Complex industrial process operation state evaluation method based on local slow feature Siamese graph convolutional network

Abstract

The invention provides a complex industrial process operation state evaluation method based on a local slow feature Siamese graph convolutional network. The method comprises the following steps: firstly, establishing a state evaluation offline model; then carrying out operation state evaluation by using online data; process data Xt and Xt-1 are obtained through data collection, and standardization processing is carried out on the process data Xt and Xt-1; and performing data division on the standardized data by using a sliding window to obtain a time sequence with the length of W. Converting the time sequence data into graph structure data through the maximum information coefficient; inputting the process data matrix and the graph structure data matrix into an operation state evaluation model, performing feature extraction on potential features of the graph structure data at two adjacent moments, constraining the potential features by using slow features, and finally obtaining posterior probabilities of the online data belonging to different operation state evaluation grades, the final evaluation result is the state level corresponding to the maximum posterior probability at the current moment; according to the method, a more accurate process operation state evaluation model can be realized.

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 a local slow feature Siamese graph convolutional network. Background Art

[0002] In the current industrial production process, maintaining the basic operation of the production system is far from meeting the manufacturing needs of high efficiency and high output. Modern industrial production pursues maximum efficiency and optimal resource utilization, which requires that the production process should not only be stable, but also be able to adapt to various changes to maintain the best operating state. However, in the actual industrial production process, due to the influence of external factors such as changes in the production environment, aging and wear of equipment, and irregular operation of personnel, the operation process is often in danger of deviating from the optimal operation track. These factors may lead to a decrease in production efficiency, fluctuations in product quality, and even safety accidents, thus failing to ensure the continuous improvement of product quality and economic benefits. Therefore, it is of great practical significance to implement timely and accurate evaluation of the operating status of the industrial process and guide production operators to effectively regulate the industrial process, improve the operating performance of the industrial process, and facilitate the production management of the enterprise.

[0003] In the evaluation of the operating status of industrial processes, existing data-driven methods often focus on extracting features from numerical data, but ignore the feature information hidden in the spatial structural relationship between time series. Therefore, systematic mining and analysis of these relationships is crucial to achieve accurate operating status evaluation. In order to make more effective use of this information, researchers began to explore the conversion of grid data types into graph structure data types, where nodes can represent different monitoring variables or process components, and edges represent the connection relationship between these nodes. This representation makes graph convolutional networks an ideal tool for analyzing such data. Graph convolutional networks can process data with non-Euclidean structures and have demonstrated their powerful classification and feature learning capabilities in many fields. However, deep learning feature extraction methods based on graph convolutional networks are all based on the global topological structure node feature relationship, ignoring the fact that the dynamic changes in the intrinsic essential characteristics of the topological relationship of the local time series have an important impact on whether the entire industrial process can always maintain the optimal operating state. Therefore, in the evaluation of the operating status of complex industrial processes, the slow changing characteristics of data and the correlation of time series are two factors that cannot be ignored. These characteristics indicate that the intrinsic essential characteristics in historical data change slowly and have local time dependence. Summary of the invention

[0004] In view of the problems existing in the above-mentioned prior art, the present invention provides a method based on a local slow feature Siamese graph convolutional network to extract slowly changing spatial structural features in complex industrial processes. In view of the complex interactive relationships of the topological structural features of industrial process data, such as linearity, nonlinearity, periodicity, etc., a local time-dependent graph structure model is constructed to improve the ability to express the most essential part of the features that affect process changes in the spatial structure. This method can learn the interactive relationships between the complex spatial features of the local time series, and extract the essential characteristics of the process state that have slow changes, thereby achieving a refined division of performance levels.

[0005] In order to achieve the above object, the present invention provides a method for evaluating the operation status of a complex industrial process based on a local slow feature Siamese graph convolutional network. The method uses offline data to train a slow feature Siamese graph convolutional network feature extraction model to establish an offline model for state evaluation. The online data is used to evaluate the operation status. The process data X at time t and t-1 obtained by data collection are t , X t-1 , and standardize it; use the sliding window to divide the standardized data to obtain a time series of length W. Then, the time series data is converted into graph structure data through the maximum information coefficient, namely Maximum Information Coefficient, referred to as MIC; the process data matrix and the graph structure data matrix are input into the operation status evaluation model, and the potential features of the graph structure data at two adjacent moments are extracted. At the same time, the slow features are used to constrain the potential features, and finally the posterior probability of the online data belonging to different operation status evaluation levels is obtained. The final evaluation result is the state level corresponding to the maximum posterior probability at the current moment; this method can learn the interaction between the complex spatial features of the local time series and extract the essential characteristics of the process state with slow changes, which helps to realize a more accurate process operation status evaluation model.

[0006] The specific steps include:

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

[0008] A1: Collecting raw data from industrial processes Where N is the number of sampling points, X = [x1, x2, ..., x N ] is the industrial process data, Y = [y1, y2, ..., y N ] is the corresponding comprehensive economic indicator data, and the original data is preprocessed, including the removal and alignment of abnormal data;

[0009] A2: Perform maximum and minimum standardization processing on the X data and Y data according to formula (1) so that the mean of the calculated data is 0 and the standard deviation is 1;

[0010]

[0011] A3: Collect two sets of time series data sets with length n at time t and time t-1 as the slow feature Siamese graph convolution network training data set, and record them as and Where s=t or t-1.

[0012] A4: Classify the operating status of the comprehensive economic indicator data Y according to expert knowledge to obtain label data

[0013] A5: Yes L Perform One-Hot encoding to obtain binary representation of level label data

[0014] A6: Introduce a sliding window technology with a fixed length of w to split the standardized training data set to obtain a sliding window set and Where w≤N,

[0015] A7: According to formula (2), MIC is used to discover the interaction relationship between process data;

[0016]

[0017] Where D is a function of the sample size, usually set to M 0.6 . M is the length of the time series in the training dataset.

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

[0019]

[0020] In the formula, represents the MIC score between the sequences of node i and node j in the kth sliding window, Represents different time series data within the kth sliding window.

[0021] A9: Set a suitable threshold, reset the elements in the weight matrix less than the threshold to 0, and reset them to 1 if they are greater than the threshold, so as to obtain an adjacency matrix that can reflect the node connection relationship and construct a series of association graphs g k; Its calculation formula is shown in formula (4):

[0022]

[0023] In the formula, Represents the adjacency matrix and ζ represents the node connection relationship score threshold determined by expert knowledge.

[0024] A10: Calculate the graph structure data set of training data in Represents the graph structure data under the kth input sample. The adjacency matrix set in the model input information can be expressed as

[0025] A11: Initialize the parameters of the slow feature Siamese graph convolutional network, forward propagate to find the intrinsic connection between nodes in the association graph, and obtain the hidden layer features and And calculate The optimal model obtained by reverse iterative optimization;

[0026] A12: Use the MIC weight matrix as the initialization parameter of the slow feature Siamese graph convolutional network to shorten the network training time. Input the sliding window data X w And the adjacency matrix A w Afterwards, the network forward propagates to obtain the output value, and combines the comprehensive economic indicator information to reversely train the slow feature Siamese graph convolutional network until the objective function converges;

[0027] A13: Using Class Label Data Introducing supervised learning into the slow feature Siamese graph convolutional network to obtain the reconstruction value of the process data At the same time, we can get the forecast value of comprehensive economic indicators

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

[0029]

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

[0031] A16: The hidden layer output h of the feature extraction model and the binary variable As input, train the Soffmax classifier to obtain the state recognition model, and obtain the output of the Softmax classifier according to formula (6):

[0032]

[0033] In the formula, c = 1, 2, ..., c represents different state levels, p(y = j|x) represents the posterior probability that the input x is state level j, is the parameter of the classifier;

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

[0035]

[0036] A18: Backward fine-tuning of the classifier model parameters until J is minimized softmax (θ) or realize J softmax The convergence of (θ) leads to the state recognition model;

[0037] A19: Cascade feature extraction models and classifiers and fine-tune parameters to build a complete offline model for running status evaluation.

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

[0039] B1: Sample and obtain online process data X, and perform standardization on X;

[0040] B2: Sliding sampling is performed on the online data using a sliding window with a window length of w to obtain sequence data with consistent length;

[0041] B3: Input the sequence data into the trained operation status evaluation model and calculate the online data x at time t t The posterior probability of belonging to different operating status levels is Among them, i∈{1, 2, ..., q}, q is the digital label representation of the state level category divided according to the comprehensive economic indicator data, and the final process operation state at time t is defined as the posterior probability set The state level corresponding to the maximum value in, that is, the process operation state level at time t is

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

[0043] The beneficial effect of the invention is: a method for evaluating the operating status of a complex industrial process based on a local slow feature Siamese graph convolutional network is provided, which transforms the problem of evaluating the operating status of a complex industrial process into a graph classification problem, and fully considers the interaction of multiple node features of time series data. The slow feature Siamese graph convolutional network learns the amount of interactive information between nodes in the time series data space through MIC, and introduces the Siamese graph convolutional network block to extract the subgraph representation based on the interaction. And the extracted potential feature information is subjected to slow feature constraints so that the model retains the feature representation that changes slowly to reflect the characteristics of the operating performance change. By embedding the local slow feature algorithm into the Siamese graph convolutional network, the graph structure data at two consecutive moments can be processed simultaneously, and the parameters of the entire network loss function can be adjusted through the Siamese graph convolutional network, and the node features after the local time-dependent process data is converted into a higher-dimensional representation. Finally, the process offline data is divided into performance levels according to the comprehensive economic indicators of the industrial process, and the state recognition model is trained using the division results combined with the hidden features learned by the feature extraction model to obtain a complete process operation status evaluation model. This method can timely and accurately evaluate the industrial operation status and guide the optimization and regulation of the process, thereby improving the operating performance of the industrial process, effectively ensuring the quality of industrial products, and improving the comprehensive economic benefits of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention is further described below with reference to the accompanying drawings and embodiments.

[0045] Figure 1 The overall flow chart of the industrial process operation status evaluation in the present invention;

[0046] Figure 2 It is a simplified flow chart of coal preparation based on the coal slime flotation process in the present invention;

[0047] Figure 3 It is a schematic diagram of the operating status evaluation structure based on the slow feature Siamese graph convolutional network model in the present invention;

[0048] Figure 4 is the variable mutual information value based on MIC in the present invention;

[0049] Figure 5 It is the online evaluation comparison result of the operating status method based on the slow feature Siamese graph convolutional network in the present invention; DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings of the specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and does not serve as any limitation on the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without carrying out creative work are within the scope of protection of the present invention.

[0051] Technologies, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.

[0052] like Figure 1 As shown, the present invention provides a complex industrial process operation status evaluation method based on a local slow feature Siamese graph convolutional network, and the slow feature Siamese graph convolutional network is a local slow feature analysis operation status evaluation method based on MIC graph structure learning.

[0053] The method specifically comprises the following steps:

[0054] Step 1: Use offline data to train the slow-feature Siamese graph convolutional network feature extraction model and classifier model to establish an offline state evaluation model;

[0055] A1: Collecting raw data from industrial processes Where N is the number of sampling points, X = [x1, x2, ..., x N ] is the industrial process data, Y = [y1, y2, ..., y N ] is the corresponding comprehensive economic indicator data, and the original data is preprocessed, including the removal and alignment of abnormal data;

[0056] A2: Perform maximum and minimum standardization processing on the X data and Y data according to formula (1) so that the mean of the calculated data is 0 and the standard deviation is 1;

[0057]

[0058] A3: Collect two sets of time series data sets with length n at time t and time t-1 as the training data set of the slow feature Siamese graph convolutional network model, and record them as and Where s=t or t-1.

[0059] A4: Classify the operating status of the comprehensive economic indicator data Y according to expert knowledge to obtain label data

[0060] A5: Yes L Perform One-Hot encoding to obtain binary representation of level label data

[0061] A6: Introduce a sliding window technology with a fixed length of w to split the standardized training data set to obtain a sliding window set and Where w≤N,

[0062] A7: Figure 3 The structure of the slow feature Siamese graph convolutional network is shown in Figure 1. The slow feature Siamese graph convolutional network constructs the graph structure based on the maximum information coefficient. By embedding the local slow feature algorithm into the Siamese graph convolutional network, it learns the interactive relationship between the complex spatial features of the local time series and extracts the essential characteristics of the process state that has slow changes.

[0063] First, according to formula (2), MIC is used to discover the interactive relationship between process data;

[0064]

[0065] Where D is a function of the sample size, usually set to M 0.6 . M is the length of the time series in the training dataset.

[0066] A8: As shown in formula (3), the interaction relationship between time series in each sliding window can be measured by MIC method to obtain the MIC weight matrix.

[0067]

[0068] In the formula, represents the MIC score between the sequences of node i and node j in the kth sliding window, Represents different time series data within the kth sliding window.

[0069] A9: Set a suitable threshold, reset the elements in the weight matrix less than the threshold to 0, and reset them to 1 if they are greater than the threshold, so as to obtain an adjacency matrix that can reflect the node connection relationship and construct a series of association graphs g k ; Its calculation formula is shown in formula (4):

[0070]

[0071] In the formula, Represents the adjacency matrix and ζ represents the node connection relationship score threshold determined by expert knowledge.

[0072] A10: Calculate the graph structure data set of training data in Represents the graph structure data under the kth input sample. The adjacency matrix set in the model input information can be expressed as

[0073] A11: Input graph structure data Enter the slow feature Siamese graph convolutional network model and obtain the implicit features at time t and t-1 in each sub-network and The calculation formula is shown in formula (5):

[0074]

[0075] In the formula, is an adjacency matrix containing self-loops, Represents the matrix elements as The degree matrix of .

[0076] A12: Initialize the parameters of the slow feature Siamese graph convolutional network, forward propagate to find the intrinsic connection between nodes in the association graph, and obtain the hidden layer features and And calculate The optimal model obtained by reverse iterative optimization; after the mapping process of the fully connected layer, the output of the feature extraction model is:

[0077]

[0078] Where σ(·) represents the activation function used to improve the nonlinear expression ability of the fully connected layer, b represents the bias vector of length l and W represents the weight matrix of dimension (m, n).

[0079] A13: Use the MIC weight matrix as the initialization parameter of the slow feature Siamese graph convolutional network to shorten the network training time and input the sliding window data X w And the adjacency matrix A w Afterwards, the network forward propagates to obtain the output value, and combines the comprehensive economic indicator information to reversely train the slow feature Siamese graph convolutional network until the objective function converges;

[0080] A14: Using Class Label Data Introducing supervised learning into the slow feature Siamese graph convolutional network to obtain the reconstruction value of the process data At the same time, we can get the forecast value of comprehensive economic indicators

[0081] A15: Calculate the loss function J of the feature extraction model according to formula (7) GCN :

[0082]

[0083] A16: After initializing the parameter matrix of the Softmax classifier, optimize the classifier model until the objective function is minimized or converged;

[0084] A17: The hidden layer output h of the feature extraction model and the binary variable As input, train the Softmax classifier to obtain the state recognition model, and obtain the output of the Softmax classifier according to formula (8):

[0085]

[0086] In the formula, c = 1, 2, ..., C represents different state levels, p(y = j|x) represents the posterior probability that the input x is state level j, is the parameter of the classifier;

[0087] A18: Calculate the classifier loss function J according to formula (9) softmax (θ):

[0088]

[0089] A19: Backward fine-tuning of the classifier model parameters until J is minimized softmax (θ) or realize J softmax The convergence of (θ) leads to the state recognition model;

[0090] A20: Cascade feature extraction model and classifier and fine-tune parameters to build a complete offline model for running status evaluation.

[0091] Step 2: Use online data to evaluate the operation status;

[0092] B1: Sample and obtain online process data X, and perform standardization on X;

[0093] B2: Sliding sampling is performed on the online data using a sliding window with a window length of w to obtain sequence data with consistent length;

[0094] B3: Input the sequence data into the trained operation status evaluation model and calculate the online data x at time t t The posterior probability of belonging to different operating status levels is Among them, i∈{1, 2, ..., q}, q is the digital label representation of the state level category divided according to the comprehensive economic indicator data, and the final process operation state at time t is defined as the posterior probability set The state level corresponding to the maximum value in, that is, the process operation state level at time t is

[0095] B4: Based on the online operation status evaluation results, guide the subsequent optimization and regulation of the industrial process.

[0096] The present invention provides a complex industrial process operation status evaluation method based on a local slow feature Siamese graph convolutional network. The method transforms the complex industrial process operation status evaluation problem into a graph classification problem, and fully considers the interaction of multiple node features of time series data. The slow feature Siamese graph convolutional network learns the amount of interactive information between nodes in the time series data space through MIC, and introduces the Siamese graph convolutional network block to extract the interaction-based subgraph representation. And the extracted potential feature information is subjected to slow feature constraints so that the model retains the slowly changing feature representation to reflect the characteristics of the operating performance change. By embedding the local slow feature algorithm into the Siamese graph convolutional network, the graph structure data at two consecutive moments can be processed simultaneously, and the parameters of the entire network loss function can be adjusted through the Siamese graph convolutional network, and the node features after the local time-dependent process data is converted into a higher-dimensional representation. Finally, the process offline data is divided into performance levels according to the comprehensive economic indicators of the industrial process, and the state recognition model is trained using the division results combined with the hidden features learned by the feature extraction model to obtain a complete process operation status evaluation model. This method can timely and accurately evaluate the industrial operation status and guide the optimization and regulation of the process, thereby improving the operating performance of the industrial process, effectively ensuring the quality of industrial products, and improving the comprehensive economic benefits of the enterprise.

[0097] The technical solution of the present invention is described in detail below in conjunction with embodiments, and its feasibility is verified.

[0098] Example:

[0099] The coal slime flotation process is a typical process industry process running in a harsh open environment. There are various interference and uncertainty factors in the operating environment, which often make the current process monitoring and operating status evaluation methods incomplete in terms of the perception of the working condition information. The present invention aims at the strong dynamics and strong nonlinearity of the data in the coal slime flotation process, and can better mine various related interaction information, and evaluate the current operating status more accurately and robustly.

[0100] The typical coal slime flotation process is first crushed and screened, the raw coal is adjusted to the appropriate size and fully mixed with the heavy medium suspension delivered by the qualified medium barrel in the mixing barrel, and the mixture is injected into the heavy medium cyclone by a pressure pump for primary sorting. Secondly, the clean coal obtained by screening is transported to the coal slime barrel, and the clean coal containing finer particles is transported to the bottom of the concentrator for thickening and adjustment to the appropriate density. Then after the concentrator is completed, the coal slime slurry is discharged from the bottom of the concentrator, fully mixed with the flotation reagent in the slurry pre-processor, and enters the flotation tank for froth flotation. Finally, the scraper device is used to remove the overflow in the flotation tank, and after dehydration and drying, the impurities such as clean coal and gangue are left at the bottom of the flotation tank and discharged as tailings.

[0101] Since the variables involved in the entire coal slime flotation process are complex and redundant, and there are some variables that are irrelevant to the operating performance evaluation index or have little correlation, the calculation burden is increased. The mutual information value of each candidate input variable is calculated according to the above MIC method, such as Figure 4 This paper selects measurable variables whose calculation results are greater than the average value as the raw data composition of the slow feature Siamese graph convolutional network model, namely the following 20 typical process variables. The specific process variables are shown in Table 1.

[0102] Table 1: Process operation variable selection

[0103]

[0104] The data of this example comes from the full-process simulation platform of coal slime flotation process (registration number: 2020SR1160052). In order to generate coal slime flotation process data of different state levels, the production site is simulated by introducing deviations into the process operation variables in the simulation platform, so that the production process deviates from the assumed optimal operating point to reflect different operating states. The disturbance generation function of variable 1 (raw coal washing amount) is: x=base+dis*(sin(T1πt)+sin(T2πt)), and the deviation ramp signal of variable 37 (thickener underflow flow) and variable 44 (flotation tank click stirring speed) is set to: y signal =base+k signal t+b signal , where base represents the optimal reference value of the operating state, dis represents the disturbance amplitude, T1, T2 and b signal represents a constant, k signal Indicates the signal slope.

[0105] 1) Establish an offline status evaluation model

[0106] After removing unstable data caused by time delay, 14,000 sets of samples collected from the full-process simulation platform of coal slime flotation process were retained as the experimental data set. In this experiment, 7,000 sets of data were selected as the offline training data set for the evaluation model, and the remaining data were used for online testing.

[0107] Combined with expert experience, the state level of the coal preparation process is divided according to the overflow ash value of the flotation tank in the clean coal product. The state level division results are shown in Table 2:

[0108] Table 2: Classification of operating status and setting of class labels

[0109]

[0110] According to the characteristics of the coal slime flotation process, the local SFA parameters of the slow feature Siamese graph convolutional network are optimized in real time during training and testing. The topological structure graph dimension of the local slow feature is set to 50*50 and the moving step size is also set to 10, which strikes a balance between capturing fine information about the process and avoiding noise interference. The outputs of the two convolutional layers under the Siamese graph convolutional network module are 11 and 4 respectively, and 6 groups of features with slow changes are retained as SFA dimensions after MSV calculation. The learning rate of the slow feature Siamese graph convolutional network training is set to 0.001 and the epoch is set to 200. In order to optimize the network parameters, the adaptive point estimation (Adam) stochastic optimization algorithm is used to minimize the cross entropy loss function, and the operating status evaluation model is established offline.

[0111] 2) Online testing of industrial process operation status evaluation method based on local slow feature Siamese graph convolutional network

[0112] Figure 5 The online evaluation results of the industrial process operation status evaluation method based on the local slow feature Siamese graph convolutional network are shown in the figure. The weighted average precision is 94.96%, and the recall rates of the four levels are 99.01%, 88.35%, 97.26%, and 96.43%, respectively. Figure 5 It can be seen that the level result given by the industrial process operation state evaluation method based on the local slow feature Siamese graph convolution network has a high degree of overlap with the level of the actual state of the current operation state, that is, the operation state evaluation method based on the local slow feature Siamese graph convolution network can effectively identify the state level to which the operation data of the coal slime flotation process belongs. At the same time, the operation state evaluation method based on the local slow feature Siamese graph convolution network has a more stable effect on the probability recognition of the operation data belonging to the corresponding operation state level. It can be seen that the present invention has a higher robustness in the recognition of each state level.

[0113] The above simulation examples show that the present invention has certain effectiveness. By implementing timely and robust operation status evaluation on the coal slime flotation process, it can provide important guidance for production adjustment of the coal slime flotation process.

[0114] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification 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 complex industrial process operation status evaluation method based on local slow feature Siamese graph convolutional network, characterized in that: The specific steps include: Step 1: Use offline data to train the local slow feature Siamese graph convolutional network feature extraction model and classifier model to establish an offline model for state evaluation; A1: Collecting raw data from industrial processes Where N is the number of sampling points, X = [x1, x2, ..., x N ] is the industrial process data, Y = [y1, y2, ..., y N ] is the corresponding comprehensive economic indicator data, and the original data is preprocessed, including the removal and alignment of abnormal data; A2: Perform maximum and minimum standardization processing on the X data and Y data according to formula (1) so that the mean of the calculated data is 0 and the standard deviation is 1; A3: Collect two sets of time series data sets with length n at time t and time t-1 as the slow feature Siamese graph convolution network training data set, and record them as and Where s = t or t-1; A4: Classify the operating status of the comprehensive economic indicator data Y according to expert knowledge to obtain label data A5: Yes L Perform One-Hot encoding to obtain binary representation of level label data A6: Introduce a sliding window technology with a fixed length of w to split the standardized training data set to obtain a sliding window set and Where w≤N, A7: According to formula (2), MIC is used to discover the interaction relationship between process data; Where D is a function of the sample size, usually set to M 0.6 ; M is the length of the time series in the training dataset; A8: As shown in formula (3), the interaction between time series in each sliding window can be measured by MIC method to obtain the MIC weight matrix; In the formula, represents the MIC score between the sequences of node i and node j in the kth sliding window, Represents different time series data in the kth sliding window; A9: Set a suitable threshold, reset the elements in the weight matrix to 0 if they are less than the threshold, and reset them to 1 if they are greater than the threshold, so as to obtain an adjacency matrix that can reflect the node connection relationship and construct a series of association graphs g k ; Its calculation formula is shown in formula (4): In the formula, Represents the adjacency matrix The elements of and ζ represents the node connection relationship score threshold determined by expert knowledge; A10: Calculate the graph structure data set of training data in Represents the graph structure data under the kth input sample; the adjacency matrix set in the model input information can be expressed as A11: Initialize the parameters of the slow feature Siamese graph convolutional network, forward propagate to find the intrinsic connection between nodes in the association graph, and obtain the hidden layer features and And calculate The optimal model obtained by reverse iterative optimization; A12: Use the MIC weight matrix as the initialization parameter of the slow feature Siamese graph convolutional network to shorten the network training time. Input the sliding window data X w And the adjacency matrix A w Afterwards, the network forward propagates to obtain the output value, and combines the comprehensive economic indicator information to reversely train the slow feature Siamese graph convolutional network until the objective function converges; A13: Using Class Label Data Introducing supervised learning into the slow feature Siamese graph convolutional network to obtain the reconstruction value of the process data At the same time, we can get the forecast value of comprehensive economic indicators A14: Select the cross entropy loss function as the optimization objective function of the network training process, and calculate the loss function J of the feature extraction model according to formula (5) GCN : A15: After initializing the parameter matrix of the Softmax classifier, optimize the classifier model until the objective function is minimized or converged; A16: The hidden layer output h of the feature extraction model and the binary variable As input, train the Soffmax classifier to obtain the state recognition model, and obtain the output of the Softmax classifier according to formula (6): In the formula, c = 1, 2, ..., C represents different state levels, p(y = j|x) represents the posterior probability that the input x is state level j, is the parameter of the classifier; A17: Calculate the classifier loss function J according to formula (7) sofotmax (θ): A18: Backward fine-tuning of the classifier model parameters until J is minimized softmax (θ) or realize J softmax The convergence of (θ) leads to the state recognition model; A19: Cascade feature extraction model and classifier and fine-tune parameters to build a complete offline model for running status evaluation; Step 2: Use online data to evaluate the operation status; B1: Sample and obtain online process data X, and perform standardization on X; B2: Sliding sampling is performed on the online data using a sliding window with a window length of w to obtain sequence data with consistent length; B3: Input the sequence data into the trained operation status evaluation model and calculate the online data x at time t t The posterior probability of belonging to different operating status levels is Among them, i∈{1, 2, ..., q}, q is the digital label representation of the state level category divided according to the comprehensive economic indicator data, and the final process operation state at time t is defined as the posterior probability set The state level corresponding to the maximum value in, that is, the process operation state level at time t is B4: Based on the online operation status evaluation results, guide the subsequent optimization and regulation of the industrial process.

Images