Complex industrial process operation state evaluation method based on SelfAtt-EvolveGCN

Abstract

The invention provides a method for evaluating the operation state of a complex industrial process based on a SelfAtt-EvolveGCN. Establishing a running state offline evaluation model; evaluating and classifying the operation state; obtaining a sequence data set X with consistent sample quantity; mining a topological structure relationship At among variables in the sliding window, and constructing association graph structure sequence data G; inputting the graph structure data into the operation state offline evaluation model for training, inputting online data into the trained evaluation model, obtaining a weight probability value belonging to each state level in the classifier, and taking the state level where the maximum probability value is located as a current prediction state; and when the operation state evaluation model evaluates that the current operation state is non-optimal, performing non-optimal factor tracing, and positioning specific variables causing the occurrence of the non-optimal state. According to the method, disturbance influence caused by a complex industrial severe environment can be resisted, effective extraction of time sequence structural features in process variable data is realized, and a complete and accurate operation state evaluation model is established.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial production process operation status evaluation, and specifically relates to a complex industrial process operation status evaluation method based on SelfAtt-EvolveGCN. Background Art

[0002] The automation level of modern industrial systems is constantly improving, and its production process involves numerous process steps and interactions between equipment, and the complexity and coupling of industrial processes have increased significantly. This complexity makes the operating status of industrial processes susceptible to interference from various factors, and various abnormal situations often occur in the production process, such as equipment failures and product quality problems. These abnormal situations may lead to problems such as reduced production efficiency, reduced product quality, increased production costs, and may even cause safety accidents. Therefore, it is necessary to be able to fully explore the hidden features in the large amount of recorded historical data of process variables. Therefore, detailed monitoring and evaluation of the status of industrial production processes is of great significance to help on-site operators understand the process status in a timely manner, provide adjustment strategies, and ensure production efficiency and product quality.

[0003] Traditional operating status evaluation methods, such as PCA, T-PLS, and ORVI, analyze the static characteristics of data and cannot capture the relevant changes in dynamic information in industrial processes. The increase in complexity and coupling in modern industry has led to strong nonlinear relationships in process variables, and traditional linear methods are difficult to achieve satisfactory results. Due to its strong nonlinear processing capabilities, neural networks have been introduced into various state evaluation methods. For example, in view of the strong nonlinearity, information redundancy, and other uncertain factors in complex industrial processes, an operating status evaluation method based on a sparse denoising autoencoder driven by comprehensive economic indicators is proposed; an operating status evaluation method based on a stacked supervised denoising autoencoder is combined with a supervised learning strategy that guides the model with state level label data, forcing the network to pay attention to the associated features with the state level during the learning process, thereby enhancing the model's ability to distinguish the state level. However, this type of autoencoder method assumes that samples are independent of each other and cannot extract the temporal dynamic relationship contained in the sample sequence. Later, the LSTM network, due to its unique gate structure, has advantages in modeling long sequence data and has been introduced into fault diagnosis, process monitoring, and state evaluation methods for industrial production, enabling the model to perceive dynamic changes in the industrial process. In response to various problems in industrial processes, a large amount of research work on LSTM networks has emerged. However, in complex industrial process variables, in addition to the temporal characteristics of the variables themselves, different variables will also affect each other, and data at different state levels contain different structural characteristics. Recurrent neural network methods for temporal sequence modeling cannot effectively extract the topological structural characteristics between process variable data, and cannot perceive the complex structural information of non-Euclidean data in the sampled samples. To address this problem, models such as graph neural networks (GNNs) can be introduced. By representing data as a graph structure and using models such as graph convolutional networks to aggregate and update nodes and edges, the topological structural characteristics between data can be effectively captured and complex structural non-Euclidean data can be processed. Although graph convolutional network models can focus on structural characteristics, they will ignore the dynamic temporal characteristics of process industries. Therefore, for complex industrial process modeling, it is necessary to consider extracting both structural characteristics and temporal characteristics to accurately describe the implicit characteristics of industrial operation. Summary of the invention

[0004] In view of the problems existing in the above-mentioned existing status evaluation methods, the present invention provides a complex industrial process operation status evaluation method based on the SelfAtt-EvolveGCN network. The method can perceive the non-Euclidean structural relationship existing in the variable data of the complex industrial process, and extract the topological structure information between the variables by constructing the association graph structure; at the same time, it pays attention to the temporal nature between the topological structure features, ensuring that the model not only learns the topological relationship between the variables, but also takes into account the dynamic time-varying characteristics of the complex industrial process, and helps the complex industrial process to establish a complete and accurate operation status evaluation model.

[0005] In order to achieve the above object, the present invention provides a method for evaluating the operating status of a complex industrial process based on SelfAtt-EvolveGCN, which specifically includes the following steps:

[0006] Step 1: Use offline data sets to train the operation status evaluation model based on the SelfAtt-EvolveGCN network;

[0007] A1: Collecting historical data of complex industrial processes Where N is the number of sampling points, x=[x1,x2,...,x N ] is the industrial process variable data, y=[y1,y2,…,y N ] is the selected comprehensive economic indicator data;

[0008] A2: Perform standardization calculation on the x data and y data according to formula (1) and unify the dimensions so that the expected value of the calculated data is 0 and the standard deviation is 1;

[0009]

[0010] In the formula, X represents the standardized data, x is the input data, μ represents the mathematical expectation of the data, and σ represents the variance of the data;

[0011] A3: Based on expert knowledge and comprehensive economic indicator data, classify y into state levels and label the data in turn to obtain

[0012] A4: Use sliding window to N×M Perform continuous sampling, where N is the sample size and M is the process variable dimension. Get sequence data Where X t ∈R L×M is the sampling data of a window, T is the number of samples in the sliding window, and L represents the length of the sliding window;

[0013] A5: Sample data X in a window t ∈R L×M , using the self-attention mechanism, Xt Map them to three different vector spaces: Q (query), K (key), and V (value), and calculate the attention weight matrix score;

[0014] Q=X t W Q +b Q (2);

[0015] K=X t W K +b K (3);

[0016] V=X t W V +b V (4);

[0017]

[0018] Where W Q , W K , W V is a learnable weight parameter, b Q , b K , b V is the corresponding bias term; the variance of the elements in α is d, divided by the root mean square of the scaling factor d, which ensures that the variance of the elements in α is 1;

[0019] A6: After obtaining the attention weight matrix score according to the self-attention mechanism, it needs to be further converted into an adjacency matrix to reflect the connection relationship of the nodes in the graph structure;

[0020]

[0021] Among them, δ is the judgment threshold set according to expert knowledge. When the element in the weight matrix is ​​greater than or equal to δ, it is considered that there is a correlation between the two nodes, and the element value is set to 1; conversely, if it is less than δ, it is considered that the two nodes are independent of each other and there is no correlation, and the element will be set to 0;

[0022] A7: Get a correlation graph structure representation G within a sliding window t =(X t ,A t ), and then obtain the set of association graph structures at all times in the data set, namely

[0023] A8: The existence of the node itself is not considered in the constructed adjacency matrix A. The adjacency matrix is ​​standardized as follows through a self-loop method:

[0024]

[0025] In the formula, represents the degree matrix, represents the normalized adjacency matrix;

[0026] A9: Input the network model and randomly initialize the EvolveGCN network parameters, including the connection weight W and bias b of each network layer;

[0027]

[0028] Where σ is the activation function, W t is the GCN weight parameter, and the initial embedding matrix comes from the node features, that is,

[0029] A10: Use LSTM to iteratively update the GCN weight parameters. The specific formula is as follows:

[0030]

[0031] 1: Function H t =f(X t )

[0032] 2:F t =σ(M F X t +U F H t-1 +B F )

[0033] 3:I t =σ(M I X t +U I H t-1 +B I )

[0034] 4:O t =σ(M O X t +U O H t-1 +B O )

[0035] 5:

[0036] 6:

[0037] 7:

[0038] 8: End function

[0039] In the formula, the current input X of the function tWith the output feature H of the previous moment t-1 Same, both are the weight parameters W of the graph convolutional network at the previous moment t-1 ; In the function formula, F, I, and O represent the update process of the three gate structures of LSTM forget gate, input gate, and output gate respectively; is the candidate state vector of the memory unit, C t Represents a memory unit; M and U represent the learnable weight parameters in each gate structure, and B represents the corresponding bias; σ and tanh are activation functions;

[0040] A11: Extract the feature vector Using the Softmax classification layer, the weighted probability that the current predicted state belongs to each state is calculated (Formula 10), and finally the maximum value of the probability of each state is taken as the final model prediction state (Formula 11);

[0041]

[0042] In the formula It means that Softmax calculates the probability set of the divided state levels. It represents the predicted state level represented by the state with the maximum probability value, p represents the state level digital label, and C represents the total number of divided state levels;

[0043] A12: Using the cross entropy loss function as the objective function of the running status offline evaluation model training process, the loss function is calculated according to formula (12);

[0044]

[0045] Where N represents the total number of labels in the corresponding sample set, y L is the real label data, is the label data predicted by the model;

[0046] A13: Use the Adam optimization algorithm that can adaptively adjust the learning rate to iteratively optimize the model's learning parameters W and B until the loss function J is loss Minimize or converge to stability, and obtain an offline evaluation model of the operating status;

[0047] Step 2: Use online data to evaluate the operation status;

[0048] B1: Sample to obtain online process variable data x and y, and standardize x to obtain X;

[0049] B2: Based on expert knowledge, the collected comprehensive economic indicator data y is classified into status levels and marked with digital labels in turn to obtain

[0050] B3: Use a sliding window with a window length of L to sample X in sequence to obtain a sequence data set

[0051] B4: For each sampled data X in the window t ∈R L×M , using the self-attention mechanism, X t Map them to three different vector spaces: Q (query), K (key), and V (value), and calculate the attention weight matrix score;

[0052] B5: After obtaining the attention weight matrix score according to the self-attention mechanism, it is further converted into an adjacency matrix A that can reflect the node connection relationship in the graph structure t ;

[0053] B6: Get a correlation graph structure representation G within a sliding window t =(X t ,A t ), then the set of association graph structures at all times in the data set is

[0054] B7: Standardize the adjacency matrix by self-looping to obtain

[0055] B8: Input the trained offline evaluation model of the running status, and the Softmax classification layer outputs the weighted probability set of the current predicted state belonging to each state The final prediction evaluation result is the operating state with the maximum probability.

[0056] When the operation status evaluation model determines that the current process is in a non-optimal state, the non-optimal factors are identified. The specific steps are as follows:

[0057] S1: Use high-quality data to train the EvolveGCN-AE model;

[0058] S11: Sampling and selecting the process variable x with the grade of "excellent" = (x1, ..., x N ), and perform standardization on x to obtain X;

[0059] S12: Use the self-attention mechanism to construct the process variable association graph structure data and standardize it, that is, the adjacency matrix A;

[0060] S13: Take X and A as the input of the EvolveGCN-AE model to obtain the output reconstructed variable data The error between input and output is calculated according to formula (13), which is the sum of the errors at all times;

[0061]

[0062] In the formula, N is the number of samples sampled at the time;

[0063] S14: Iteratively optimize the EvolveGCN-AE model, continuously update the weight terms and bias terms, minimize the loss function or stabilize the convergence, and obtain a trained non-optimal factor tracing model;

[0064] S2: Use the “non-optimal” status data as the input of the traceability model to identify non-optimal factors:

[0065] S21: Collect and select non-optimal process variable data and perform standardization processing;

[0066] S22: Use the self-attention mechanism to calculate the node correlation and obtain the standardized adjacency matrix;

[0067] S23: The standardized process variables and adjacency matrix are input into the trained EvolveGCN-AE model. According to the SPE contribution graph definition, the contribution of each process variable to the non-optimal state is calculated. The formula is as follows:

[0068]

[0069] In the formula, x t i represents the i-th process variable at time t. Similarly, represents the corresponding reconstruction value;

[0070] S24: Calculate the contribution of each variable and compare them. The variable with a large contribution rate is defined as the causal variable that leads to the non-optimal state.

[0071] The present invention calculates the correlation degree between variables by using the self-attention mechanism and constructs the process variable association graph structure; introduces the EvolveGCN network, so that the established model can perceive the non-Euclidean structure information existing in the industrial data, learn the topological structure relationship between variables, pay attention to the dynamic time-varying characteristics of the industrial process, and comprehensively extract the hidden characteristics of the temporal structure in the factory data. According to the comprehensive economic indicators that can reflect the benefits and production efficiency of the enterprise, the different operating state levels existing in the industrial production process are defined and divided. Then, the selected comprehensive economic indicators and the divided state level supervision training model are used to establish an accurate industrial process operation state evaluation model. When the state evaluation model determines that the current state level is "non-optimal", the non-optimal data is reconstructed by the EvolveGCN-AE model, and the contribution degree of each process variable to the non-optimal state is calculated according to the contribution graph algorithm. By comparing the size of the contribution rate of each variable, the main process variables are located, and the on-site management personnel can find the cause in time and provide guidance strategies for regulating the equipment, which is helpful to improve the production quality and efficiency of the product. The method can perceive the current operating state of the complex industrial process in a detailed and accurate manner, and find the changes and deviations of the operating state in time; and when the state level is non-optimal, the main cause process variables can be traced back, which is convenient for the production management of the enterprise. This method can effectively monitor the subtle changes in complex industrial production, improve the productivity of high-quality finished products in industrial production, discover and reduce unnecessary raw material consumption, and optimize the overall economic benefits of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 The overall flow chart of the industrial process operation status evaluation in the present invention;

[0073] Figure 2 It is a schematic diagram of a complex industrial process operation status evaluation method based on the SelfAtt-EvolveGCN network in the present invention;

[0074] Figure 3 It is the online evaluation result of the operation status method based on SelfAtt-EvolveGCN in the present invention;

[0075] Figure 4 It is the model evaluation result when the variable is missing in the present invention compared with the normal situation;

[0076] Figure 5 is the model evaluation result under different sample loss levels in the present invention;

[0077] Figure 6 It is the result of tracing the non-optimal factors in the present invention. DETAILED DESCRIPTION

[0078] like Figure 2As shown, the present invention provides a method based on introducing a self-attention mechanism for evolving graph convolutional networks (Self-Attention mechanism introduced for Evolving Graph Convolutional Networks, SelfAtt-EvolveGCN), and the method specifically includes the following steps:

[0079] Step 1: Use offline data sets to train the operation status evaluation model based on the SelfAtt-EvolveGCN network;

[0080] A1: Collecting historical data of complex industrial processes Where N is the number of sampling points, x=[x1,x2,…,x N ] is the industrial process variable data, y=[y1,y2,…,y N ] is the selected comprehensive economic indicator data;

[0081] A2: Perform standardization calculation on the x data and y data according to formula (1) and unify the dimensions so that the expected value of the calculated data is 0 and the standard deviation is 1;

[0082]

[0083] In the formula, X represents the standardized data, x is the input data, μ represents the mathematical expectation of the data, and σ represents the variance of the data;

[0084] A3: Based on expert knowledge and comprehensive economic indicator data, classify y into state levels and label the data in turn to obtain

[0085] A4: Use sliding window to N×M Perform continuous sampling, where N is the sample size and M is the process variable dimension. Get sequence data Where X t ∈R L×M is the sampling data of a window, T is the number of samples in the sliding window, and L represents the length of the sliding window;

[0086] A5: Sample data X in a window t ∈R L×M , using the self-attention mechanism, X t Map them to three different vector spaces: Q (query), K (key), and V (value), and calculate the attention weight matrix score;

[0087] Q=X t W Q +b Q (2);

[0088] K=X t W K +b K (3);

[0089] V=X t W V +b V (4);

[0090]

[0091] Where W Q , W K , W V is a learnable weight parameter, b Q 、b K 、b V is the corresponding bias term; the variance of the elements in α is d, divided by the root mean square of the scaling factor d, which ensures that the variance of the elements in α is 1;

[0092] A6: After obtaining the attention weight matrix score according to the self-attention mechanism, it needs to be further converted into an adjacency matrix to reflect the connection relationship of the nodes in the graph structure;

[0093]

[0094] Among them, δ is the judgment threshold set according to expert knowledge. When the element in the weight matrix is ​​greater than or equal to δ, it is considered that there is a correlation between the two nodes, and the element value is set to 1; conversely, if it is less than δ, it is considered that the two nodes are independent of each other and there is no correlation, and the element will be set to 0;

[0095] A7: Get a correlation graph structure representation G within a sliding window t =(X t ,A t ), and then obtain the set of association graph structures at all times in the data set, namely

[0096] A8: The existence of the node itself is not considered in the constructed adjacency matrix A. The adjacency matrix is ​​standardized as follows through a self-loop method:

[0097]

[0098] In the formula, represents the degree matrix, represents the normalized adjacency matrix;

[0099] A9: Input the network model and randomly initialize the EvolveGCN network parameters, including the connection weight W and bias b of each network layer;

[0100]

[0101] Where σ is the activation function, W t is the GCN weight parameter, and the initial embedding matrix comes from the node features, that is,

[0102] A10: Use LSTM to iteratively update the GCN weight parameters. The specific formula is as follows:

[0103]

[0104] 1: Function H t =f(X t )

[0105] 2:F t =σ(M F X t +U F H t-1 +B F )

[0106] 3:I t =σ(M I X t +U I H t-1 +B I )

[0107] 4:O t =σ(M O X t +U O H t-1 +B O )

[0108] 5:

[0109] 6:

[0110] 7:

[0111] 8: End function

[0112] In the formula, the current input X of the function t With the output feature H of the previous moment t-1 Same, both are the weight parameters W of the graph convolutional network at the previous moment t-1 ; In the function formula, F, I, and O represent the update process of the three gate structures of LSTM forget gate, input gate, and output gate respectively; is the candidate state vector of the memory unit, C t Represents a memory unit; M and U represent the learnable weight parameters in each gate structure, and B represents the corresponding bias; σ and tanh are activation functions;

[0113] A11: Extract the feature vector Using the Softmax classification layer, the weighted probability that the current predicted state belongs to each state is calculated (Formula 10), and finally the maximum value of the probability of each state is taken as the final model prediction state (Formula 11);

[0114]

[0115] In the formula It means that Softmax calculates the probability set of the divided state levels. It represents the predicted state level represented by the state with the maximum probability value, p represents the state level digital label, and C represents the total number of divided state levels;

[0116] A12: Using the cross entropy loss function as the objective function of the running status offline evaluation model training process, the loss function is calculated according to formula (12);

[0117]

[0118] Where N represents the total number of labels in the corresponding sample set, y L is the real label data, is the label data predicted by the model;

[0119] A13: Use the Adam optimization algorithm that can adaptively adjust the learning rate to iteratively optimize the model's learning parameters W and B until the loss function J is loss Minimize or converge to stability, and obtain an offline evaluation model of the operating status;

[0120] Step 2: Use online data to evaluate the operation status;

[0121] B1: Sample to obtain online process variable data x and y, and standardize x to obtain X;

[0122] B2: Based on expert knowledge, the collected comprehensive economic indicator data y is classified into status levels and marked with digital labels in turn to obtain

[0123] B3: Use a sliding window with a window length of L to sample X in sequence to obtain a sequence data set

[0124] B4: For each sampled data X in the window t ∈RL×M , using the self-attention mechanism, X t Map them to three different vector spaces: Q (query), K (key), and V (value), and calculate the attention weight matrix score;

[0125] B5: After obtaining the attention weight matrix score according to the self-attention mechanism, it is further converted into an adjacency matrix A that can reflect the node connection relationship in the graph structure t ;

[0126] B6: Get a correlation graph structure representation G within a sliding window t =(X t ,A t ), then the set of association graph structures at all times in the data set is

[0127] B7: Standardize the adjacency matrix by self-looping to obtain

[0128] B8: Input the trained offline evaluation model of the running status, and the Softmax classification layer outputs the weighted probability set of the current predicted state belonging to each state The final prediction evaluation result is the operating state with the maximum probability.

[0129] When the operation status evaluation model determines that the current process is in a non-optimal state, the non-optimal factors are identified. The specific steps are as follows:

[0130] S1: Use high-quality data to train the EvolveGCN-AE model;

[0131] S11: Sampling and selecting the process variable x with the grade of "excellent" = (x1, ..., x N ), and perform standardization on x to obtain X;

[0132] S12: Use the self-attention mechanism to construct the process variable association graph structure data and standardize it, that is, the adjacency matrix A;

[0133] S13: Take X and A as the input of the EvolveGCN-AE model to obtain the output reconstructed variable data The error between input and output is calculated according to formula (13), which is the sum of the errors at all times;

[0134]

[0135] In the formula, N is the number of samples sampled at the time;

[0136] S14: Iteratively optimize the EvolveGCN-AE model, continuously update the weight terms and bias terms, minimize the loss function or stabilize the convergence, and obtain a trained non-optimal factor tracing model;

[0137] S2: Use the “non-optimal” status data as the input of the traceability model to identify non-optimal factors:

[0138] S21: Collect and select non-optimal process variable data and perform standardization processing;

[0139] S22: Use the self-attention mechanism to calculate the node correlation and obtain the standardized adjacency matrix;

[0140] S23: The standardized process variables and adjacency matrix are input into the trained EvolveGCN-AE model. According to the SPE contribution graph definition, the contribution of each process variable to the non-optimal state is calculated. The formula is as follows:

[0141]

[0142] In the formula, represents the i-th process variable at time t. Similarly, represents the corresponding reconstruction value;

[0143] S24: Calculate the contribution of each variable and compare them. The variable with a large contribution rate is defined as the causal variable that leads to the non-optimal state.

[0144] The present invention calculates the correlation degree between variables by using the self-attention mechanism to construct the process variable association graph structure; introduces EvolveGCN, so that the established model can perceive the non-Euclidean structure information existing in industrial data, learn the topological structure relationship between variables, pay attention to the dynamic time-varying characteristics of the industrial process, and comprehensively extract the hidden characteristics of the temporal structure in the factory data. According to the comprehensive economic indicators that can reflect the benefits and production efficiency of the enterprise, the different operating state levels existing in the industrial production process are defined and divided. Then, the selected comprehensive economic indicators and the divided state level supervision training model are used to establish an accurate industrial process operation state evaluation model. When the state evaluation model determines that the current state level is "non-optimal", the non-optimal data is reconstructed by the EvolveGCN-AE model, and the contribution degree of each process variable to the non-optimal state is calculated according to the contribution graph algorithm. By comparing the size of the contribution rate of each variable, the main process variables are located, and the on-site management personnel can find the cause in time and provide guidance strategies for regulating the equipment, which is helpful to improve the production quality and efficiency of the product. The method can perceive the current operating state of the complex industrial process in a detailed and accurate manner, and find the changes and deviations of the operating state in time; and when the state level is non-optimal, the main cause process variables can be traced back, which is convenient for the production management of the enterprise. This method can effectively monitor the subtle changes in complex industrial production, improve the productivity of high-quality finished products in industrial production, discover and reduce unnecessary raw material consumption, and optimize the overall economic benefits of the enterprise.

[0145] The technical solution of the present invention is described in detail below in conjunction with embodiments, and its feasibility is verified.

[0146] Example:

[0147] The heavy medium coal preparation process is a typical complex industrial process in a harsh open environment. With the various interference and uncertainty factors in production operation, the process variables before and after the process affect each other. Traditional evaluation methods often focus on the temporal changes of the variables themselves, while ignoring the structural characteristics and change information between the variables. The present invention aims at the problem that the non-Euclidean structure of the heavy medium coal preparation process data is difficult to perceive, and the topological structure relationship between process variables and the dynamic time-varying characteristics of the structure are difficult to learn. It can extract the implicit characteristics of the industrial process more completely and accurately, and has a certain robustness in the case of missing variables and missing samples.

[0148] The industrial process selected is a two-product conical heavy medium coal preparation process. The main equipment used in this process includes raw coal particle size grading screen, mixing barrel, heavy medium cyclone, magnetic separator, de-mediation screen, qualified medium barrel, etc. First, the raw coal is crushed into coarse coal of different particle sizes by a coal gangue crusher, and transported to the primary grading screen and the secondary grading screen by a conveyor belt, where the particle size is graded to obtain raw coal of different particle sizes; then the raw coal of appropriate particle size in the lower screen is collected, and mixed with the heavy medium suspension in the qualified medium barrel in a proportion in the mixing barrel to obtain a coal-medium mixture of a certain concentration; then, the material is sent to the heavy medium cyclone by a pressure pump for sorting, and the overflow and underflow products of the cyclone are dehydrated and de-medium treated to obtain clean coal with ash content that meets the standard and waste materials such as gangue with a high ash content. The magnetic medium in the heavy medium cyclone is screened out by the magnetic separator, and the dilute medium is sent to the qualified medium barrel, replenished with new magnetic medium, and water is added to the appropriate density to form a new heavy medium suspension, which continues to participate in the new coal washing cycle.

[0149] Through the understanding of the heavy medium coal preparation process and the analysis of the factors affecting coal products, 11 variables were finally selected as process variables. The specific process variables are shown in Table 1. The quality judgment index of the final coal product is generally the ash content of the clean coal, so the ash value is selected as a reference index for judging the operating status level of the coal preparation process.

[0150] Table 1: Process operation variable selection

[0151]

[0152] The data of this example comes from the full-process simulation platform of the heavy medium coal preparation process (registration number: 2020SR1160052). In order to generate heavy medium coal preparation process data of different state levels, three main operating variables in the coal preparation process are set in the full-process simulation platform of the heavy medium coal preparation process: the raw coal feed amount of the coal preparation plant (variable 1), the cyclone feed pressure (variable 7) and the density of the heavy medium suspension output from the qualified medium barrel (variable 9). By introducing fluctuations to simulate the changes in state levels, the fluctuation generating functions of variables 1 and 9 are: x=base+dis*(sin(T1πt)+sin(T2πt)), and the fluctuation generating function of variable 7 is:

[0153] x=base-dis*(sin(T1πt)+sin(T2πt)), where base is the optimal variable benchmark value of the operating state and dis is the fluctuation amplitude. T1 and T2 are 0.02 and 0.0095 respectively, and the results of the non-optimal factor setting are shown in Table 2.

[0154] Table 2: Non-optimal factor settings in the dataset generated by the coal preparation process simulation platform

[0155]

[0156] 1) Establish an offline status evaluation model

[0157] The offline data of the heavy medium coal preparation process mechanism model were collected, and 5000 groups of data were selected. The first 2500 groups of data were divided into offline data as the training data set of the state evaluation model, and the remaining data were used as the test data set for online evaluation of the model.

[0158] Combined with expert experience, the state level of the coal preparation process is divided according to the ash value in the clean coal product. The state level division results are shown in Table 2:

[0159] Table 3: Classification of operating status and setting of level labels

[0160]

[0161] The experiment uses two layers of EvolveGCN to build the model. The dimensions of the input and output of the first layer are 11 and 17 respectively; the input and output of the second layer are set to 17 and 4. Among them, 11 is the input dimension of the experimental data, and 4 is the state dimension of the output. The number of hidden layer neurons is adjusted to 17 according to the model evaluation effect. The optimizer uses the Adam optimizer commonly used in neural network models, and the learning rate parameter is set to 0.01.

[0162] 2) Online testing of the industrial process operation status evaluation method based on SelfAtt-EvolveGCN

[0163] Figure 4 The online evaluation results of the industrial process operation status evaluation method based on SelfAtt-EvolveGCN are shown in the figure. The weighted average of precision is 0.9428, the weighted average of recall is 0.9436, and the weighted average of F1-score is 0.9431. Figure 4 It can be seen that the industrial process operation status evaluation method based on the SelfAtt-EvolveGCN network basically matches and coincides with the actual result even when the operation status fluctuates and in the unstable state transition interval. The operation status evaluation method based on the SelfAtt-EvolveGCN network can effectively identify the state level to which the heavy medium coal preparation process operation data belongs.

[0164] 3) Model evaluation results when variables are missing

[0165] In order to verify that the proposed model has certain resistance and advantages to the problem of partial missing process variables in the harsh environment of complex industrial scenarios. In the normal and disturbed parts of the data set, the values ​​of some process variables are set to 0. The specific intervals and variable settings refer to Table 4. In order to show the advantages of the model, the long short-term memory network (LSTM) with advantages in time series prediction, signal processing and other tasks; the graph convolutional network (GCN) that focuses on extracting structural feature information; and the fully connected neural network (FCN) with a simple structure and good at handling nonlinear problems are selected as comparison models. The comparison results are shown in Figure 4. Figure 4 .

[0166] Table 4 Missing variables and interval settings

[0167]

[0168] 4) Model evaluation results when samples are missing

[0169] To further verify whether the model still has advantages in anti-interference and recognition accuracy when facing sample (all variables) loss, a certain proportion of random loss is set in the normal and disturbance intervals of data, respectively. The specific intervals and proportions are shown in Table 5.

[0170] Table 5 Sample loss setting interval and proportion

[0171]

[0172] It can be seen that due to its focus on structural features, GCN has good robustness regardless of the situation where process variables are partially missing or when samples are lost. GCN uses the connection structure between nodes to propagate information and learn features. Therefore, even if the feature information of some nodes is missing, GCN can infer the features of the missing nodes through the feature information of neighboring nodes. Based on the structural characteristics of GCN, the SelfAtt-EvolveGCN model proposed in the present invention comprehensively extracts the structural features containing time series characteristics in process data, and has highly accurate recognition capabilities in the environment of missing variables and sample loss.

[0173] 5) Tracing of non-optimal factors.

[0174] When the complex industrial process operation status model based on SelfAtt-EvolveGCN judges that the industrial process status level is non-optimal, it is necessary to trace the cause variables of the non-optimal status.

[0175] Input the non-optimal data of the test set into the non-optimal cause tracing model based on EvolveGCN-AE. Figure 6The contribution rate of each operating variable to the non-optimal operating state in the 7th, 24th, 98th, 1150th, 1312th, 1449th, 2007th, and 2250th groups of samples in the test data set is shown. Figure 6 It can be seen that the main non-optimal process variable of the 7th, 24th, and 98th samples is variable 1, the main non-optimal process variable of the 1150th, 1312th, and 1449th samples is variable 7, and the main non-optimal process variable of the 2007th and 2250th samples is variable 9, which is consistent with the disturbance variables of the corresponding intervals set in Table 2;

[0176] The present invention provides a novel evaluation method to carefully judge the operating status of the heavy medium coal preparation process and promptly discover non-optimal factors. The effectiveness of the method has been proven through simulation experiments. The method can provide an important guiding basis for production adjustments in the heavy medium coal preparation process, helping production personnel to promptly discover problems and make adjustments, thereby improving production efficiency and product quality.

[0177] It should be particularly pointed out that the above-described embodiment is only one of the preferred implementations of the present invention and does not limit the protection scope of the present invention. Therefore, any simple modification and equivalent changes made to the above embodiment based on the technical essence of the present invention shall be deemed to fall within the protection scope of the present invention.

Claims

1. A complex industrial process operation status evaluation method based on SelfAtt-EvolveGCN, specifically comprising the following steps: Step 1: Use offline data sets to train the operation status evaluation model based on the SelfAtt-EvolveGCN network; A1: Collecting historical data of complex industrial processes Where N is the number of sampling points, x=[x1,x2,…,x N ] is the industrial process variable data, y=[y1,y2,…,y N ] is the selected comprehensive economic indicator data; A2: Perform standardization calculation on the x data and y data according to formula (1) and unify the dimensions so that the expected value of the calculated data is 0 and the standard deviation is 1; In the formula, X represents the standardized data, x is the input data, μ represents the mathematical expectation of the data, and σ represents the variance of the data; A3: Based on expert knowledge and comprehensive economic indicator data, classify y into state levels and label the data in turn to obtain A4: Use sliding window to N×M Perform continuous sampling, where N is the sample size and M is the process variable dimension, to obtain sequence data Where X t ∈R L×M is the sampling data of a window, T is the number of samples in the sliding window, and L represents the length of the sliding window; A5: Sample data X in a window t ∈R L×M , using the self-attention mechanism, X t Map them to three different vector spaces: Q (query), K (key), and V (value), and calculate the attention weight matrix score; Q=X t W Q +b Q (2); K=X t W K +b K (3); V=X t W V +b V (4); Where W Q , W K , W V is a learnable weight parameter, b Q 、b K 、b V is the corresponding bias term; the variance of the elements in α is d, divided by the root mean square of the scaling factor d, which ensures that the variance of the elements in α is 1; A6: After obtaining the attention weight matrix score according to the self-attention mechanism, it needs to be further converted into an adjacency matrix to reflect the connection relationship of the nodes in the graph structure; Among them, δ is the judgment threshold set according to expert knowledge. When the element in the weight matrix is ​​greater than or equal to δ, it is considered that there is a correlation between the two nodes, and the element value is set to 1; conversely, if it is less than δ, it is considered that the two nodes are independent of each other and there is no correlation, and the element will be set to 0; A7: Get a correlation graph structure representation G within a sliding window t =(X t ,A t ), and then obtain the set of association graph structures at all times in the data set, namely A8: The existence of the node itself is not considered in the constructed adjacency matrix A. The adjacency matrix is ​​standardized as follows through a self-loop method: In the formula, represents the degree matrix, represents the normalized adjacency matrix; A9: Input the network model and randomly initialize the EvolveGCN network parameters, including the connection weight W and bias b of each network layer; Where σ is the activation function, W t is the GCN weight parameter, and the initial embedding matrix comes from the node features, that is, A10: Use LSTM to iteratively update the GCN weight parameters. The specific formula is as follows: 1:Function H t =f(X t ) 2:F t =σ(M F X t +U F H t-1 +B F ) 3:I t =σ(M I X t +U I H t-1 +B I ) 4:O t =σ(M O X t +U O H t-1 +B O ) 5: 6: 7: 8: End function In the formula, the current input X of the function t With the output feature H of the previous moment t-1 Same, both are the weight parameters W of the graph convolutional network at the previous moment t-1 ; In the function formula, F, I, and O represent the update process of the three gate structures of LSTM forget gate, input gate, and output gate respectively; is the candidate state vector of the memory unit, C t Represents a memory unit; M and U represent the learnable weight parameters in each gate structure, and B represents the corresponding bias; σ and tanh are activation functions; A11: Extract the feature vector Using the Softmax classification layer, the weighted probability that the current predicted state belongs to each state is calculated (Formula 10), and finally the maximum value of the probability of each state is taken as the final model prediction state (Formula 11); In the formula It means that Softmax calculates the probability set of the divided state levels. It represents the predicted state level represented by the state with the maximum probability value, p represents the state level digital label, and C represents the total number of divided state levels; A12: Using the cross entropy loss function as the objective function of the running status offline evaluation model training process, the loss function is calculated according to formula (12); Where N represents the total number of labels in the corresponding sample set, y L is the real label data, is the label data predicted by the model; A13: Use the Adam optimization algorithm that can adaptively adjust the learning rate to iteratively optimize the model's learning parameters W and B until the loss function J is loss Minimize or converge to stability, and obtain an offline evaluation model of the operating status; Step 2: Use online data to evaluate the operation status; B1: Sample to obtain online process variable data x and y, and standardize x to obtain X; B2: Based on expert knowledge, the collected comprehensive economic indicator data y is classified into status levels and marked with digital labels in turn to obtain B3: Use a sliding window with a window length of L to sample X in sequence to obtain a sequence data set B4: For each sampled data X in the window t ∈R L×M , using the self-attention mechanism, X t Map them to three different vector spaces: Q (query), K (key), and V (value), and calculate the attention weight matrix score; B5: After obtaining the attention weight matrix score according to the self-attention mechanism, it is further converted into an adjacency matrix A that can reflect the node connection relationship in the graph structure t ; B6: Get a correlation graph structure representation G within a sliding window t =(X t ,A t ), then the set of association graph structures at all times in the data set is B7: Standardize the adjacency matrix by self-looping to obtain B8: Input the trained offline evaluation model of the running status, and the Softmax classification layer outputs the weighted probability set of the current predicted state belonging to each state The final prediction evaluation result is the operating state with the maximum probability. When the operation status evaluation model determines that the current process is in a non-optimal state, the non-optimal factors are identified. The specific steps are as follows: S1: Use high-quality data to train the EvolveGCN-AE model; S11: Sampling and selecting the process variable x with the grade of "excellent" = (x1, ..., x N ), and perform standardization on x to obtain X; S12: Use the self-attention mechanism to construct the process variable association graph structure data and standardize it, that is, the adjacency matrix A; S13: Take X and A as the input of the EvolveGCN-AE model to obtain the output reconstructed variable data The error between input and output is calculated according to formula (13), which is the sum of the errors at all times; In the formula, N is the number of samples sampled at the time; S14: Iteratively optimize the EvolveGCN-AE model, continuously update the weight terms and bias terms, minimize the loss function or stabilize the convergence, and obtain a trained non-optimal factor tracing model; S2: Use the "non-optimal" status data as the input of the traceability model to identify non-optimal factors: S21: Collect and select non-optimal process variable data and perform standardization processing; S22: Use the self-attention mechanism to calculate the node correlation and obtain the standardized adjacency matrix; S23: The standardized process variables and adjacency matrix are input into the trained EvolveGCN-AE model. According to the SPE contribution graph definition, the contribution of each process variable to the non-optimal state is calculated. The formula is as follows: In the formula, represents the i-th process variable at time t. Similarly, represents the corresponding reconstruction value; S24: Calculate the contribution of each variable and compare them. The variable with a large contribution rate is defined as the causal variable that leads to the non-optimal state.

Images