A method for evaluating running state of complex industrial process based on VAGCN

Abstract

A kind of complex industrial process running state evaluation method based on VAGCN, process data are collected and standardized processing is carried out;Data division is carried out using sliding window, and time series are obtained;Time series data are converted into cross-time space variable interaction graph structure data by variable interaction graph structure, then moving window is used to capture variable interaction features and local space-time dependence in each time window;Time pooling is used to extract high-level features of variable interaction and then input running state evaluation model, to obtain the posterior probability that online data belongs to different running state evaluation grades, and the final evaluation result is the state grade corresponding to the maximum posterior probability at the current time;Similarity analysis is carried out between variable interaction graph structure under non-optimal state and optimality graph structure under optimal state to evaluate the degree of variable dynamic deviation under non-optimal state.The method can capture the interaction between variables at different time points, which helps to realize more accurate process running state evaluation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of industrial production process running state evaluation, and specifically relates to a complex industrial process running state evaluation method based on VAGCN. BACKGROUND

[0002] In current industrial production processes, maintaining the basic operation of the production system has far failed to meet the manufacturing requirements of high efficiency and high output. Modern industrial production pursues maximum efficiency and optimal resource utilization, which requires the production process not only to have high stability, but also to be able to adapt to various changes to maintain the best running state. However, in actual industrial production processes, due to the influence of external factors such as changes in the production environment, aging and wear of equipment, and non-standard personnel operation, the running process often faces the situation of deviating from the optimal running track. These factors can lead to a decrease in production efficiency, fluctuations in product quality, and even serious safety accidents, thereby failing to guarantee the continuous improvement of product quality and economic benefits. Therefore, timely and accurate evaluation of the running state of the industrial process, and then guiding the production operators to effectively regulate and control the industrial process, have important practical significance for improving the running performance of the industrial process and facilitating the production management of enterprises.

[0003] Due to the complex process mechanism and different production modes, the collected industrial data set contains 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 widely existing but rarely solved process problem should also be considered, that is, the variable dynamic shift caused by the time delay between the process variables and the quality variables. For the same batch of production materials, the variable time delay is caused by different measurement positions, which is caused by the natural process structure and the transmission time of the production materials. However, the industrial data acquisition system collects data from all variables at the same time, which means that the variable time delay itself exists in the collected data set. Time delay will significantly affect the data distribution pattern and the dynamic dependence between variables, which will in turn lead to poor performance of the prediction model. Although many existing GNN-based models can solve the spatiotemporal dependence problem of multivariate data to some extent, they usually separate the time and spatial dependence for processing, and fail to capture the dynamic interaction between different time points and different variables. The strong coupling, multivariate dynamic dependence and time delay in industrial processes often lead to changes in the interaction between variables with the delay of time. If the cross-time and spatial variable interaction features between variables are not adequately considered, it will be difficult to accurately reflect the actual running state of the system, thereby affecting the stability and accuracy of performance evaluation. Therefore, when dealing with the problem of multivariate dynamic shift in complex industrial processes, the existing methods show certain limitations. SUMMARY

[0004] In view of the problems existing in the prior art, the present application provides a complex industrial process running state evaluation method based on VAGCN, which has the advantages of simple implementation process, low implementation cost, timely and accurate industrial running state evaluation, reliable guidance for process optimization and control, and help to significantly improve the running efficiency of industrial processes and ensure the quality of industrial products.

[0005] In order to achieve the above purpose, the present application provides a complex industrial process running state evaluation method based on VAGCN, comprising the following steps:

[0006] Step 1: training a variable interaction graph structure learning feature extraction model and a classifier model using offline data to establish a state evaluation offline model;

[0007] A1: first use a plurality of sensors arranged at key nodes in the industrial production system to collect original data of the actual industrial process Wherein, N is the number of sampling points, X = [x1, x2,..., x N ] is the process data of the industry, Y = [y1, y2,..., yN ] for corresponding comprehensive economic index data;

[0008] A2: Preprocessing the original data, first removing abnormal data and aligning the original data, and then performing maximum and minimum standardization processing on X data and Y data according to formula (1), so that the calculated data mean is 0 and the standard deviation is 1;

[0009]

[0010] where 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 index data from the preprocessed data according to the specific production process, then combine expert knowledge, and divide the selected data into training data sets and test data sets according to different industrial running state levels, then mark different state level categories as digital labels in turn to get real label data y L ;

[0012] A4: Introduce a time series slice technology with a fixed length of H to segment the training data set, and get the data set as where the data of 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, represents 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), learn the features of each variable slice, and extract the local features of each variable, i.e. t ′ ,i = f c (x t,i |W c );

[0014] A6: Use position encoding to supplement the relative position information of different time point graph structures at the same time; for the i-th variable Enhanced features z t,i = f p (t) + x t ′ ,i are obtained by adding position encoding, as shown in formula (2);

[0015]

[0016] where p i is the position encoding, and m represents the m-th feature among the variable features; ω k is the frequency;

[0017] A7: The 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 )g s (z r,j ) T , where t, r ∈ [1, T], i, j ∈ [1, N], and the function g s (z) = zW s is used to enhance the expression capability, where W s is a learnable weight;

[0018] A8: The softmax function is used to limit the correlation to [0, 1] at the same time; finally, the graph structure with the interaction connection between different variables at different time points is obtained, that is, the variable interaction graph structure G t = (Z t , A t ), where A t is represented as the adjacency matrix of the variable interaction graph structure within the time window W(t);

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

[0020]

[0021] A10: Through the multi-layer graph convolution operation, the neighbor information of the node is gradually aggregated; for the input features and the adjacency matrix within each time window W(t), the update process of the graph convolution layer is shown in formula (4), where represents the initial features of each node within 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 first layer graph convolution, and the initial value is W (l) denotes the learnable parameter matrix of the l-th layer; and σ is a nonlinear activation function.

[0024] A11: The features of different times are summarized using average pooling. After time pooling, the feature matrix is flattened into a vector

[0025] A12: The flattened features are linearly transformed using a fully connected layer, and the final classification output is generated through an activation function;

[0026] A13: The softmax function is applied to calculate the weight probability of the current predicted state belonging to each state, and the maximum value of each state probability is taken as the final model prediction state;

[0027] A14: The cross-entropy loss function is selected as the optimization objective function in the network training process, as shown in equation (5);

[0028]

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

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

[0031] A16: The hidden layer output h of the feature extraction model and the binary variable are taken as inputs to train the Softmax classifier to obtain a state recognition model, and the output of the Softmax classifier is obtained according to equation (6);

[0032]

[0033] where c = 1, 2, …, C represents different state levels, p(y = j | x) represents the posterior probability of input x being state level j, is the parameter of the classifier;

[0034] A17: The classifier loss function J softmax (θ) is calculated according to equation (7);

[0035]

[0036] A18: The classifier model parameters are fine-tuned in the backward direction until J softmax (θ) is minimized or J softmaxconverge of (θ) to obtain and save the state recognition model;

[0037] A19: Fine-tune parameters after cascading feature extraction model and classifier model to build complete running state evaluation offline model;

[0038] Step two: use online data to evaluate running state;

[0039] B1: Use several sensors arranged at key nodes in the industrial production system to perform online sampling to obtain online process data X, and perform standardization processing on X;

[0040] B2: Slide sampling the online data after standardization processing with a sliding window of window length H to obtain sequence data of consistent length;

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

[0042]

[0043] B4: According to the online running state evaluation result, guide the subsequent optimization and control of the industrial process;

[0044] When the running state evaluation offline model evaluates the current running state as non-optimal, perform non-optimal state variable dynamic offset evaluation, and the specific steps are as follows:

[0045] S1: Extract the variable interaction graph topological structure feature G non-opt in the time window under the non-optimal state;

[0046] S2: Calculate the norm size of the difference matrix ΔG between G opt and the optimality graph structure G

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

[0048]

[0049] The application provides a complex industrial process running state evaluation method based on a VAGCN, which first effectively captures the dynamic dependence relationship between different time points and variables by constructing a cross-time and space variable interaction graph structure, greatly improving the accuracy and robustness of the running state evaluation. Under the framework of the variable interaction graph convolution network model, the interactive features between variables are first extracted in depth by combining the graph convolutional neural network (GCN), and the online performance is constrained in real time by introducing the comprehensive economic index (CEI), and an accurate and efficient online running state evaluation model is constructed by virtue of the variable interaction features. Further, the graph convolutional neural network is used for deep learning of the variable interaction graph structure, and in this process, the variable nodes at the previous moment can aggregate the multi-variable characteristics at the subsequent moment, so that the information interaction between variables can be maintained even in the case of variable offset, greatly relieving the information loss problem caused by variable offset. In addition, by extracting the variable interaction graph topological structure features in the time window under the non-optimal state, and evaluating the similarity with the optimality graph structure under the optimal state, the dynamic offset degree of variables under the non-optimal state can be effectively quantified, providing a new idea for the non-optimal state evaluation in complex industrial systems.

[0050] The method has simple implementation process and low implementation cost, can effectively extract variable dynamic deviation information caused by time delay and capture the interaction between variables at different time points, and can timely and accurately evaluate the industrial running state, thereby providing reliable guidance for process optimization and control, helping to realize more accurate process running state evaluation, and thus can significantly improve the running efficiency of the industrial process, effectively guarantee the quality of the industrial products, and lay a solid foundation for improving the comprehensive economic benefits of enterprises. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The flowchart of the industrial process running state evaluation in the application;

[0052] Figure 2 The running state evaluation structure diagram based on the VAGCN model in the application;

[0053] Figure 3 The state online evaluation result based on the VAGCN model in the application;

[0054] Figure 4 The ROC curve diagram of the online evaluation result of the complex industrial process running state evaluation model based on the VAGCN in the application;

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

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

[0057] Figure 7 This is the dynamic deviation curve of the variable under the influence of the 7 disturbances.

[0058] Figure 8 This is the dynamic deviation curve of the variable under the influence of the 9 disturbances. Detailed Implementation

[0059] The invention will now be further described with reference to the accompanying drawings.

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

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

[0062] A1: First, use several sensors deployed at key nodes in the industrial production system to collect raw data from the actual industrial process. Where N is the number of sampling points, X = [x1, x2, ..., x N [This represents industrial process data, Y = [y1, y2, ..., y]] N [This refers to the corresponding comprehensive economic indicator data;]

[0063] A2: Preprocess the original data. First, remove and align the abnormal data in the original data. Then, perform maximum and minimum standardization on the X and Y data according to formula (1) so that the mean of the calculated data is 0 and the standard deviation is 1.

[0064]

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

[0066] A3: First, select suitable comprehensive economic indicator data from the preprocessed data based on the specific production process. Then, combining expert knowledge, divide the selected data into training and test datasets proportionally according to different industrial operation status levels. Subsequently, label the different status level categories with numerical labels to obtain the true label data y. L ;

[0067] A4: Introduce a time series slicing technique with a fixed length of H to segment the training dataset, resulting in the following dataset: The data for each time slice is as follows: Each Xt contains time slice samples from N variables, i.e. t represents the window index of the time window, represents 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), learn the features of each variable segment, and extract the local features of each variable, i.e. t ′ ,i = f c (x t,i |W c );

[0069] A6: Use position encoding to supplement the relative position information of different time point graph structures at the same time; for the i-th variable Enhanced features z t,i = f p (t) + x t ′ ,i are obtained by adding position encoding, as shown in equation (2);

[0070]

[0071] where p i is the position encoding, m represents the m-th feature among the variable features; ω k is the frequency;

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

[0073] A8: Use the softmax function to limit the correlation to [0, 1] at the same time; finally obtain the graph structure with different time point and different variable interaction connections, i.e. variable interaction graph structure G t = (Z t , A t ), where A tan adjacency matrix representing the variable interaction graph structure within a time window W(t);

[0074] A9: using a decay matrix c tr,ij = w t-r to adjust the variable interaction correlation graph G t = (Z t , A t ) between different time points, as shown in equation (3);

[0075]

[0076] A10: by multi-layer graph convolution operation, step by step aggregation of neighbor information of nodes; for each input feature within a time window W(t) and the adjacency matrix The update process of the graph convolution layer is shown in equation (4), where represents the initial feature of each node within the time window W(t);

[0077]

[0078] In the formula: 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 l-layer graph convolution, initially W (l) represents the learnable parameter matrix of the l-th layer; σ is a nonlinear activation function;

[0079] A11: using average pooling to summarize the features of different times, after time pooling, the feature matrix is flattened into a vector

[0080] A12: using a fully connected layer (FC layer) to perform linear transformation on the flattened features, and generating the final classification output through an activation function;

[0081] A13: applying the softmax function to calculate the weight probability of the current predicted state belonging to each state, and taking the maximum value of each state probability as the final model prediction state;

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

[0083]

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

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

[0086] A16: The hidden layer output h of the feature extraction model and the binary variable As input, the Softmax classifier is trained to obtain the state recognition model, and the output of the Softmax classifier is obtained according to formula (6);

[0087]

[0088] where c = 1, 2, …, C represents different state levels, and p (y = j | x) represents the posterior probability of input x being state level j, is the parameter of the classifier;

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

[0090]

[0091] A18: Backward fine-tune the classifier model parameters until J softmax (θ) is minimized or J softmax (θ) converges, obtain and save the state recognition model;

[0092] A19: Fine-tune the parameters after cascading the feature extraction model and the classifier model to build a complete running state evaluation offline model;

[0093] Step two: use online data to evaluate the running state;

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

[0095] B2: Slide sampling on the standardized online data with a sliding window of window length H to obtain sequence data of consistent length;

[0096] B3: Input the sequence data into the trained running state evaluation offline model to calculate the online data x t at time t, and the posterior probability of the online data x at time t belonging to different running state levels is The state level corresponding to the maximum value, i.e., the process running state level at time t is

[0097]

[0098] B4: According to the online running state evaluation result, guiding the subsequent optimization control of the industrial process;

[0099] When the running state evaluation offline model evaluates the current running state as non-optimal, non-optimal state variable dynamic offset evaluation is performed, and the specific steps are as follows:

[0100] S1: Extracting the variable interaction graph topological structure feature G in the time window under the non-optimal state non-opt ;

[0101] S2: Using the Frobenius norm to calculate the norm size of the difference matrix AG between it and the optimality graph structure G opt ;

[0102] S3: According to formula (8), the similarity score S is obtained, and the similarity score S is used to measure the closeness of the non-optimal state and the optimal state, and to quantify the variable offset degree under the non-optimal state and the corresponding adjustment strategy, so as to guide the operator to accurately and efficiently adjust the production process;

[0103]

[0104] Embodiment:

[0105] The heavy medium coal preparation process is a typical process industry process running in a harsh open environment, and various disturbances and uncertainty factors exist in the running environment, which often makes the current process monitoring and running state evaluation method not comprehensive in process condition information perception. The present application can more effectively extract the dynamic interaction features between variables, more accurately and robustly evaluate the current running state, aiming at the strong dynamic, strong nonlinearity and dynamic offset problems between variables existing in the data of the slime flotation process.

[0106] The selected industrial process is a 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, medium screen, qualified medium barrel, etc. The coal preparation plant uses an automatic feeder to transport raw coal to the washing process. When passing through the double-layer grading screen, coal particles with a particle size of 25 mm or more are sent to other separation equipment such as a drum separator for separation. Medium-sized raw coal particles (6 mm to 25 mm) are transported to the mixing tank and mixed with the heavy medium suspension and pumped into the heavy medium cyclone. The coal-medium mixture completes the density stratification under the action of the centrifugal field in the cyclone and is discharged from the underflow port and the overflow port, respectively, and sent to the medium screen for medium removal treatment. At the same time, the magnetic separator is used to recover the magnetic weighting agent. The recovered weighting agent and the additional magnetic weighting agent are adjusted in the qualified medium barrel. The input amount of water and weighting agent is adjusted to keep the liquid level of the qualified medium barrel and the density of the heavy medium suspension output within a suitable range, and then transported to the mixing tank 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 strong nonlinear characteristics. In order to ensure the quality of coal, the process parameters must be continuously detected and adjusted in actual operation. According to the process mechanism and actual situation, 11 process variables are determined, and the ash content index is determined as the output variable. The specific process variables are shown in Table 1.

[0108] Table 1: Selection of process operation variables

[0109]

[0110] The data in this example is from the heavy medium coal preparation process full-process simulation platform (registration number: 2020SR1160052). In order to generate heavy medium coal preparation process data of different state levels, three main operating variables in the heavy medium coal preparation process are set in the heavy medium coal preparation process full-process simulation platform: raw coal feeding amount of the coal preparation plant (variable 1), cyclone feeding pressure (variable 7), and heavy medium suspension density output from the qualified medium barrel (variable 9). The variable 1 and variable 9 are generated by introducing fluctuations to simulate the change of state level. The variable generation function is: x = base + dis * (sin(T1πt) + sin(T2πt)), and the variable 7 is generated by introducing fluctuations to simulate the change of state level. The variable generation function is:

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

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

[0113]

[0114] 1) Establishing an offline state evaluation model

[0115] In the data set generated by the mechanism model of the dense medium coal preparation process, 5000 groups of data were selected, and the first 2500 groups of data were divided into offline data as the training data set of the state evaluation model, and the remaining data were used as the test data set for online testing of the model.

[0116] Combined with expert experience, the state level of the coal preparation process was divided according to the ash value in the clean coal product, and the state level division results are shown in Table 3:

[0117] Table 3: State level division and corresponding digital label

[0118]

[0119] The time window is set to 5, and the variable interaction graph structure is constructed in the small window, which can extract local features more finely and reduce the calculation amount. The moving step is set to 1, which ensures that the window will completely cover all time step data each time. Because the time distance between nodes is considered, the decay coefficient is set to 0.4, which directly affects the importance of the interaction between time steps. Because the graph convolution is prone to over-smoothing problem, the experiment only uses two one-dimensional convolution 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 expression ability of the model, and this dimension can improve the model's ability to capture complex patterns, while not causing the model to overfit. The pooling method uses average pooling, the optimizer uses Adam optimizer, and the learning rate is set to 0.001.

[0120] 2) Online testing of the industrial process running state 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, i.e. the 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 running state fluctuates and is in the unstable state transition interval, the judgment result is basically matched with the true result.

[0122] Figure (4) shows the ROC curve of the online evaluation results of the complex industrial process running state evaluation model based on variable interaction graph structure learning. The ROC curve is usually located in the upper left of the coordinate system, and the closer 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 classification effect of the model, and vice versa. As can be seen, the robustness of the present application for each state grade recognition is higher.

[0123] Figure 5 is the confusion matrix of the model. The numbers on the diagonal of the confusion matrix reflect the similarity between the model recognition results and the true state grade. The closer the number on the diagonal to 1, the higher the recognition accuracy of the model. As can be seen from the confusion matrix, the VAGCN model has higher prediction rate, recall rate and evaluation accuracy.

[0124] 3) Non-optimal state dynamic deviation evaluation.

[0125] When the industrial process running state model based on variable interaction graph neural network is non-optimal, the dynamic deviation of the variable needs to be analyzed.

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

[0127] The effectiveness of the present application can be seen from the above simulation examples. Through timely and robust running state evaluation of the slime flotation process, important guidance can be provided for production adjustment of the slime flotation process.

[0128] The application provides a complex industrial process running state evaluation method based on a VAGCN, which first effectively captures the dynamic dependence relationship between different time points and variables by constructing a cross-time and space variable interaction graph structure, greatly improving the accuracy and robustness of the running state evaluation. Under the framework of the variable interaction graph convolution network model, the interactive features between variables are first extracted in combination with the graph convolution neural network (GCN), and the online performance is constrained in real time by introducing the comprehensive economic index (CEI), and a precise and efficient online running state evaluation model is constructed by virtue of the variable interaction features. Further, the graph convolution neural network is used for deep learning of the variable interaction graph structure, in this process, the variable nodes at the previous time can aggregate the multi-variable characteristics at the subsequent time, even in the case of variable offset, the information interaction between variables can still be maintained, greatly alleviating the information missing problem caused by variable offset. In addition, by extracting the variable interaction graph topological structure features in the time window under the non-optimal state, and evaluating the similarity with the optimality graph structure under the optimal state, the dynamic deviation degree of variables under the non-optimal state can be effectively quantified, providing a new idea for the non-optimal state evaluation in the complex industrial system.

[0129] The method has simple implementation process and low implementation cost, can effectively extract variable dynamic deviation information caused by time delay, capture the interaction between variables at different time points, and then timely and accurately evaluate the industrial running state, providing reliable guidance for process optimization and control, helping to realize more accurate process running state evaluation, thereby significantly improving the running efficiency of the industrial process, effectively guaranteeing the quality of the industrial products, and laying a solid foundation for improving the comprehensive economic benefits of enterprises.

[0130] The above is only a preferred embodiment of the application, and does not limit the application in any form, any simple modification and equivalent change made on the basis of the technical essence of the application to the above embodiment fall within the protection scope of the application.

Claims

1. A method for evaluating the running state of a complex industrial process based on VAGCN, characterized in that, Comprising the following steps: Step one: using offline data to train variable interaction graph structure learning feature extraction model and classifier model, and establish state evaluation offline model; A1: first collect the original data of the actual industrial process by using several sensors arranged at key nodes in the industrial production system wherein N is the number of sampling points, X = [x1, x2,..., x N ] is the process data of the industry, and Y = [y1, y2,..., y N ] is the corresponding comprehensive economic index data; A2: preprocessing the original data, first removing abnormal data and aligning the original data, then normalizing the X data and Y data according to formula (1), so that the calculated data has an expectation of 0 and a standard deviation of 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 index data from the pretreated data according to the specific production process, then combine the expert knowledge, divide the selected data into training data set and test data set according to different industrial running state grades, then mark different state grade categories as digital labels in turn to obtain real label data y L ; A4: Introduce a time series slicing technique with a fixed length H, divide the training data set to obtain a data set wherein the data of each time slice is Each X t contains time slice samples from N variables, that is t represents the window index of the time window, represents the number of time slices, A5: using an encoder f c (·|W c ) to process the variable data samples in each window W(t), feature learning is performed on each piece of variable data, and the local features of each variable are extracted, i.e., x t ′ ,i = f c (x t,i |W c ) A6: Simultaneous use of position encoding to complement relative position information of different time point graph structures; for the i-th variable Obtain enhanced features z by adding position encoding t,i = f p (t) + x t ' ,i As shown in equation (2); where p i is the position encoding, m indicates the mth feature among the variable features; ω k is the frequency; A7: The 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 )g s (z r,j ) T where t, r ∈ [1, T], i, j ∈ [1, N], and the function g s (z) = zW s is used to enhance the expressiveness, where W s is a learnable weight; A8: The correlation is limited within [0, 1] by using a softmax function at the same time; finally, a graph structure with interactive connections between different variables at different time points is obtained, that is, a variable interaction graph structure G t = (Z t , A t ), wherein, A t is expressed as an adjacency matrix of the variable interaction graph structure within the time window W(t); A9: Using a decay matrix c tr,ij = w t-r to adjust the variable interaction graph G t = (Z t , A t ) between different time points, as shown in equation (3); A10: Step by step aggregate the neighbor information of nodes by multi-layer graph convolution operation; aggregate the input features in each time window W(t) and adjacency matrix aggregate, the update process of graph convolution layer is shown in formula (4), wherein, denotes the initial features of each node in the time window W(t); In the formula: 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 lth layer graph convolution, and initially W (l) represents the learnable parameter matrix of the lth layer; and σ is a nonlinear activation function. A11: aggregate features at different times using average pooling, after time pooling, the feature matrix is flattened into a vector A12: using a fully connected layer to perform linear transformation on the flattened features, and generating the final classification output through an activation function; A13: applying the softmax function to calculate the weight probability of the current predicted state belonging to each state, and taking the maximum value of each state probability as the final model predicted state; A14: select cross entropy loss function as the optimization objective function in the network training process, as shown in formula (5); In the formula, N represents the total number of labels corresponding to the 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 converges, and then save the classifier model; A16: the hidden layer output h of the feature extraction model and the binary variable As input, the Softmax classifier is trained to obtain a state recognition model, and the output of the Softmax classifier is obtained according to formula (6); where c = 1, 2,..., C represents different state levels, p(y = j | x) represents the posterior probability of input x being state level j, is a parameter of the classifier; A17: Calculate the classifier loss function J according to formula (7) softmax (θ); A18: Backward fine-tuning the classifier model parameters until J is minimized softmax (θ) or convergence of J softmax (θ) is achieved, resulting in and saving the state recognition model; A19: fine-tune the parameters after cascading the feature extraction model and the classifier model to build a complete offline running state evaluation model; Step two: using online data to evaluate the running state; B1: use several sensors arranged at key nodes in the industrial production system to perform online sampling to obtain online process data X, and standardize X; B2: sliding window sampling of the standardized online data with window length H to obtain sequence data with consistent length; B3: input the sequence data into the trained running state evaluation offline model to calculate the online data x at time t t The posterior probability of different running state grades is Wherein, i∈{1,2,...,q}, q is the numerical label representing the state grade category divided according to the comprehensive economic index data, and the process running state at the final time t is defined as the posterior probability set The state grade corresponding to the maximum value in the maximum value of the posterior probability set is the process running state grade at time t B4: according to the online running state evaluation result, guide the subsequent optimization and control of the industrial process; When the running state evaluation offline model evaluates the current running state as non-optimal, perform non-optimal state variable dynamic offset evaluation, the specific steps are as follows: S1: extract the variable interaction graph topology structure features G in the time window under the non-optimal state non-opt ; S2: Compute the norm size of the difference matrix AG between it and the optimality graph structure G using the Frobenius norm; opt between it and the optimality graph structure G using the Frobenius norm; S3: according to formula (8), obtain the similarity score S, and use the similarity score S to measure the closeness of the non-optimal state and the optimal state, and quantify the variable offset degree and the corresponding adjustment strategy in the non-optimal state, to guide the operator to accurately and efficiently adjust the production process;

Images