Complex industrial process running state evaluation method based on local slow feature siamese graph convolution network

Abstract

The application provides a complex industrial process running state evaluation method based on a local slow feature Siamese graph convolution network, first, an offline state evaluation model is established; then, online data is used for running state evaluation; process data X t , X t‑1 is obtained through data collection, and standardized treatment is performed; the standardized data is divided by using a sliding window, and time series with a length of W is obtained. Then, the time series data is converted into graph structure data by using a maximum information coefficient; the process data matrix and the graph structure data matrix are input into a running state evaluation model, feature extraction is performed on the potential features of the graph structure data at two adjacent time points, the potential features are constrained by using slow features, and finally, the posterior probability that the online data belongs to different running state evaluation grades is obtained; the final evaluation result is the state grade corresponding to the maximum posterior probability at the current time; and the method is helpful to realize a more accurate process running state evaluation model.

Description

Technical Field

[0001] This invention belongs to the field of industrial production process operation status evaluation technology, specifically involving a method for evaluating the operation status of complex industrial processes based on Siamese graph convolutional networks with local slow features. Background Technology

[0002] In current industrial production processes, simply maintaining the basic operation of the production system is far from sufficient to meet the demands of high efficiency and high output. Modern industrial production pursues maximum efficiency and optimal resource utilization, requiring the production process not only to be stable but also to adapt to various changes in order to maintain optimal operating conditions. However, in actual industrial production processes, due to external factors such as changes in the production environment, aging and wear of equipment, and non-standard personnel operations, the operation process often faces the risk of deviating from its optimal operating trajectory. These factors may lead to decreased production efficiency, fluctuations in product quality, and even safety accidents, thus failing to guarantee the continuous improvement of product quality and economic benefits. Therefore, timely and accurate evaluation of the operating status of industrial processes, guiding production operators to effectively control the industrial processes, is of significant practical importance for improving the operating performance of industrial processes and facilitating enterprise production management.

[0003] In the evaluation of the operational status of industrial processes, existing data-driven methods often focus on extracting features from numerical data, neglecting the hidden feature information in the spatial structural relationships between time series. Therefore, systematically mining and analyzing these relationships is crucial for achieving accurate operational status evaluation. To utilize this information more effectively, researchers have begun exploring the conversion of grid data types into graph-structured data types, where nodes can represent different monitoring variables or process components, and edges characterize the connections between these nodes. This representation makes graph convolutional networks (GCNNs) an ideal tool for analyzing such data. GCNNs can handle non-Euclidean data and have demonstrated powerful classification and feature learning capabilities in various fields. However, deep learning feature extraction methods based on GCNNs are all based on global topological node feature relationships, neglecting the significant impact of the dynamic changes in the intrinsic characteristics of local time series topological relationships on whether the entire industrial process can maintain optimal operational status. Therefore, in the evaluation of the operational status of complex industrial processes, the slow change characteristics of data and time series correlation are two factors that cannot be ignored. These characteristics indicate that the intrinsic characteristics of historical data change slowly and have local time dependence. Summary of the Invention

[0004] To address the problems of existing technologies, this invention provides a method for extracting slowly changing spatial structural features in complex industrial processes using a Siamese graph convolutional network based on local slow features. It constructs a local time-dependent graph structure model based on the complex interactions (linear, nonlinear, periodic, etc.) of the topological structure features of industrial process data, thereby improving the ability to represent the most essential features in the spatial structure that influence process changes. This method can learn the interactions between complex spatial features of local time series and extract the essential features of slowly changing process states, thus achieving refined classification of performance levels.

[0005] To achieve the above objectives, this invention provides a method for evaluating the operational status of complex industrial processes based on a Siamese graph convolutional network with local slow features. The method trains a Siamese graph convolutional network feature extraction model using offline data to establish an offline operational status evaluation model. Online data is then used for operational status evaluation. Process data Xt and Xt-1 at times t and t-1 are collected and standardized. A sliding window is used to partition the standardized data, resulting in a time series of length W. The time series data is then converted into graph-structured data using the Maximum Information Coefficient (MIC). The process data matrix and the graph-structured data matrix are input into the operational status evaluation model to extract latent features from the graph-structured data at two adjacent time points. Slow features are used to constrain the latent features, and finally, the posterior probability of the online data belonging to different operational status evaluation levels is obtained. The final evaluation result is the state level corresponding to the maximum posterior probability at the current time. This method can learn the interactions between complex spatial features of local time series and extract the essential features of slowly changing process states, thus contributing to a more accurate process operational status evaluation model.

[0006] Specifically, the following steps are included:

[0007] Step 1: Use offline data to train a local slow feature Siamese graph convolutional network feature extraction model and a classifier model to establish an offline state evaluation model;

[0008] A1: Collect raw data from industrial processes Where N is the number of sampling points, For industrial process data, For the corresponding comprehensive economic indicator data, the raw data is preprocessed, including the removal and alignment of outlier data;

[0009] A2: According to formula (1) Data and The data is standardized by min-max, resulting in a mean of 0 and a standard deviation of 1. and Representing industrial process data The maximum and minimum values ​​in;

[0010] (1);

[0011] A3: Collect two sets of data at time t and t-1, both of which have a length of... The time series dataset was used as a training dataset for a slow-feature Siamese graph convolutional network, and after max-min normalization, it was denoted as... and ,in .

[0012] A4: Based on expert knowledge, the comprehensive economic indicator data Tag data is obtained by classifying the operational status levels. ;

[0013] A5: Yes One-Hot encoding is performed to obtain the grade label data in binary representation. ;

[0014] A6: Introducing a fixed length The sliding window technique splits the standardized training dataset to obtain a sliding window set. and ,in , .

[0015] A7: Based on formula (2), use MIC to discover the interaction relationship between process data;

[0016] (2);

[0017] In the formula, It is a function of the sample size, and is usually set as . It is the length of the time series in the training dataset.

[0018] A8: As in formula (3), the interaction between time series within each sliding window can be measured using the MIC method to obtain the MIC weight matrix.

[0019] (3);

[0020] In the formula, Indicates the first Nodes in a sliding window and nodes MIC scores between sequences Representative at the Different time series data within a sliding window.

[0021] A9: Set an appropriate threshold, reset elements in the weight matrix that are less than the threshold to 0, and those that are greater than the threshold to 1, thus obtaining an adjacency matrix that reflects the connection relationships between nodes, and construct a series of association graphs. Its calculation formula is shown in formula (4):

[0022] (4);

[0023] In the formula, Representing the adjacency matrix elements and This represents the score threshold for node connectivity determined through expert knowledge.

[0024] A10: The graph structured dataset obtained from the training data. ,in Representing the The graph structure data under each input sample. The set of adjacency matrices in the model input information can be represented as... .

[0025] A11: Initialize the parameters of the slow feature Siamese graph convolutional network, perform forward propagation to discover the intrinsic relationships between nodes in the association graph, and obtain the hidden layer features. and And calculate The optimal model is obtained through reverse iterative optimization;

[0026] A12: Use the MIC weight matrix as the initialization parameters for the slow-feature Siamese graph convolutional network to shorten the network's training time, and input sliding window data. and adjacency matrix Afterwards, the network forward propagates to obtain the output value, and combines it with comprehensive economic indicator information to back-train the slow feature Siamese graph convolutional network until the objective function converges.

[0027] A13: Utilizing grade label data Supervised learning is introduced into the Siamese graph convolutional network for slow features to obtain reconstructed values ​​of the process data. At the same time, the predicted values ​​of comprehensive economic indicators are obtained. ;

[0028] A14: Calculate the loss function of the feature extraction model according to formula (5). :

[0029] (5);

[0030] A15: After initializing the parameter matrix of the Softmax classifier, optimize the classifier model until the objective function is minimized or converges;

[0031] A16: Output the hidden layer of the feature extraction model and binary variables As input, a Softmax classifier is trained to obtain a state recognition model. The output of the Softmax classifier is obtained according to formula (6):

[0032] (6);

[0033] In the formula, Indicates different state levels, Indicates input Status level The posterior probability, These are the parameters of the classifier;

[0034] A17: Calculate the classifier loss function according to formula (7). :

[0035] (7);

[0036] A18: Reverse fine-tuning of the classifier model parameters until minimized. Or achieve The convergence of the state recognition model is obtained.

[0037] A19: Fine-tune the parameters after cascading the feature extraction model and classifier to build a complete offline model for evaluating operational status.

[0038] Step 2: Use online data to evaluate operational status;

[0039] B1: Online process data obtained through sampling and to Standardization processing is required;

[0040] B2: For online data, the window length is... Sliding sampling using a sliding window yields sequence data of consistent length;

[0041] B3: Input the sequence data into the trained runtime evaluation model to calculate... Real-time online data The posterior probability of belonging to different operating state levels is ,in, , The final numerical label representation of the status level categories based on comprehensive economic indicator data. The process state at a given time is defined as the posterior probability set. The state level corresponding to the maximum value, i.e. The process operation status level at time is ;

[0042] B4: Based on the online operation status evaluation results, guide the subsequent optimization and control of industrial processes.

[0043] The beneficial effects of this invention are as follows: It provides a method for evaluating the operational status of complex industrial processes based on Siamese graph convolutional networks with local slow features. This method transforms the evaluation problem of complex industrial process operational status into a graph classification problem, fully considering the interaction of multi-class node features in time-series data. The slow-feature Siamese graph convolutional network learns the interaction information between spatial nodes in time-series data through MIC (Multi-Minute Interaction), and introduces Siamese graph convolutional network blocks to extract interaction-based subgraph representations. Slow-feature constraints are applied to the extracted latent feature information, allowing the model to retain slowly changing feature representations to reflect operational performance changes. By embedding the local slow-feature algorithm into the Siamese graph convolutional network, graph-structured data at two consecutive time points can be processed simultaneously. Furthermore, the parameters of the entire network loss function can be adjusted through the Siamese graph convolutional network, enabling a higher-dimensional representation of node features after the transformation of locally time-dependent process data. Finally, the offline process data is classified into performance levels based on the comprehensive economic indicators of the industrial process. The classification results are combined with the hidden features learned by the feature extraction model to train a state recognition model, resulting in a complete process operational status evaluation model. This method can evaluate the industrial operating status in a timely and accurate manner, guide the optimization and control of processes, thereby improving the operating performance of industrial processes, effectively ensuring the quality of industrial products, and improving the overall economic benefits of enterprises. Attached Figure Description

[0044] Figure 1 A flowchart illustrating the overall evaluation of industrial process operation status in this invention;

[0045] Figure 2 This is a simplified flow chart of coal preparation based on coal slime flotation process in this invention;

[0046] Figure 3 This is a schematic diagram of the running state evaluation structure of the Siamese graph convolutional network model based on slow features in this invention;

[0047] Figure 4 This refers to the mutual information value of variables based on MIC in this invention;

[0048] Figure 5 This is the online evaluation and comparison result of the running state method based on the slow feature Siamese graph convolutional network in this invention. Detailed Implementation

[0049] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is not intended to limit the present invention or its application or use in any way. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0050] Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification.

[0051] like Figure 1 As shown, this invention provides a method for evaluating the operational status of complex industrial processes based on a Siamese graph convolutional network with local slow features. The Siamese graph convolutional network with slow features is a method for evaluating operational status based on local slow feature analysis using MIC graph structure learning.

[0052] The method specifically includes the following steps:

[0053] Step 1: Use offline data to train a slow feature Siamese graph convolutional network feature extraction model and a classifier model to establish an offline state evaluation model;

[0054] A1: Collect raw data from industrial processes Where N is the number of sampling points, For industrial process data, For the corresponding comprehensive economic indicator data, the raw data is preprocessed, including the removal and alignment of outlier data;

[0055] A2: According to formula (1) Data and The data is standardized by min-max, resulting in a mean of 0 and a standard deviation of 1. and Representing industrial process data The maximum and minimum values ​​in;

[0056] (1);

[0057] A3: Collect two sets of data at time t and t-1, both of which have a length of... The time series dataset was used as the training dataset for the slow feature Siamese graph convolutional network model. After min-max normalization, it was denoted as... and ,in .

[0058] A4: Based on expert knowledge, the comprehensive economic indicator data Tag data is obtained by classifying the operational status levels. ;

[0059] A5: Yes One-Hot encoding is performed to obtain the grade label data in binary representation. ;

[0060] A6: Introducing a fixed length The sliding window technique splits the standardized training dataset to obtain a sliding window set. and ,in , .

[0061] A7: Figure 3 The structure of the slow feature Siamese graph convolutional network is described. The slow feature Siamese graph convolutional network is constructed based on the maximum information coefficient. By embedding the local slow feature algorithm into the Siamese graph convolutional network, it learns the interaction relationship between complex spatial features of local time series and extracts the essential features of the process state with slow changes.

[0062] First, based on formula (2), the interaction relationship between process data is discovered using MIC;

[0063] (2);

[0064] In the formula, It is a function of the sample size, and is usually set as . It is the length of the time series in the training dataset.

[0065] A8: As in formula (3), the interaction between time series within each sliding window can be measured using the MIC method to obtain the MIC weight matrix.

[0066] (3);

[0067] In the formula, Indicates the first Nodes in a sliding window and nodes MIC scores between sequences Representative at the Different time series data within a sliding window.

[0068] A9: Set an appropriate threshold, reset elements in the weight matrix that are less than the threshold to 0, and those that are greater than the threshold to 1, thus obtaining an adjacency matrix that reflects the connection relationships between nodes, and construct a series of association graphs. Its calculation formula is shown in formula (4):

[0069] (4);

[0070] In the formula, Representing the adjacency matrix elements and This represents the score threshold for node connectivity determined through expert knowledge.

[0071] A10: The graph structured dataset obtained from the training data. ,in Representing the The graph structure data under each input sample. The set of adjacency matrices in the model input information can be represented as... .

[0072] A11: Input graph structure data Entering the slow-feature Siamese graph convolutional network model, we obtain the sub-network values ​​of each sub-network. and Implicit characteristics of time and The calculation formula is shown in formula (5):

[0073] (5);

[0074] In the formula, It is an adjacency matrix containing self-loops. The representative matrix element is The degree matrix.

[0075] A12: Initialize the parameters of the slow feature Siamese graph convolutional network, perform forward propagation to discover the intrinsic relationships between nodes in the association graph, and obtain the hidden layer features. and And calculate The optimal model is obtained through reverse iterative optimization; after mapping processing by a fully connected layer, the output of the feature extraction model is:

[0076] (6);

[0077] In the formula, This represents the activation function used to improve the nonlinear expressive power of fully connected layers. The representative length is The deviation vector and The dimension is The weight matrix.

[0078] A13: Use the MIC weight matrix as the initialization parameters for the slow-feature Siamese graph convolutional network to shorten the network's training time, and input sliding window data. and adjacency matrix Afterwards, the network forward propagates to obtain the output value, and combines it with comprehensive economic indicator information to back-train the slow feature Siamese graph convolutional network until the objective function converges.

[0079] A14: Utilizing grade label data Supervised learning is introduced into the Siamese graph convolutional network for slow features to obtain reconstructed values ​​of the process data. At the same time, the predicted values ​​of comprehensive economic indicators are obtained. ;

[0080] A15: Calculate the loss function of the feature extraction model according to formula (7). :

[0081] (7);

[0082] A16: After initializing the parameter matrix of the Softmax classifier, optimize the classifier model until the objective function is minimized or converges;

[0083] A17: Output the hidden layer of the feature extraction model and binary variables As input, a Softmax classifier is trained to obtain a state recognition model. The output of the Softmax classifier is obtained according to formula (8):

[0084] (8);

[0085] In the formula, Indicates different state levels, Indicates input Status level The posterior probability, These are the parameters of the classifier;

[0086] A18: Calculate the classifier loss function according to formula (9) :

[0087] (9);

[0088] A19: Reverse fine-tuning of the classifier model parameters until minimized. Or achieve The convergence of the state recognition model is obtained.

[0089] A20: Fine-tune the parameters after cascading the feature extraction model and classifier to build a complete offline model for evaluating operational status.

[0090] Step 2: Use online data to evaluate operational status;

[0091] B1: Online process data obtained through sampling and to Standardization processing is required;

[0092] B2: For online data, the window length is... Sliding sampling using a sliding window yields sequence data of consistent length;

[0093] B3: Input the sequence data into the trained runtime evaluation model to calculate the online data at time t. The posterior probability of belonging to different operating state levels is ,in, , The state level categories are represented by numerical labels based on comprehensive economic indicator data. The final process operating state at time t is defined as the posterior probability set. The state level corresponding to the maximum value, that is, the process running state level at time t, is: ;

[0094] B4: Based on the online operation status evaluation results, guide the subsequent optimization and control of industrial processes.

[0095] This invention provides a method for evaluating the operational status of complex industrial processes based on Siamese graph convolutional networks with local slow features. This method transforms the evaluation of complex industrial process operational status into a graph classification problem, fully considering the interaction of features among multiple nodes in time-series data. The slow-feature Siamese graph convolutional network learns the interaction information between spatial nodes in time-series data through MIC (Minimum Interaction Model), and introduces Siamese graph convolutional network blocks to extract interaction-based subgraph representations. Slow-feature constraints are applied to the extracted latent feature information, ensuring the model retains slowly changing feature representations to reflect operational performance changes. By embedding the local slow-feature algorithm into the Siamese graph convolutional network, graph-structured data at two consecutive time points can be processed simultaneously. Furthermore, the parameters of the entire network loss function can be adjusted through the Siamese graph convolutional network, enabling a higher-dimensional representation of node features after the transformation of locally time-dependent process data. Finally, the offline process data is classified into performance levels based on comprehensive economic indicators of the industrial process. The classification results are combined with the hidden features learned by the feature extraction model to train a state recognition model, resulting in a complete process operational status evaluation model. This method can evaluate the industrial operating status in a timely and accurate manner, guide the optimization and control of processes, thereby improving the operating performance of industrial processes, effectively ensuring the quality of industrial products, and improving the overall economic benefits of enterprises.

[0096] The technical solution of the present invention will be described in detail below with reference to the embodiments, and its feasibility will be verified.

[0097] Example:

[0098] Coal slime flotation is a typical process industry process operating in a harsh, open environment. The operating environment is subject to various disturbances and uncertainties, often resulting in incomplete perception of operating conditions by current process monitoring and operational status evaluation methods. This invention addresses the highly dynamic and nonlinear nature of data in the coal slime flotation process, enabling better mining of various interrelationships and interactions, and providing a more accurate and robust evaluation of the current operating status.

[0099] A typical coal slime flotation process begins with crushing and screening to adjust the raw coal to an appropriate size. This raw coal is then thoroughly mixed with a heavy medium suspension transported from a qualified medium tank in a mixing tank. A pressure pump injects the mixture into a heavy medium hydrocyclone for initial separation. Next, the screened clean coal is transported to a coal slime tank. The finer particles in the clean coal are then conveyed to the bottom of a thickener for thickening, adjusting the density to a suitable level. After the thickener completes its work, the coal slime slurry is discharged from the bottom of the thickener and thoroughly mixed with flotation reagents in a slurry pre-processor before entering the flotation cell for froth flotation. Finally, a scraper device removes the overflow from the flotation cell. After dewatering and drying, the remaining clean coal, gangue, and other impurities are discharged as tailings at the bottom of the flotation cell.

[0100] The complex and redundant variables involved in the entire coal slime flotation process, including some variables that are irrelevant or have very little correlation with operational performance evaluation indicators, increase the computational burden. The mutual information value of each candidate input variable is calculated using the MIC method described above, such as... Figure 4 As shown in Table 1, this paper selects measurable variables whose calculated results are greater than the average value as the raw data components of the slow feature Siamese graph convolutional network model. These 20 typical process variables are listed below.

[0101] Table 1: Selection of Process Operation Variables

[0102]

[0103] The data in this example comes from a full-process simulation platform for coal slime flotation (Registration No.: 2020SR1160052). To generate coal slime flotation process data at different state levels, deviations are introduced into the process operation variables in the simulation platform to simulate the production site, causing the production process to deviate from the assumed optimal operating point to reflect different operating states. The disturbance generation function for variable 1 (raw coal washing amount) is: The deviation ramp signals for variables 37 (thickener underflow rate) and 44 (flotation cell stirring speed) are set as follows: ,in This represents the optimal baseline value for the operating state. Indicates the magnitude of the disturbance. , and Represents a constant. This indicates the slope of the signal.

[0104] 1) Establish an offline status evaluation model

[0105] After removing unstable data caused by time delays, 14,000 samples were collected from the full-process simulation platform of coal slime flotation and retained as the experimental dataset. In this experiment, 7,000 datasets were selected as the offline training dataset for the evaluation model, and the remaining data were used for online testing.

[0106] Based on expert experience, the state levels of the coal preparation process are classified according to the ash content of the flotation cell overflow in the clean coal product. The results of the state level classification are shown in Table 2.

[0107] Table 2: Classification of Operating Status Levels and Setting of Level Labels

[0108]

[0109] Based on the characteristics of the coal slime flotation process, the local SFA parameters of the slow feature Siamese graph convolutional network were optimized in real time during training and testing. The topology graph dimension of the local slow features was set to 50*50, and the stride was also set to 10, which achieved a balance between capturing fine information about the process and avoiding noise interference. Furthermore, the outputs of the two convolutional layers under the Siamese graph convolutional network module were 11 and 4, respectively, and six sets of slowly changing features were retained as the SFA dimension after MSV calculation. The learning rate for training the slow feature Siamese graph convolutional network was set to 0.001, and the epoch was set to 200. To optimize the network parameters, the adaptive point estimation (Adam) stochastic optimization algorithm was used to minimize the cross-entropy loss function, and an offline operational status evaluation model was established.

[0110] 2) Online testing of an industrial process operation status evaluation method based on Siamese graph convolutional networks with local slow features.

[0111] Figure 5 The online evaluation results for an industrial process operation status assessment method based on a Siamese graph convolutional network with local slow features are presented. The weighted average precision is calculated to be 94.96%, and the recall rates for the four levels are 99.01%, 88.35%, 97.26%, and 96.43%, respectively. Further combined with... Figure 5 It can be seen that the evaluation method for the operational status of industrial processes based on Siamese graph convolutional networks with local slow features shows a high degree of overlap between the evaluation results and the actual operational status. This means that the evaluation method can effectively identify the operational status level of the coal slime flotation process data. Furthermore, the evaluation method based on Siamese graph convolutional networks with local slow features demonstrates more stable probability identification of operational data belonging to the corresponding operational status level. Therefore, this invention exhibits higher robustness in identifying each operational status level.

[0112] The simulation examples above demonstrate the effectiveness of this invention. Timely and robust operational status evaluation of the coal slime flotation process provides crucial guidance for production adjustments.

[0113] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications and equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for evaluating the operational status of complex industrial processes based on Siamese graph convolutional networks with local slow features, characterized in that, Specifically, the following steps are included: Step 1: Use offline data to train a local slow feature Siamese graph convolutional network feature extraction model and a classifier model to establish an offline state evaluation model; A1: Collect raw data from industrial processes Where N is the number of sampling points, For industrial process data, For the corresponding comprehensive economic indicator data, the raw data is preprocessed, including the removal and alignment of outlier data; A2: According to formula (1) Data and The data is subjected to min-max standardization so that the mean of the calculated data is 0 and the standard deviation is 1. （1）； A3: Collect two sets of data at time t and t-1, both of which have a length of... The time series dataset is used as a training dataset for a slow-feature Siamese graph convolutional network. After max-min normalization, it is denoted as... and ,in ; A4: Based on expert knowledge, the comprehensive economic indicator data Tag data is obtained by classifying the operational status levels. ; A5: Yes One-Hot encoding is performed to obtain the grade label data in binary representation. ; A6: Introducing a fixed length The sliding window technique splits the standardized training dataset to obtain a sliding window set. and ,in , ; A7: Based on formula (2), use MIC to discover the interaction relationship between process data; （2）； In the formula, It is a function of the sample size, and is usually set as ; It is the length of the time series in the training dataset; A8: As shown in formula (3), the interaction between time series within each sliding window can be measured using the MIC method to obtain the MIC weight matrix; （3）； In the formula, Indicates the first Nodes in a sliding window and nodes MIC scores between sequences Representative at the Different time series data within a sliding window; A9: Set an appropriate threshold, reset elements in the weight matrix that are less than the threshold to 0, and those that are greater than the threshold to 1, thus obtaining an adjacency matrix that reflects the connection relationships between nodes, and construct a series of association graphs. The calculation formula is shown in formula (4): （4）； In the formula, Representing the adjacency matrix elements and The threshold score represents the node connection relationship determined by expert knowledge. A10: The graph structured dataset obtained from the training data. ,in Representing the The graph structure data under each input sample; the set of adjacency matrices in the model input information can be represented as... ; A11: Initialize the parameters of the slow feature Siamese graph convolutional network, perform forward propagation to discover the intrinsic relationships between nodes in the association graph, and obtain the hidden layer features. and And calculate The optimal model is obtained through reverse iterative optimization; A12: Use the MIC weight matrix as the initialization parameters for the slow-feature Siamese graph convolutional network to shorten the network's training time, and input sliding window data. and adjacency matrix Afterwards, the network forward propagates to obtain the output value, and combines it with comprehensive economic indicator information to back-train the slow feature Siamese graph convolutional network until the objective function converges. A13: Utilizing grade label data Supervised learning is introduced into the Siamese graph convolutional network for slow features to obtain reconstructed values ​​of the process data. At the same time, the predicted values ​​of comprehensive economic indicators are obtained. ; A14: Select the cross-entropy loss function as the optimization objective function for the network training process, and calculate the loss function of the feature extraction model according to formula (5). : （5）； A15: After initializing the parameter matrix of the Softmax classifier, optimize the classifier model until the objective function is minimized or converges; A16: Output the hidden layer of the feature extraction model and binary variables As input, a Softmax classifier is trained to obtain a state recognition model. The output of the Softmax classifier is obtained according to formula (6): （6）； In the formula, Indicates different state levels, Indicates input Status level The posterior probability, These are the parameters of the classifier; A17: Calculate the classifier loss function according to formula (7). : （7）； A18: Reverse fine-tuning of the classifier model parameters until minimized. Or achieve The convergence of the state recognition model is obtained. A19: Fine-tune the parameters after cascading the feature extraction model and classifier to build a complete offline model for evaluating the running status; Step 2: Use online data to evaluate operational status; B1: Online process data obtained through sampling and to Standardize the process; B2: For online data, the window length is... Sliding sampling using a sliding window yields sequence data of consistent length; B3: Input the sequence data into the trained runtime evaluation model to calculate... Real-time online data The posterior probability of belonging to different operating state levels is ,in, , The final numerical label representation of the status level categories based on comprehensive economic indicator data. The process state at a given time is defined as the posterior probability set. The state level corresponding to the maximum value, i.e. The process operation status level at time is ; B4: Based on the online operation status evaluation results, guide the subsequent optimization and control of industrial processes.

Images