the node with the smallest distance to B i-9 , the four nodes with the smallest distance from small to large are B i-9 , B i-6 , B i-4 , and B i-2 . However, since Bi-9 In B i-7 Before production, B i-7 →B i-9 The directed edge between does not exist. If the conditions permit, the node has K out-degree neighbor nodes, and K is a self-selected parameter.

[0111] After constructing the directed edge between the labeled nodes, it is necessary to determine the directed edge between the current labeled node and the unlabeled node. Since the known information of the unknown label node is less, the directed edge is determined in combination with the process knowledge. According to the process characteristics of the hot strip rolling process, the adjacent production batches have similarity in equipment process conditions and rolling state, and there is a time sequence correlation; the same steel grade and specification of the strip steel show similar rolling characteristics in heating parameters, material properties, chemical composition and rolling process. Therefore, for the unknown label node, the nodes of the previous two batches of adjacent production are pointed to it. At the same time, the Euclidean distance of the embedding vector of the production plan information of the node and the embedding vector of the production plan information of the known label node is calculated, and the node with the minimum embedding vector Euclidean distance is pointed to the node. As shown in Figure 5 B i-4 Node is the neighbor node determined according to the rolling plan embedding vector, B i-1 and B i-2 are the nodes corresponding to the two batches of adjacent production.

[0112] As described above, the embedding features of the known label node are [V i p ;H i ], and the features of the unknown label node are V i p The construction steps of the dynamic directed graph with K out-degree neighbor nodes of the node are as follows:

[0113] (1) For the current time slice, initialize an adjacency matrix A∈R (n+1)×(n+1) , where n is the number of labeled nodes in the current graph, and the node with the latest time occurrence is the node with unknown label.

[0114] (2) Calculate the embedding features [V i p ;H ithe distance between the node and other nodes from small to large, starting from the node with the smallest distance, judging whether the time sequence of the node is later than the current node. When the condition is met, the element at the corresponding position in the adjacency matrix is assigned a value of 1. Judge whether the out-degree neighbor node number of the current node reaches the set K, if yes, perform the same operation of step (2) on the next node, if not, continue to determine the out-degree neighbor of the current node until the out-degree neighbor node number of the current node reaches the set K. The self-loop of each node is retained, that is, the diagonal element of the adjacency matrix is 1.

[0115] (3) Calculate the embedding feature V of the unknown label node p Calculate the Euclidean distance of the corresponding part of the embedding feature of the known label node. Select the first two nodes adjacent in time sequence with the smallest distance as the in-degree neighbor nodes of the unknown label node, and record the elements at the corresponding positions of the adjacency matrix A as 1.

[0116] As shown in Figure 6 , it is a structure diagram of the directed graph of the adjacent two time slices constructed using the method proposed in the embodiment. The dynamic directed graph construction method introduced here will play a role in the construction of the graph neural network model in the subsequent running condition prediction task.

[0117] S3, propagate and update the node features in the dynamic graph to obtain the missing features of the unknown label nodes;

[0118] Among them, the embodiment uses the LGASNN module to propagate and update the node embedding features on the basis of the constructed dynamic graph. The structure of the LGASNN module is as shown in Figure 7 .

[0119] Since the related information of the key variables of the batch to be predicted is unknown, the propagation and update of the graph neural network are needed to obtain the missing features, so as to perform the subsequent running condition prediction. For the nodes with unknown labels, the production plan information is known, that is, the embedding features V i p of the production plan related information are known, so only the embedding features H i of the key variables of the known label nodes and the evaluation labels y i need to be propagated and updated. In the dynamic graph, according to the n historical batches and the future one batch B i-n …B i to be predicted, the constructed graph propagates and updates the node features.

[0120] GAT is a graph network that conducts inductive and transductive learning. It retains the first-order approximation localization property of graph convolution while adding adaptive edge weight coefficients. The importance of adjacent nodes is calculated in parallel before obtaining overall information by stacking these hidden self-attention layers.

[0121] The original GAT designs a shared linear transformation for each node through the weight matrix W, and then calculates the attention coefficient between the node and its directed neighbor based on the shared attention mechanism. First, the attention coefficient of node i and its in-degree neighbor node j and the node itself is calculated:

[0122]

[0123] In the formula, N i represents the in-degree neighbor node of node i, a is a single-layer feedforward neural network, and W is the weight matrix. The node is n K Head attention aggregation:

[0124]

[0125] In the formula, refers to the n K Head attention aggregation feature splicing, and σ is the sigmoid function.

[0126] As Figure 8 shown, it is a multi-head attention diagram of n K =2. The updated node features are obtained by splicing the features weighted by multiple attention mechanisms.

[0127] In order to capture the timing dynamic characteristics, LSTM is often used for modeling. The structure of the LSTM layer is as Figure 9 shown. It is composed of input gate, output gate, forget gate and memory cell, four gates.

[0128] The input Z t and the hidden layer feature h t-1 pass through the input gate to obtain

[0129]

[0130] After the forget gate, we get:

[0131]

[0132] After the output gate, we get:

[0133]

[0134] After the memory unit, we get:

[0135]

[0136] The cell state is updated as:

[0137]

[0138] The final output is:

[0139]

[0140] The use of LSTM layers is denoted as f LSTM , and the calculation process of the internal structure is not described. After that, the features of the initial node embedding are captured through the LSTM layer to capture the dynamic changing features, which are superimposed with the results aggregated by GAT after passing through a linear mapping layer. To stabilize the training process, a layer normalization layer (LN) is added. After the node features are propagated and updated, the features of the nodes are:

[0141]

[0142] After that, the node features are processed in the same way as the previous layer to obtain That is:

[0143]

[0144] At the same time, the same structure is used in parallel, and GraphSAGE is used instead of GAT for node feature propagation and update. GAT uses attention to propagate and update node features, and the mean aggregation of GraphSAGE provides stable features, and the use of both can obtain more comprehensive node feature information. Moreover, both are suitable for inductive learning and dynamic graph scenarios. GraphSAGE is biased towards uniform aggregation, and GAT is biased towards attention screening. Parallel use is similar to model integration, in order to expect to improve the generalization ability of the model.

[0145] GraphSAGE performs mean aggregation on the current representation of the node and the aggregated neighborhood vector, i.e.

[0146] [H i ;y i ]′ SAGE = σ (W1[H i ;y i ]+ W2 · mean j∈N(i) [H j ;y j ]) (22)

[0147] In the formula, W1 and W2 are weight matrices. Similarly, they are fused with the dynamic features extracted by LSTM and subjected to layer normalization, i.e.

[0148]

[0149] In the above formula, f with the same upper index LSTM represent the sharing of the LSTM layer.

[0150] Therefore, after the LGASNN module, the results of node feature propagation and update based on the GAT branch and the results of node feature propagation and update based on the GraphSAGE branch are obtained. For the unknown label node, the node features obtained by the graph neural network update contain and As the input of feature fusion and working condition prediction, the node embedding vector of the production plan information (including the set value and the steel grade) of the unknown label node and the node embedding vector of the production plan information (including the set value and the steel grade) of the unknown label node work together to obtain the prediction result of the running working condition.

[0151] S4, combining the missing features and embedding features of the unknown label node, realizes the industrial process running working condition prediction.

[0152] After the node feature propagation and update, the features of the unknown label node updated based on the GAT branch and the GraphSAGE branch and the embedding features of the production plan related information of the node need to be fused to obtain the predicted running working condition level. The embedding features V i p , the features based on the GAT branch and the features based on the GraphSAGE branch are spliced, and through the full connection layer and the Softmax layer, the evaluation level of the running working condition to be predicted is obtained, that is:

[0153]

[0154]

[0155] For training the working condition prediction model, first, the feature extraction model based on CVAE is trained, and the loss function is:

[0156]

[0157] The feature embedding of the production plan information is trained together with the dynamic LGASNN module, and the loss function The cross-entropy function is used, that is:

[0158]

[0159] In the formula, N is the number of training samples, n c is the number of categories, p ij is the probability that sample i belongs to category j, and y is the real label.

[0160] In summary, the embodiment provides a data-driven industrial process operating condition prediction method, which uses a dynamic LGASNN model to realize industrial process operating condition prediction. The dynamic LGASNN model extracts features of key variables based on CVAE for nodes with known labels, i.e., historical batches with known operating condition levels. The extracted features and true labels are part of the node embedding features. The steel grades and set values of known labels and unknown labels are known. The character embedding and trainable position embedding of the steel grade, which is a string, obtain the vector representation of the steel grade. The set values of the production plan are spliced with the vectorized representation of the steel grade. The feature embedding is performed using a fully connected layer to obtain the embedding features of the nodes with unknown labels, which are part of the embedding features of the nodes with known labels. The key variables and production plan related information (steel grade and production plan set value variables) are extracted separately in different ways. Not only does it solve the problem of feature embedding for nodes with known labels and nodes with unknown labels, but it also solves the problem of feature embedding for multiple heterogeneous inputs. Based on the constructed dynamic directed graph and embedded node features, the node feature propagation and update based on the LGASNN module are performed. The LGASNN module realizes the aggregation of node features and the capture of dynamic change features of nodes, and propagates and updates the node features to obtain the missing features of the unlabeled nodes. Finally, the obtained features of the unlabeled nodes and the embedded features of the production plan information are fused to obtain the operating condition prediction result of a batch to be produced in the future. For the problem of predicting the operating condition of a batch to be produced in the future, the process variables of which are unknown and some features are missing, a method and idea for predicting the operating condition are proposed.

[0161] For the dynamic LGASNN model, the embedding feature extraction of key variables based on CVAE is separately trained to avoid excessive uncertainty of node embedding features, which affects model convergence. The CVAE is used for feature extraction to avoid extracting features that are irrelevant to the operating condition category. The feature embedding process of the steel grade and production plan set information is learned through training, and it participates in the construction of the dynamic graph together with the embedding features of the key variables extracted by CVAE. Although the construction rule of the dynamic graph structure is determined, the directed graph structure corresponding to the time slice still needs to be calculated. The feature embedding of the steel grade and production plan set information, the LGASNN module, and the feature fusion for operating condition prediction are backpropagated together. The goal of training and learning is to make the loss function of formula (28) smaller. The model training process is shown in Figure 10 .

[0162] During the test process, the node embedding module, LGASNN module and working condition prediction module of each dynamic graph are determined. After node feature embedding, the relevant calculations are performed according to the rules of directed graph construction to construct the graph. For each time slice of the static graph, the feature propagation and update are performed through the LGASNN module to obtain the node features updated by the unknown label, and the prediction result of the running condition is finally obtained by combining the embedding features of the unknown label node itself.

[0163] Based on the above, the embodiment proposes a feasible solution for running condition prediction. First, the working condition prediction problem to be solved is described and the corresponding variables are defined. Second, the setting information of the production plan, the string form input, and how to perform node feature embedding of key variables are introduced. When embedding features of labeled nodes and unlabeled nodes, the CVAE-based feature extraction method is used for key variables, the character embedding and position embedding method is used for string form input, and the vector converted from the string and the production plan setting value related variable are jointly embedded. Then, the method of constructing dynamic graph and batch label is described, and how the proposed dynamic LGASNN module uses node feature propagation to update the missing features of the unlabeled node. The embedding features of the production plan information of the unlabeled node and the features obtained by propagation and update are fused to obtain the final working condition prediction result. Finally, the proposed dynamic LGASNN model is analyzed.

[0164] Next, the method proposed in the embodiment is used to predict the running condition and analyze the results.

[0165] 1. Data Description

[0166] The running condition prediction method proposed in the embodiment is applied to the finishing process of the hot strip rolling process for verification. The production plan information, incoming material information, load distribution, equipment state and outlet quality related variables are used, and the specific variables are shown in Table 1.

[0167] Table 1 Variables used for running condition prediction

[0168]

[0169] The data for verification comes from the real production data of a steel plant 2150 production line in January 2023. A total of 5034 batches of data are derived, and each batch of data corresponds to a sample after processing. The 5034 batches correspond to 75 kinds of steel, that is, there are 75 kinds of strings representing steel grades. 3995 batches of data are used for training, and the training set contains 3985 samples, of which 871 samples belong to excellent, 1303 samples belong to good, 798 samples belong to medium, and 1013 samples belong to poor. 1039 batches of data are used for testing, and a total of 1029 samples are used as test samples, of which 183 samples belong to excellent, 278 samples belong to good, 221 samples belong to medium, and 347 samples belong to poor. In addition to the steel grade, the mean and standard deviation of the training sample are used to standardize the sample.

[0170] The operator intervention variable is too uncertain, and the processing of the slab data will weaken the characteristics of the operator intervention variable. For the difficult task of running condition prediction, the construction of the slab label and the working condition prediction model takes into account the operator intervention variable, which may cause the operator intervention variable related features to be unable to be effectively learned. Therefore, the label of the running condition prediction task mainly considers the thickness, crown, and width of the three quality indicators.

[0171] 2. Prediction model training

[0172] Although the running condition prediction is a prediction, it is actually a classification task, so the indicators used by the model are still the indicators used in the evaluation task, namely accuracy, recall, precision, and F1-score.

[0173] For the training of the dynamic LGASNN model, similar to the hyperparameter optimization method, the interval of the parameter is set manually, and the appropriate hyperparameters are selected according to the classification effect of the test set. The number of training rounds of the graph neural network model and the CVAE is 50. The number of historical nodes in each graph of the dynamic graph is 10. The out-degree neighbor number K of the labeled node is 3. The number of heads of the first attention layer is 4, and the number of heads of the second attention layer is 1. The learning rate η of the graph neural network model is determined as g and the batch number N of model training bg . η g The interval {0.01, 0.005, 0.001, 0.0005, 0.0001} is traversed, as shown in Figure 11 (a), when η g = 0.01, the indicators of the model, namely accuracy, recall, precision, and F1-score, are the best. The learning rate of the graph neural network model is finally determined as 0.01. N bgtraverse {80, 120, 160, 200, 240, 280, 320}, as shown in (a) of Figure 11 (b), when the classification result is good, N bg = 200. Finally, the training batch number of the graph neural network model is determined to be 200.

[0174] The learning rate η c of the node feature embedding module CVAE and the training batch number N bc are determined. c traverse {0.01, 0.005, 0.001, 0.0005, 0.0001}, as shown in (c) of Figure 11 , when η c = 0.001, the classification accuracy is the highest, and the learning rate of CVAE is determined to be 0.001. N bc traverse {40, 80, 120, 160, 200}, as shown in (d) of Figure 11 , when N bc = 160, the model classification performance is better. Therefore, the training batch number of CVAE is 160.

[0175] The node feature embedding dimension is determined. The dimension d v of the node feature embedding of the steel grade and the set value is determined. v traverse {12, 24, 36, 48, 60, 72}, as shown in (e) of Figure 11 , when d v = 24, the indicators are higher. The node feature embedding dimension of the steel grade and the set value is taken as 24. The node feature embedding dimension d H of the key variable is determined. H traverse {12, 16, 20, 24, 28, 30}, as shown in (f) of Figure 11 , when d H = 24, the classification effect is better. Therefore, the node feature embedding dimension of the key variable is taken as 24.

[0176] The branch output feature dimension d M of the LGASNN model is determined. M traverse {14, 21, 28, 35, 42}, as shown in (g) of Figure 11 , when d M = 28, the indicators are the highest. Therefore, the feature output dimension of each branch of the LGASNN model is taken as 28.

[0177] The number of hidden layer units d o of the feature fusion full connection layer is determined. o traverse {12, 24, 36, 48, 60}, as shown in (h) of Figure 11As shown in (h), when d o The classification performance is best when the number of hidden units in the fully connected layer for feature fusion is 12.

[0178] After determining the hyperparameters of the model, the loss function curves of the graph neural network model and the CVAE module training process are as follows: Figure 12 As shown. Cross-entropy loss of model class prediction. The total loss L of the CVAE module cvae The KL divergence loss L of CVAE k and reconstruction loss L r All of them were able to decrease and eventually stabilize.

[0179] 3. Prediction Results and Analysis

[0180] The operational prediction results of the trained dynamic LGASNN model on the test set are as follows: Figure 13 As shown, when the batch data includes multiple steel types and specifications, the actual labels mostly overlap with the predicted labels, verifying the effectiveness of the proposed method. However, there are still instances of prediction errors regarding operating conditions, indicating that there is room for further improvement in the algorithm's structure or training performance.

[0181] Table 2 Dynamic LGASN Ablation Experiment

[0182]

[0183] Table 3 Comparison of Results of Operating Condition Prediction Methods

[0184]

[0185]

[0186] The ablation experiment of the dynamic LGASNN model included: removing the LSTM module, the branch containing GAT, the branch containing GraphSAGE, the CVAE feature embedding module, and the input of steel grade information. Table 2 shows the operating condition prediction results of the ablation experiment of the dynamic LGASNN model. Considering all indicators, the node feature embedding module (CVAE module) and the input of steel grade information have a significant impact on the model. After removing the LSTM module, GAT branch, and GraphSAGE branch of LGASNN, the various indicators of the operating condition prediction also decreased to varying degrees. This ablation experiment demonstrates the necessity of each part of the dynamic LGASNN model. Simultaneously, the experimental results verify the effectiveness of the multi-source heterogeneous data (string and numerical) feature extraction method proposed in this paper.

[0187] Comparative experiments of the dynamic LGASNN model include static GCN and label propagation algorithms. Static GCN uses a static fully connected graph and employs AE for key variable embedding feature extraction. The embedding feature extraction modules for steel grade and production information setpoints are retained. A two-layer GCN is used to train the model, yielding the predicted operating conditions. Label propagation is then applied to the graph for each time slice, resulting in a dynamic graph-based label propagation algorithm for operating condition prediction. Figure 14 The table shows the runtime prediction results for the static GCN and label propagation algorithms. Table 3 shows the various metrics for the runtime prediction results of the dynamic LGASNN method, the static GCN method, and the label propagation algorithm. Considering all four metrics, the dynamic LGASNN method outperforms both the static GCN and the dynamic graph-based label propagation algorithm. This demonstrates that the proposed dynamic LGASNN model can improve the accuracy of runtime prediction to a certain extent.

[0188] Through the above process, this embodiment verifies the application of the proposed operating condition prediction method in the hot strip rolling process. Addressing the problem of operating condition prediction, experiments verify the effectiveness of the proposed dynamic LGASNN method under batch data with multiple steel grades and specifications. Ablation experiments demonstrate the necessity of each module in the model. Compared with static GCN and label propagation algorithms based on dynamic graphs, the accuracy of operating condition prediction is improved by an average of 7%. The proposed method can effectively extract evaluation category-related features contained in batch information and obtain the missing features of unproduced batches through directed propagation and updating of node embedded features on the dynamic graph, providing a feasible solution to the difficulty of operating condition prediction.

[0189] Second Embodiment

[0190] This embodiment provides an electronic device, such as... Figure 15 As shown, the electronic device includes a processor and a memory; wherein the processor and the memory can be connected via a communication bus; the memory stores at least one instruction, which is loaded and executed by the processor to implement the method of the first embodiment described above. Furthermore, the electronic device may also include a transceiver, the processor and the transceiver can be connected via a communication bus, and the transceiver is used to communicate with other devices.

[0191] Below, in conjunction with Figure 15 A detailed introduction to each component of this electronic device is provided below:

[0192] The processor is the control center of the electronic device. The electronic device can include multiple processors. Each of the processors can be a single-CPU or a multi-CPU. The processor can be one processor or a collective term of multiple processing elements. For example, the processor can be one or more central processing units (CPUs), other general purpose processors, application specific integrated circuits (ASICs), or one or more integrated circuits configured to implement one or more embodiments of the present application, such as one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or the like. The general purpose processor can be a microprocessor or any conventional processor, or the like. The processor can perform various functions of the electronic device by running or executing software programs stored in the memory and calling data stored in the memory.

[0193] In a specific implementation, as an embodiment, the processor can include one or more CPUs, such as CPU0 and CPU1 shown in FIG. 1, of course, this is only an exemplary description. Figure 15

[0194] The memory is used to store software programs for implementing the solution of the present application, and is controlled by the processor to perform the implementation. The specific implementation can refer to the method embodiments described above, and will not be described here.

[0195] ​Optionally, the memory may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory may be integrated with the processor or may exist independently, and may be accessed through the interface circuit of the electronic device (…). Figure 15 (Not shown in the image) is coupled to the processor; however, this embodiment of the invention does not impose specific limitations on this.

[0196] The transceiver may include a receiver and a transmitter. Figure 15 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function. The transceiver can be integrated with the processor or exist independently, and is connected through the interface circuit of the electronic device (…). Figure 15 (Not shown in the image) is coupled to the processor, and this embodiment of the invention does not specifically limit this.

[0197] In addition, it should be noted that, Figure 15 The structure of the electronic device shown is not intended to limit the device. Actual devices may include more or fewer components than shown, or combine certain components, or have different component arrangements. Furthermore, the technical effects achieved by this electronic device when performing the method of the first embodiment described above can be referenced to the technical effects described in the first embodiment; therefore, they will not be repeated here.

[0198] Third Embodiment

[0199] This embodiment provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the method of the first embodiment described above. The computer-readable storage medium may be a ROM, random access memory, CD-ROM, magnetic tape, floppy disk, or optical data storage device, etc. The instruction stored therein can be loaded and executed by a processor in a terminal.

[0200] Moreover, it should be noted that the present application can be provided as a method, an apparatus or a computer program product. Therefore, the embodiments of the present application can take the form of an entirely or partially hardware embodiment, an entirely or partially software embodiment or an embodiment combining software and hardware aspects. Furthermore, when implemented in software, the embodiments of the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, a computer diskette, an optical storage medium, a magnetic storage medium or any other medium that can be used to carry or store computer-usable program codes). The computer program product includes one or more computer-usable program codes. When the computer-usable program codes are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are produced. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable apparatus. The computer-usable program codes can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer-usable program codes can be transferred from one website, computer, server or data center to another website, computer, server or data center through a wired (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center or the like containing one or more available medium sets. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD) or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0201] It should also be noted that the terms "comprising", "including", or any other variant thereof are intended to cover a non-exclusive inclusion, such that processes, methods, articles or terminal devices including a series of elements not only include those elements, but also include other elements not explicitly listed or inherent to such processes, methods, articles or terminal devices. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or terminal device including the element.

[0202] Those skilled in the art can realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0203] Finally, it should be noted that the above description is only the preferred embodiment of the application, it should be pointed out that although the preferred embodiment of the application has been described, for those skilled in the art, once the basic creative concept of the application is known, several improvements and refinements can be made without departing from the principles of the application, and these improvements and refinements should also be considered as the protection scope of the application. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the application.

Claims

1. A data-driven method for predicting the operating conditions of an industrial process, characterized in that, include: Each production batch is treated as a node, and the embedding features of the node are obtained and node feature embedding is performed. Among them, the production batch includes historical batches with known working conditions and future batches with unknown working conditions. The nodes corresponding to the historical batches are known label nodes, and the nodes corresponding to the future batches are unknown label nodes. Construct a dynamic graph containing multiple known label nodes and one unknown label node; In the dynamic graph, node features are propagated and updated to obtain the missing features of unknown label nodes; By combining the missing and embedded features of unknown label nodes, the operating conditions of industrial processes can be predicted. Construct a dynamic graph containing multiple known label nodes and one unknown label node, including: For the current time slice, initialize an adjacency matrix A∈R with all elements equal to 0. (n+1)×(n+1) Where n is the number of known label nodes in the current graph, and the node whose time occurs latest is the unknown label node; Calculate the Euclidean distance between the embedding features of the first n nodes. Sort the Euclidean distances between the current node and other nodes in ascending order. Starting from the node with the smallest distance, determine if its temporal sequence is later than the current node. If its temporal sequence is later than the current node, set the corresponding element in the adjacency matrix to 1. Determine if the number of out-degree neighbors of the current node has reached the set value K. If yes, perform the same operation on the next node. If not, continue to determine the out-degree neighbors of the current node until the number of out-degree neighbors reaches the set value K. Preserve the self-loop of each node, i.e., set the diagonal elements of the adjacency matrix to 1. Calculate the Euclidean distance between the embedding features of the unknown label node and the embedding features of the production plan information of the known label node; select the node corresponding to the minimum Euclidean distance and the two nodes that are adjacent to the unknown label node in time sequence as the in-degree neighbors of the unknown label node, and denote the corresponding element of the adjacency matrix A as 1; In the dynamic graph, node features are propagated and updated to obtain the missing features of nodes with unknown labels, including: The dynamic graph is input into the LGASNN module, which then uses the LGASNN module to propagate and update node features in the dynamic graph to obtain the missing features of unknown labeled nodes. The LGASNN module includes the GAT branch and the GraphSAGE branch; the process of using the LGASNN module to propagate and update node features in a dynamic graph includes: In the GAT branch, the node's embedded features are input into LSTM and GAT respectively. LSTM captures dynamically changing features, which are then aggregated by GAT. The dynamically changing features captured by LSTM are passed through a linear mapping layer and superimposed with the result of aggregation by GAT. The superimposed result is input into a normalization layer. The output of the normalization layer is processed by ReLU to obtain the first feature. The first feature is then input into LSTM and GAT respectively. LSTM captures dynamically changing features, which are then aggregated by GAT. The dynamically changing features captured by LSTM are passed through a linear mapping layer and superimposed with the result of aggregation by GAT. The superimposed result is input into a normalization layer. The output of the normalization layer is processed by ReLU to obtain the updated node features, denoted as the GAT branch features. In the GraphSAGE branch, the node's embedded features are input into LSTM and GraphSAGE respectively. LSTM captures dynamically changing features, and GraphSAGE performs mean aggregation. The dynamically changing features captured by LSTM are passed through a linear mapping layer and then superimposed with the result of mean aggregation by GraphSAGE. The superimposed result is input into a normalization layer. The output of the normalization layer is passed through a ReLU mechanism to obtain the second feature. The second feature is then input into LSTM and GraphSAGE respectively. LSTM captures dynamically changing features, and GraphSAGE performs mean aggregation. The dynamically changing features captured by LSTM are passed through a linear mapping layer and then superimposed with the result of mean aggregation by GraphSAGE. The superimposed result is input into a normalization layer. The output of the normalization layer is passed through a ReLU mechanism to obtain the updated node features, denoted as the GraphSAGE branch features. GAT branch features and GraphSAGE branch features together constitute the missing features of unknown label nodes.

2. The data-driven industrial process operating condition prediction method as described in claim 1, characterized in that, For nodes with unknown labels, the process of obtaining the node's embedding features includes: Collect the product model and process settings corresponding to unknown tag nodes; Obtain the vector representation of the product model, then concatenate the vector representation of the product model with the process setting value, and input the concatenation result into a fully connected layer for feature embedding to obtain the embedded features of the unknown label node.

3. The data-driven industrial process operation condition prediction method as described in claim 2, characterized in that, For a known labeled node, the process of obtaining the node's embedding features includes: Collect the product model, process setting value, preset index variable and operating condition label corresponding to the known label node; Feature extraction is performed on the preset indicator variables to obtain the feature data of the preset indicator variables; Obtain the vector representation of the product model, concatenate the vector representation of the product model with the process setting value, and input the concatenation result into the fully connected layer for feature embedding to obtain the production plan information embedding feature; The embedded features of the production plan information corresponding to the known label node, the feature data of the preset indicator variables, and the working condition label are concatenated to obtain the embedded features corresponding to the known label node.

4. The data-driven industrial process operating condition prediction method as described in claim 3, characterized in that, The feature data of the preset indicator variables are extracted through a conditional variational autoencoder.

5. The data-driven industrial process operating condition prediction method as described in claim 3, characterized in that, In the aforementioned dynamic graph, the directed edges of known label nodes are determined as follows: Calculate the Euclidean distance between the current known label node and the embedding features of every other known label node, and select the known label nodes whose corresponding Euclidean distance is less than a preset threshold as candidate nodes; Determine whether the index of each candidate node is greater than the index of the currently known label node; If the index of a candidate node is greater than the index of a currently known label node, then there exists a directed edge from the currently known label node to the corresponding candidate node; otherwise, there is no directed edge from the currently known label node to the corresponding candidate node.

6. The data-driven industrial process operating condition prediction method as described in claim 3, characterized in that, In the aforementioned dynamic graph, the directed edges for unknown label nodes are determined as follows: For an unknown tag node, point the nodes corresponding to the two nearest-produced batches to it; Calculate the Euclidean distance between the embedding features of the unknown label node and the production plan information embedding features of each known label node, and point the known label node corresponding to the smallest Euclidean distance to this unknown label node.

7. The data-driven industrial process operating condition prediction method as described in claim 5 or 6, characterized in that, In the dynamic graph, a node has K out-degree neighbor nodes; K is a parameter that can be selected and set by the user.

8. The data-driven industrial process operating condition prediction method as described in claim 1, characterized in that, By combining the missing and embedding features of unknown label nodes, the operating conditions of industrial processes can be predicted, including: The missing features and embedded features of the unknown label nodes are concatenated, and the concatenated features are input into a fully connected layer. Through the fully connected layer and the Softmax layer, the prediction results of the industrial process operation conditions are obtained.

Images