VAGCN-based complex industrial process operation state evaluation method

Abstract

The invention discloses a VAGCN-based complex industrial process operation state evaluation method. The method comprises the following steps: collecting process data and carrying out standardization processing; performing data division on the data by using a sliding window to obtain a time sequence; the method comprises the following steps: converting time sequence data into cross-time-space variable interaction graph structure data through a variable interaction graph structure, and then capturing variable interaction characteristics and local space-time dependency in each time window by using a moving window; time pooling is used, advanced features of variable interaction are extracted and then input into the operation state evaluation model, posterior probabilities of the online data belonging to different operation state evaluation grades are obtained, and a final evaluation result is a state grade corresponding to the maximum posterior probability at the current moment; and performing similarity analysis on the variable interaction graph structure in the non-optimal state and the optimality graph structure in the optimal state so as to evaluate the variable dynamic offset degree in the non-optimal state. According to the method, the interaction among the variables at different time points can be captured, and more accurate process operation state evaluation can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial production process operation status evaluation, and specifically is a complex industrial process operation status evaluation method based on VAGCN. 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 must not only have high stability, 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 often faces the situation of deviation from the optimal operation track. These factors may lead to a decrease in production efficiency, fluctuations in product quality, and even serious safety accidents, thus failing to ensure the continuous improvement of product quality and economic benefits. Therefore, timely and accurate evaluation of the operating status of the industrial process, and then guiding production operators to effectively regulate the industrial process, is of great practical significance for improving the operating performance of the industrial process and facilitating the production management of the enterprise.

[0003] Due to the complex process mechanisms and different production modes, the collected industrial data sets contain many different types of features. Among the common data features, nonlinearity and dynamics are two features that exist in almost all collected industrial data sets. In order to build a reliable evaluation model, these data features should be carefully considered and processed. In addition to the nonlinearity and dynamics of the data, a widespread but rarely addressed process problem should actually be considered, namely the dynamic offset of variables caused by the time delay between process variables and quality variables. For the same batch of production materials, variable time delays are caused by different measurement locations, which is caused by the natural process structure and the transmission time of production materials. However, industrial data acquisition systems collect data from all variables at the same time, which means that variable time delays themselves exist in the collected data sets. Time lags can significantly affect the data distribution pattern and the dynamic dependencies between variables, which in turn leads to poor performance of the prediction model. Although many existing GNN-based models can solve the spatiotemporal dependency problem of multivariate data to a certain extent, they usually separate the temporal and spatial dependencies and fail to capture the dynamic interactions between different time points and different variables. Factors such as strong coupling, multivariate dynamic dependencies, and time delays in industrial processes often cause the interaction relationship between variables to change with time delays. If the variable interaction characteristics across time and space are not fully considered, it will be difficult to accurately reflect the actual operating status of the system, which will affect the stability and accuracy of performance evaluation. Therefore, when dealing with the problem of multi-variable dynamic offset in complex industrial processes, existing methods show certain limitations. Summary of the invention

[0004] In view of the problems existing in the above-mentioned prior art, the present invention provides a method for evaluating the operating status of complex industrial processes based on VAGCN. The method has a simple implementation process and low implementation cost. It can evaluate the industrial operating status in a timely and accurate manner, provide reliable guidance for the optimization and regulation of the process, help to significantly improve the operating efficiency of the industrial process, and ensure the quality of industrial products.

[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 VAGCN, comprising the following steps:

[0006] Step 1: Use offline data to train the feature extraction model and classifier model based on the variable interaction graph structure learning to establish an offline model for state evaluation;

[0007] A1: First, use several sensors placed at key nodes in the industrial production system to collect the original data of the actual industrial process. Where N is the number of sampling points, X = [x1, x2, ..., x N ] is the industrial process data, Y=[y1,y2,...,yN ] is the corresponding comprehensive economic indicator data;

[0008] A2: Preprocess the original data by removing abnormal data and aligning the original data, and then perform maximum and minimum standardization on the X data and Y data according to formula (1) so that the calculated data has a mean of 0 and a standard deviation of 1;

[0009]

[0010] In the formula, x is the input data, μ represents the mean of the input data, and σ is the standard deviation of the input data;

[0011] A3: First, select the appropriate comprehensive economic indicator data from the pre-processed data according to the specific production process, and then divide the selected data into training data sets and test data sets in proportion according to different industrial operation status levels in combination with expert knowledge. Then, mark the different status level categories with digital labels in turn to obtain the real label data y L ;

[0012] A4: Introduce the time series slicing technology with a fixed length of H to split the training data set, and obtain the data set as The data for each time slice is Each X t contains time slice samples from N variables, i.e. t represents the window index of the time window, Indicates the number of time slices,

[0013] A5: Use an encoder f c (·|W c ) to process the variable data samples in each window W(t), perform feature learning on each variable segment, and extract the local features of each variable, that is, x t ' ,i =f c (x t,i |W c );

[0014] A6: Use position encoding to supplement the relative position information of graph structures at different time points; for the i-th variable By adding position encoding, we can get enhanced feature z t,i =f p (t)+x t ' ,i , as shown in formula (2);

[0015]

[0016] In the formula, p i is the position encoding, m represents the mth feature among the variable features; ω k is the frequency;

[0017] A7: Dot product similarity is used to quantify the similarity between two variables at different time points, defined as e tr,ij =g s (z t,i ) s (z r,j ) T , where t,r∈[1,T], i,j∈[1,N], and function g s (z)=zW s is used to enhance the expressive power, where W s are learnable weights;

[0018] A8: Use the softmax function to limit the correlation to [0, 1]; finally, we get a graph structure with interactive connections between different variables at different time points, i.e., the variable interaction graph structure G. t =(Z t ,A t ),in, A t It is represented as the adjacency matrix of the variable interaction graph structure within the time window W(t);

[0019] A9: Use the decay matrix c tr,ij =w t-r To adjust the variable interaction graph G t =(Z t ,A t ) is the edge weight value between different time points, as shown in formula (3);

[0020]

[0021] A10: Through multi-layer graph convolution operations, the neighbor information of the node is gradually aggregated; for each input feature in the time window W(t) and the adjacency matrix After aggregation, the update process of the graph convolution layer is shown in formula (4), where: Represents the initial features of each node in the time window W(t);

[0022]

[0023] Where: represents the variable interaction adjacency matrix constructed within the time window W(t); is the degree matrix of the adjacency matrix; is the updated node feature matrix after the l-th layer of graph convolution. Initially W (l) represents the learnable parameter matrix of the lth layer; σ is the nonlinear activation function;

[0024] A11: Use average pooling to aggregate features at different times. After time pooling, the feature matrix is flattened into a vector

[0025] A12: Use a fully connected layer to perform linear transformation on the flattened features and generate the final classification output through an activation function;

[0026] A13: Apply the softmax function to calculate the weighted probability that the current predicted state belongs to each state, and take the maximum value of the probability of each state as the final model prediction state;

[0027] A14: Select the cross entropy loss function as the optimization objective function of the network training process, as shown in formula (5);

[0028]

[0029] In the formula, N represents the total number of labels of the corresponding sample set. is the label data predicted by the model;

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

[0031] A16: 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 (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(θ) converges, and the state recognition model is obtained and saved;

[0037] A19: Cascade feature extraction model and classifier model 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: Use several sensors placed at key nodes in the industrial production system to perform online sampling, obtain online process data X, and perform standardized processing on X;

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

[0041] B3: Input the sequence data into the trained offline model for running status evaluation 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]

[0043] B4: Based on the online operation status evaluation results, guide the subsequent optimization and control of the industrial process;

[0044] When the offline model for running status evaluation evaluates the current running status as non-optimal, a dynamic offset evaluation of non-optimal state variables is performed. The specific steps are as follows:

[0045] S1: Extract the topological structure features G of the variable interaction graph within the time window under non-optimal conditions non-opt ;

[0046] S2: Use the Frobenius norm to calculate its optimality graph structure G opt The norm of the difference matrix ΔG between them;

[0047] S3: Obtain the similarity score S according to formula (8), and use the similarity score S to measure the closeness between the non-optimal state and the optimal state, and quantify the variable deviation degree and corresponding adjustment strategy under the non-optimal state to guide the operator to adjust the production process accurately and efficiently;

[0048]

[0049] The present invention provides a method for evaluating the operation status of a complex industrial process based on VAGCN. The method first effectively captures the dynamic dependency between different time points and variables by constructing a variable interaction graph structure across time and space, greatly improving the accuracy and robustness of the operation status evaluation. Under the framework of the variable interaction graph convolutional network model, the graph convolutional neural network (GCN) is first combined to deeply extract the interaction features between variables, and the online performance is constrained in real time by introducing the comprehensive economic index (CEI). An accurate and efficient online operation status evaluation model is constructed based on the variable interaction features. The graph convolutional neural network is further used to perform deep learning on the variable interaction graph structure. In this process, the variable nodes at the previous moment can aggregate the multi-variable characteristics of the subsequent moments. Even in the face of variable offset, the information interaction between variables can still be maintained, which greatly alleviates the problem of information loss caused by variable offset. In addition, by extracting the topological structure features of the variable interaction graph in the time window under the non-optimal state and evaluating the similarity with the superiority graph structure under the optimal state, the degree of dynamic offset of variables under the non-optimal state can be effectively quantified and evaluated, providing a new idea for the evaluation of non-optimal states in complex industrial systems.

[0050] The implementation process of this method is simple and the implementation cost is low. It can effectively extract the dynamic deviation information of variables caused by time delays and capture the interaction between variables at different time points, so as to timely and accurately evaluate the industrial operation status, provide reliable guidance for the optimization and regulation of the process, and help to achieve more accurate process operation status evaluation, thereby significantly improving the operating efficiency of the industrial process, effectively ensuring the quality of industrial products, and laying a solid foundation for improving the comprehensive economic benefits of enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The overall flow chart of the industrial process operation status evaluation in the present invention;

[0052] Figure 2 It is a schematic diagram of the operating status evaluation structure based on the VAGCN model in the present invention;

[0053] Figure 3 It is the online evaluation result of the state based on the VAGCN model in the present invention;

[0054] Figure 4 It is a ROC curve diagram of the online evaluation results of the complex industrial process operation status evaluation model based on VAGCN in the present invention;

[0055] Figure 5 It is the confusion matrix diagram of the VAGCN model;

[0056] Figure 6 It is the variable dynamic deviation curve under the disturbance of variable 1;

[0057] Figure 7 It is the variable dynamic deviation curve under the disturbance of variable 7;

[0058] Figure 8 It is the variable dynamic deviation curve under the disturbance of variable 9. DETAILED DESCRIPTION

[0059] The present invention will be further described below in conjunction with the accompanying drawings.

[0060] like Figures 1 to 8 As shown, the present invention provides a method for evaluating the operating status of a complex industrial process based on VAGCN (Variable-Aware Graph Convolutional Network), comprising the following steps:

[0061] Step 1: Use offline data to train the feature extraction model and classifier model based on the variable interaction graph structure learning to establish an offline model for state evaluation;

[0062] A1: First, use several sensors placed at key nodes in the industrial production system to collect the original data of the actual industrial process. 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;

[0063] A2: Preprocess the original data by removing abnormal data and aligning the original data, and then perform maximum and minimum standardization on the X data and Y data according to formula (1) so that the calculated data has a mean of 0 and a standard deviation of 1;

[0064]

[0065] In the formula, x is the input data, μ represents the mean of the input data, and σ is the standard deviation of the input data;

[0066] A3: First, select the appropriate comprehensive economic indicator data from the pre-processed data according to the specific production process, and then divide the selected data into training data sets and test data sets in proportion according to different industrial operation status levels in combination with expert knowledge. Then, mark the different status level categories with digital labels in turn to obtain the real label data y L ;

[0067] A4: Introduce the time series slicing technology with a fixed length of H to split the training data set, and obtain the data set as The data for each time slice is Each Xt contains time slice samples from N variables, i.e. t represents the window index of the time window, Indicates the number of time slices,

[0068] A5: Use an encoder f c (·|W c ) to process the variable data samples in each window W(t), perform feature learning on each variable segment, and extract the local features of each variable, that is, x t ' ,i =f c (x t,i |W c );

[0069] A6: Use position encoding to supplement the relative position information of graph structures at different time points; for the i-th variable By adding position encoding, we can get enhanced feature z t,i =f p (t)+x t ' ,i , as shown in formula (2);

[0070]

[0071] In the formula, p i is the position encoding, m represents the mth feature among the variable features; ω k is the frequency;

[0072] A7: Dot product similarity is used to quantify the similarity between two variables at different time points, defined as e tr,ij =g s (z t,i ) s (z r,j ) T , where t,r∈[1,T], i,j∈[1,N], and function g s (z)=zW s is used to enhance the expressive power, where W s are learnable weights;

[0073] A8: Use the softmax function to limit the correlation to [0, 1]; finally, we get a graph structure with interactive connections between different variables at different time points, i.e., the variable interaction graph structure G. t =(Z t ,A t ),in, A tIt is represented as the adjacency matrix of the variable interaction graph structure within the time window W(t);

[0074] A9: Use the decay matrix c tr,ij =w t-r To adjust the variable interaction graph G t =(Z t ,A t ) is the edge weight value between different time points, as shown in formula (3);

[0075]

[0076] A10: Through multi-layer graph convolution operations, the neighbor information of the node is gradually aggregated; for each input feature in the time window W(t) and the adjacency matrix After aggregation, the update process of the graph convolution layer is shown in formula (4), where: Represents the initial features of each node in the time window W(t);

[0077]

[0078] Where: represents the variable interaction adjacency matrix constructed within the time window W(t); is the degree matrix of the adjacency matrix; is the updated node feature matrix after the l-th layer of graph convolution. Initially W (l) represents the learnable parameter matrix of the lth layer; σ is the nonlinear activation function;

[0079] A11: Use average pooling to aggregate features at different times. After time pooling, the feature matrix is flattened into a vector

[0080] A12: Use the fully connected layer (FC layer) to perform linear transformation on the flattened features and generate the final classification output through the activation function;

[0081] A13: Apply the softmax function to calculate the weighted probability that the current predicted state belongs to each state, and take the maximum value of the probability of each state as the final model prediction state;

[0082] A14: Select the cross entropy loss function as the optimization objective function of the network training process, as shown in formula (5);

[0083]

[0084] In the formula, N represents the total number of labels of the corresponding sample set. is the label data predicted by the model;

[0085] A15: After initializing the parameter matrix of the Softmax classifier, optimize the classifier model until the objective function is minimized or converged, and then save the classifier model;

[0086] A16: 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 (6);

[0087]

[0088] 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;

[0089] A17: Calculate the classifier loss function J according to formula (7) softmax (θ);

[0090]

[0091] A18: Backward fine-tuning of the classifier model parameters until J is minimized softmax (θ) or realize J softmax (θ) converges, and the state recognition model is obtained and saved;

[0092] A19: Cascade feature extraction model and classifier model and fine-tune parameters to build a complete offline model for running status evaluation;

[0093] Step 2: Use online data to evaluate the operation status;

[0094] B1: Use several sensors placed at key nodes in the industrial production system to perform online sampling, obtain online process data X, and perform standardized processing on X;

[0095] B2: Sliding sampling is performed on the online data after standardization with a sliding window of window length H to obtain sequence data of consistent length;

[0096] B3: Input the sequence data into the trained offline model for running status evaluation 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

[0098] B4: Based on the online operation status evaluation results, guide the subsequent optimization and control of the industrial process;

[0099] When the offline model for running status evaluation evaluates the current running status as non-optimal, a dynamic offset evaluation of non-optimal state variables is performed. The specific steps are as follows:

[0100] S1: Extract the topological structure features G of the variable interaction graph within the time window under non-optimal conditions non-opt ;

[0101] S2: Use the Frobenius norm to calculate its optimality graph structure G opt The norm of the difference matrix ΔG between them;

[0102] S3: Obtain the similarity score S according to formula (8), and use the similarity score S to measure the closeness between the non-optimal state and the optimal state, and quantify the variable deviation degree and corresponding adjustment strategy under the non-optimal state to guide the operator to adjust the production process accurately and efficiently;

[0103]

[0104] Example:

[0105] The heavy medium coal preparation 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, strong nonlinearity and dynamic offset between variables in the data of the coal slime flotation process, and can more effectively extract the dynamic interaction characteristics between variables, and evaluate the current operating status more accurately and robustly.

[0106] The industrial process selected is the heavy medium two-product conical heavy medium cyclone coal preparation process. The main equipment includes: raw coal particle size grading screen, mixing barrel, heavy medium cyclone, magnetic separator, de-mediation screen, qualified medium barrel, etc. The coal preparation plant uses an automatic feeder to transport the raw coal to the washing process. After passing through the double-layer grading screen, coal particles with a particle size of more than 25mm are sent to other sorting equipment such as drum separators for sorting. Medium-sized raw coal particles (6mm~25mm) are transported to the mixing box to mix with the heavy medium suspension and pumped into the heavy medium cyclone. The coal-medium mixture completes density stratification under the action of the centrifugal field in the cyclone, and is discharged from the bottom flow port and overflow port respectively, and sent to the de-mediation screen for de-mediation treatment. At the same time, the magnetic weighting material is recovered by the magnetic separator. The recovered weighting material and the supplemented magnetic weighting material are mixed in the qualified medium barrel. By adjusting the input amount of water and weighting material, the liquid level of the qualified medium barrel and the output density of the heavy medium suspension are kept within a suitable range, and then transported to the mixing barrel for washing.

[0107] Ash content is usually used as a basic indicator for monitoring coal quality. The higher the ash content, the worse the coal quality. In the heavy medium coal preparation process, due to the involvement of a large number of equipment and instruments, the heavy medium coal preparation process has a strong nonlinear characteristic. In order to ensure the quality of coal, the parameters of the coal preparation process must be continuously tested and adjusted in actual operation. According to the process mechanism and actual conditions, 11 process variables were determined, and the ash content index was determined as the output variable. The specific process variables are shown in Table 1.

[0108] Table 1: Process operation variable selection

[0109]

[0110] 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:

[0111] x=base-dis*(sin(T1πt)+sin(T2πt)), where base is the optimal variable reference value for the operating state and dis is the fluctuation amplitude. T1 and T2 are constant parameters, set to 0.02 and 0.0096 respectively. The results of the non-optimal factor settings are shown in Table 2.

[0112] Table 2: Set disturbance variables and corresponding sampling periods

[0113]

[0114] 1) Establish an offline status evaluation model

[0115] 5000 sets of data were selected from the data set generated by the heavy medium coal preparation process mechanism model. The first 2500 sets 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 testing of the model.

[0116] 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 3:

[0117] Table 3: Status level classification and corresponding digital labels

[0118]

[0119] The time window is set to 5, and the variable interaction graph structure is constructed within a small window, which can extract local features more finely and reduce the amount of calculation. The moving step is set to 1, which ensures that the window will completely cover all time step data every time. Considering the time distance between nodes, the attenuation coefficient is set to 0.4, which directly affects the importance of the interaction between time steps. Because graph convolution is prone to over-smoothing problems, the experiment only uses two layers of one-dimensional convolutional layers to build the model. The hidden layer dimension is set to 35, which means that when the graph neural network is processed, the features of each sensor node will be mapped to a 35-dimensional hidden representation. This dimension directly affects the complexity and expressiveness of the model. This dimension can improve the model's ability to capture complex patterns without causing the model to overfit. Average pooling is used as the pooling method, and the optimizer uses the Adam optimizer, with a learning rate of 0.001.

[0120] 2) Online testing of industrial process operation status evaluation method based on VAGCN

[0121] The evaluation results of the trained model are shown in Figure (3), where the red triangle represents the true label of the data, that is, the current true state level of the complex industrial process; the blue circle is the judgment result obtained by the evaluation model. It can be seen that even if the operating state fluctuates and in the unstable state transition range, the judgment result basically matches and coincides with the true result.

[0122] Figure (4) shows the ROC curve of the online evaluation results of the complex industrial process operation status evaluation model based on variable interaction graph structure learning. The ROC curve is usually located in the upper left of the coordinate system. The closer it is to the point (0,1), the better the classification effect. In addition, the area under the curve (AUC) is also used to measure the performance of the model. The closer the AUC value is to 1, the better the model classification effect is, and vice versa. It can be seen that the present invention has a higher robustness in identifying each state level.

[0123] Figure 5 This is the confusion matrix of the model. The numbers on the diagonal in the confusion matrix reflect the similarity between the model recognition results and the true state level. The closer the numbers on the diagonal are to 1, the higher the recognition accuracy of the model. From the confusion matrix, we can see that the VAGCN model has higher prediction rate, recall rate and evaluation accuracy.

[0124] 3) Dynamic deviation assessment of non-optimal states.

[0125] When the evaluation result of the industrial process operation status model based on the variable interaction graph neural network is non-optimal, it is necessary to analyze the dynamic deviation of the variables.

[0126] The variable interaction graph structures in multiple windows under the optimal state are averaged to obtain an optimality graph structure representing the optimal state. Then, the variable interaction graph structure under the non-optimal state is analyzed for similarity with the optimality graph structure under the optimal state, and the corresponding similarity scores are obtained. These scores are used to quantify the degree of deviation of the dynamic behavior of the variables under different disturbances, and then the corresponding adjustment strategies can be obtained according to the degree of deviation to guide the operators to adjust the production process accurately and efficiently. In order to reduce the uncertainty of the comparison results, five groups of experiments were carried out with the same experimental settings. The variable interaction graph structure of the same window in each group of experiments was extracted, and the uncertainty of the comparison results was described by the confidence interval. The above similarity analysis results are shown in Figure (6). The variable dynamic deviation curves under three different disturbances show the time variation characteristics of the degree of variable deviation in the system. Through the changing trends of these curves, it can be observed that under the three different disturbances, the degree of deviation and stability of the non-optimal state are different.

[0127] The effectiveness of the present invention can be seen from the above simulation examples. By implementing timely and robust operation status evaluation on the coal slime flotation process, it is possible to provide important guidance for production adjustment of the coal slime flotation process.

[0128] The present invention provides a method for evaluating the operation status of a complex industrial process based on VAGCN. The method first effectively captures the dynamic dependency between different time points and variables by constructing a variable interaction graph structure across time and space, greatly improving the accuracy and robustness of the operation status evaluation. Under the framework of the variable interaction graph convolutional network model, the graph convolutional neural network (GCN) is first combined to deeply extract the interaction features between variables, and the online performance is constrained in real time by introducing the comprehensive economic index (CEI). An accurate and efficient online operation status evaluation model is constructed based on the variable interaction features. The graph convolutional neural network is further used to perform deep learning on the variable interaction graph structure. In this process, the variable nodes at the previous moment can aggregate the multi-variable characteristics of the subsequent moments. Even in the face of variable offset, the information interaction between variables can still be maintained, which greatly alleviates the problem of information loss caused by variable offset. In addition, by extracting the topological structure features of the variable interaction graph in the time window under the non-optimal state and evaluating the similarity with the superiority graph structure under the optimal state, the degree of dynamic offset of variables under the non-optimal state can be effectively quantified and evaluated, providing a new idea for the evaluation of non-optimal states in complex industrial systems.

[0129] The implementation process of this method is simple and the implementation cost is low. It can effectively extract the dynamic deviation information of variables caused by time delays and capture the interaction between variables at different time points, so as to timely and accurately evaluate the industrial operation status, provide reliable guidance for the optimization and regulation of the process, and help to achieve more accurate process operation status evaluation, thereby significantly improving the operating efficiency of the industrial process, effectively ensuring the quality of industrial products, and laying a solid foundation for improving the comprehensive economic benefits of enterprises.

[0130] 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 VAGCN, characterized in that: The following steps are involved: Step 1: Use offline data to train the feature extraction model and classifier model based on the variable interaction graph structure learning to establish an offline model for state evaluation; A1: First, use several sensors placed at key nodes in the industrial production system to collect the original data of the actual industrial process. 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; A2: Preprocess the original data by removing abnormal data and aligning the original data, and then perform maximum and minimum standardization on the X data and Y data according to formula (1) so that the calculated data expectation is 0 and the standard deviation is 1; In the formula, x is the input data, μ represents the mean of the input data, and σ is the standard deviation of the input data; A3: First, select the appropriate comprehensive economic indicator data from the pre-processed data according to the specific production process, and then divide the selected data into training data sets and test data sets in proportion according to different industrial operation status levels in combination with expert knowledge. Then, mark the different status level categories with digital labels in turn to obtain the real label data y L ; A4: Introduce the time series slicing technology with a fixed length of H to split the training data set, and obtain the data set as The data for each time slice is Each X t contains time slice samples from N variables, i.e. t represents the window index of the time window, Indicates the number of time slices, A5: Use an encoder f c (·|W c ) to process the variable data samples in each window W(t), perform feature learning on the fragments of each variable, and extract the local features of each variable, that is, x t ' ,i =f c (x t,i |W c ); A6: Use position encoding to supplement the relative position information of graph structures at different time points; for the i-th variable By adding position encoding, we can get enhanced feature z t,i =f p (t)+x t ' ,i , as shown in formula (2); In the formula, p i is the position encoding, m represents the mth feature among the variable features; ω k is the frequency; A7: Dot product similarity is used to quantify the similarity between two variables at different time points, defined as e tr,ij =g s (z t,i ) s (z r,j ) T , where t,r∈[1,T], i,j∈[1,N], and function g s (z)=zW s is used to enhance the expressive power, where W s are learnable weights; A8: Use the softmax function to limit the correlation to [0, 1]; finally, we get a graph structure with interactive connections between different variables at different time points, i.e., the variable interaction graph structure G. t =(Z t ,A t ),in, A t It is represented as the adjacency matrix of the variable interaction graph structure within the time window W(t); A9: Use the decay matrix c tr,ij =w t-r To adjust the variable interaction graph G t =(Z t ,A t ) is the edge weight value between different time points, as shown in formula (3); A10: Through multi-layer graph convolution operations, the neighbor information of the node is gradually aggregated; for each input feature in the time window W(t) and the adjacency matrix After aggregation, the update process of the graph convolution layer is shown in formula (4), where: Represents the initial features of each node in the time window W(t); Where: represents the variable interaction adjacency matrix constructed within the time window W(t); is the degree matrix of the adjacency matrix; is the updated node feature matrix after the l-th layer of graph convolution. Initially W (l) represents the learnable parameter matrix of the lth layer; σ is the nonlinear activation function; A11: Use average pooling to aggregate features at different times. After time pooling, the feature matrix is flattened into a vector A12: Use a fully connected layer to perform linear transformation on the flattened features and generate the final classification output through an activation function; A13: Apply the softmax function to calculate the weighted probability that the current predicted state belongs to each state, and take the maximum value of the probability of each state as the final model prediction state; A14: Select the cross entropy loss function as the optimization objective function of the network training process, as shown in formula (5); In the formula, N represents the total number of labels of the corresponding sample set. is the label data predicted by the model; A15: After initializing the parameter matrix of the Softmax classifier, optimize the classifier model until the objective function is minimized or converged, and then save the classifier model; A16: 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 (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) softmax (θ); A18: Backward fine-tuning of the classifier model parameters until J is minimized softmax (θ) or realize J softmax (θ) converges, and the state recognition model is obtained and saved; A19: Cascade feature extraction model and classifier model 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: Use several sensors placed at key nodes in the industrial production system to perform online sampling, obtain online process data X, and perform standardized processing on X; B2: Sliding sampling is performed on the online data after standardization with a sliding window of window length H to obtain sequence data of consistent length; B3: Input the sequence data into the trained offline model for running status evaluation 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 control of the industrial process; When the offline model for running status evaluation evaluates the current running status as non-optimal, a dynamic offset evaluation of non-optimal state variables is performed. The specific steps are as follows: S1: Extract the topological structure features G of the variable interaction graph within the time window under non-optimal conditions non-opt ; S2: Use the Frobenius norm to calculate its optimality graph structure G opt The norm of the difference matrix ΔG between them; S3: Obtain the similarity score S according to formula (8), and use the similarity score S to measure the closeness between the non-optimal state and the optimal state, and quantify the variable deviation degree and corresponding adjustment strategy under the non-optimal state to guide the operator to adjust the production process accurately and efficiently;

Images