Fault-tolerant prediction method of silicon content in blast furnace digital twin system based on graph network

Abstract

The present invention discloses a fault-tolerant prediction method for silicon content in a blast furnace digital twin system based on a graph network. The method is divided into three parts: credibility assessment, data quality enhancement, and time scale aggregation. First, in view of the point deviation and sequence drift that may occur in the sensor measurement process, a point-level credibility calculation based on an attention mechanism and a sequence-level credibility assessment based on graph structure comparison are developed, and the identification of low-quality data is realized by integrating the point-sequence evaluation results. Afterwards, data enhancement based on a graph neural network is carried out in a targeted manner, and the low-quality data is enhanced using the features of adjacent sensors. Finally, time scale aggregation is used to mine the dynamics in the data and complete the prediction of silicon content. Through the two steps of "evaluation-enhancement", this method realizes the fault-tolerant prediction of the model in the case of inaccurate sensor measurement, improves the method's ability to resist interference, makes it more adaptable in complex industrial environments, and improves the accuracy of the method.

Description

Technical Field

[0001] The present invention belongs to the field of soft sensor modeling in process industries, and in particular relates to a fault-tolerant prediction method for silicon content in a blast furnace digital twin system based on a graph network, which solves the problem of prediction caused by inaccurate sensor measurement. Background Art

[0002] Blast furnace ironmaking plays an important role in the development of the national economy and production construction. It is very important to ensure the smooth and high-quality operation of the blast furnace ironmaking process. The silicon content in the molten iron is closely related to the blast furnace condition and product quality. By observing the changes in the silicon content in the molten iron, the current reaction status inside the blast furnace can be inferred, so that the operator can adjust the relevant operating parameters in time to keep the blast furnace in a relatively ideal state. In view of the limitations of current measurement methods, the silicon content in actual production can generally only be obtained by manual measurement. This method not only cannot guarantee timeliness, but also causes a waste of human resources. With the development of intelligent algorithms, data-driven soft measurement prediction methods have emerged. Rich process variable data has been accumulated in the blast furnace production process, reflecting the operation status of the physical production line. As a virtual-real interactive system, the blast furnace digital twin system collects ironmaking entity data, establishes a virtual simulation model, and can complete the tasks of remote monitoring, optimization control, and interactive feedback. According to the data of the digital twin system data platform, the mapping relationship between process variables and silicon content is constructed, and then the silicon content prediction based on data is realized. The prediction results are uploaded to the data platform for reference by operators and interact with the services of other digital twin systems.

[0003] However, the ironmaking environment of blast furnaces is harsh, and sensors are exposed to high temperature, high pressure, gas-solid-liquid mixture, strong dust, and multiple interferences for a long time. The measurement accuracy is easily affected, making the measurement results inaccurate. In addition to the more obvious outliers that can be identified and processed by the outlier processing algorithm, there are many deviations that are difficult to identify, which are manifested as values ​​within a reasonable range of variables, but the measurements are inaccurate. For example, noise is scattered throughout the data set and lasts for a period of time; in addition to this, there is also long-term drift. For example, after interference is attached to the outer layer of the sensor, the measurement value has a long-term deviation, but this deviation is difficult to distinguish from the new working conditions. These inaccurate measurements are mixed in the data set, which reduces the quality of the data and introduces a lot of interference, which brings challenges to the training of prediction methods.

[0004] Previously, some scholars have conducted research on the noise problem and achieved noise reduction from the perspectives of loss function, probability distribution, etc. However, little attention has been paid to the long-term drift problem, and no explicit judgment of data quality has been made, ignoring the targeted improvement of data quality. This invention fully considers instantaneous noise and long-term drift, and designs a fault-tolerant prediction method that can not only adaptively identify deviations, but also achieve data enhancement for low-quality data. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a fault-tolerant prediction method for silicon content in a blast furnace digital twin system based on a graph network.

[0006] The technical solution of the present invention is as follows:

[0007] A fault-tolerant prediction method for silicon content in a blast furnace digital twin system based on a graph network, the steps are as follows:

[0008] Step 1: Based on the established blast furnace digital twin system, obtain the blast furnace ironmaking process variable data from the system's data platform, and use the time window to divide the original sequence into time series as the input of the prediction method;

[0009] Step 2: Aiming at the point deviation and sequence drift that may occur during the sensor measurement process, point-sequence evaluation is carried out to adaptively identify point deviation and sequence deviation. When a sensor makes an inaccurate measurement, the credibility of the measurement value decreases. Point deviation identification refers to the use of point-by-point analysis based on the attention mechanism to quantify the difference between sample points in the variable, and then evaluate the credibility of the point. Sequence deviation identification is achieved by comparing the graph structure.

[0010] Step 3: Use the process variable relationship learning module to obtain the dependency relationship between variables and express it as a process variable graph structure;

[0011] Step 4: Graph structure comparison compares the differences between variables and completes the sequence credibility assessment; then, the point-sequence assessment results are merged to enable the method to fully identify the sensor's measurement noise and drift;

[0012] Step 5: Carry out targeted data quality improvement based on the above evaluation results. When the data of a certain variable is unreliable, use the message propagation in the graph neural network to transmit reliable information to the unreliable variable under the guidance of the adjacency matrix. Under the representation of the message propagation operator, the unreliable data is enhanced.

[0013] Step 6: Based on the enhanced data, multi-channel gated sequential convolution is used on a time scale to complete dynamic feature extraction and finally predict the silicon content; the predicted silicon content data is uploaded to the digital twin platform.

[0014] Step 1: Obtain data of 10 process variables, including furnace gas volume, blast kinetic energy, coal injection volume, oxygen enrichment flow rate, cold air pressure, top temperature, top pressure, silicon content at the last moment, theoretical combustion temperature, and cold air flow rate, in the data platform, and use the time sliding window to divide the original data into time series X = {x 1 ,x 2 ,…,x T},in P=10, indicating the dimension of the process variable, T indicating the size of the time window, X being a two-dimensional matrix of P×T, and the predicted value being the silicon content in the molten iron at time T.

[0015] Step 2: Conduct point-to-sequential evaluation to adaptively evaluate the credibility of each variable at each moment. The purpose of point-level credibility is to evaluate The credibility of the measured value of a single process variable at a single moment is set by the credibility coefficient matrix A pnt , the shape is exactly the same as X, and each element in the matrix and One to one correspondence, reflect The coefficient weights the original input X and obtains in Represents the point-by-point multiplication of matrix elements; the credibility coefficient matrix is ​​calculated based on the attention mechanism.

[0016] Step 3: The graph structure between variables is learned through the process variable learning module. Each node in the graph structure represents a process variable, and the edges between variables represent the relationship between process variables. This method uses the improved Gaussian radial basis distance to measure this relationship:

[0017]

[0018] where α pq Represents the weight of the edge between variable p and variable q in the graph structure. z is a trainable matrix. To ensure the sparsity of the graph structure, the first g edges with the largest weights are retained, and the rest are set to 0.

[0019] The task of sequence-level credibility assessment in step 4 is to quantify the entire time series x of a single process variable p The credibility of the graph structure is learned based on the features of the entire sequence, reflecting the dependencies between variables in the entire sequence. Afterwards, the results of the point-level evaluation and the sequence-level evaluation are taken as the union to obtain the point-sequence combined evaluation result.

[0020] Step 5: Carry out targeted data quality improvement based on the above evaluation results. When a variable data is unreliable, use the message propagation in the graph neural network to transmit reliable information for the unreliable variable under the guidance of the adjacency matrix; under the representation of the message propagation operator, the unreliable data is enhanced. The corresponding feature of each node is In order to enhance the representation ability of the feature, this feature is mapped to a high-dimensional space to obtain a high-dimensional feature representation; thereafter, for the data evaluated as unreliable, message passing is carried out to control the proportion of each part of the information in the data enhancement process of the unreliable variable. Through the above process, the unreliable variable selectively introduces the trusted information of the related process variables to achieve data enhancement.

[0021] The beneficial effects of the present invention are as follows: the instantaneous noise and continuous drift generated during the operation of the sensor are fully considered, and solutions are proposed for these two situations. The measurement deviation is explicitly identified through sequence analysis based on point-by-point analysis of the attention mechanism and graph structure comparison, and the dynamics within the variables and the associations between variables are fully utilized. This explicit identification lays the foundation for refined modeling. Based on the above evaluation results, a customized message propagation operator is used to achieve data enhancement of untrusted variables. After the "evaluation-enhancement" steps, the method enhances the inclusiveness and robustness of untrusted data, and the disturbed sensors are more reliably represented, which improves the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A block diagram of the method is shown.

[0023] Figure 2 Shown is the message propagation operator.

[0024] Figure 3 Shown is a dilated convolutional layer.

[0025] Figure 4 Shown is a multi-channel gated sequential convolution module.

[0026] Figure 5 The figure shows the prediction effect of this method on silicon content in molten iron. DETAILED DESCRIPTION

[0027] Firstly, the basic principle, structure and steps of the present invention are described.

[0028] Based on the established blast furnace digital twin system, the blast furnace ironmaking process variable data is obtained from the system's data platform. These data are collected from the blast furnace physical production line and reflect the operation process of the blast furnace. However, due to the harsh on-site environment of the blast furnace, the accuracy of the sensor is easily damaged, and the collected data has noise and drift. These noise and drift are difficult to identify and process, making the input data quality low. The original sequence is divided into time series using a time window as the input of the prediction method. Secondly, facing the point deviation and sequence drift that will occur in the sensor measurement process, a point-sequence evaluation is carried out to adaptively identify the point deviation and sequence deviation. When a sensor measurement is inaccurate, the credibility of the measurement value decreases. Based on this premise, point deviation identification refers to the use of point-by-point analysis based on the attention mechanism to quantify the difference between sample points in the variable, and then evaluate the credibility of the point. Sequence deviation identification is achieved based on graph structure comparison. The process variable relationship learning module is used to obtain the dependency relationship between variables and represent it as a process variable graph structure. The purpose of graph structure comparison is to compare the differences between variables and complete the sequence credibility evaluation. After that, the point-sequence evaluation results are merged so that the method can fully identify the measurement noise and drift of the sensor. Next, we will carry out targeted data quality improvement based on the above evaluation results. When the data of a variable is unreliable, we use the message propagation in the graph neural network to transmit reliable information to the unreliable variable under the guidance of the adjacency matrix. Under the characterization of the message propagation operator, the unreliable data is enhanced. Finally, based on the enhanced data, multi-channel gated temporal convolution is used on the time scale to complete dynamic feature extraction and finally predict the silicon content. The measured silicon content is uploaded to the digital twin data platform for operator reference and interaction with other services. The steps are as follows:

[0029] Step 1: Obtain data of 10 process variables, including furnace gas volume, blast kinetic energy, coal injection volume, oxygen enrichment flow rate, cold air pressure, top temperature, top pressure, silicon content at the previous moment, theoretical combustion temperature, and cold air flow rate, in the data platform, and use the time sliding window to divide the original data into time series X = {x 1 ,x 2 ,…,x T},in P = 10, indicating the dimension of the process variable, T indicating the size of the time window, and X being a two-dimensional matrix of P × T. The predicted value is the silicon content in the molten iron at time T.

[0030] Step 2: Conduct point-sequential evaluation to adaptively evaluate the credibility of each variable at each moment. The purpose of point-level credibility calculation is to evaluate The credibility of (the measured value of a single process variable at a single moment). Set the credibility coefficient matrix A pnt , whose shape is exactly the same as X, and each element in the matrix and One to one correspondence. Can be reflected The coefficient weights the original input X and obtains in Represents the point-by-point multiplication of matrix elements. The credibility coefficient matrix is ​​calculated based on the attention mechanism, and the specific calculation method is:

[0031]

[0032]

[0033] Where W A , U A , K A are all trainable weight matrices. When the credibility coefficient is within the range of η1 to η2, the data is considered credible, otherwise it is unreliable. η1 and η2 are both hyperparameters.

[0034] Step 3: Learn the graph structure between variables through the process variable learning module. Each node in the graph structure represents a process variable, and the edges between variables represent the relationship between process variables. This method uses the improved Gaussian radial basis distance to measure this relationship:

[0035]

[0036] where α pq represents the weight of the edge between variable p and variable q in the graph structure, and z is a trainable matrix. To ensure the sparsity of the graph structure, the first g edges with the largest weights are retained, and the rest are set to 0.

[0037] Step 4: The task of sequence-level credibility assessment is to quantify the entire time series x of a single process variable p The graph structure is learned based on the features of the entire sequence and reflects the dependencies between variables in the entire sequence. If the sensor of a variable has a long-term drift error, the dependency between the variable and other variables will be abnormal. By comparing the similarities and differences between the graph structure learned based on the data and the graph structure in the standard graph library, the untrustworthy sequence and process variables are located. After that, the results of the point-level evaluation and the sequence-level evaluation are taken as the union to obtain the point-sequence combined evaluation result.

[0038] Step 5: Carry out targeted data quality improvement based on the above evaluation results. When the data of a variable is unreliable, use the message propagation in the graph neural network to transmit reliable information to the unreliable variable under the guidance of the adjacency matrix. Under the representation of the message propagation operator, the enhancement of unreliable data is completed. The corresponding feature of each node is In order to enhance the representation ability of the feature, this feature is mapped to a high-dimensional space to obtain a high-dimensional feature representation Where l represents the number of layers of the graph neural network, Rp is a learnable mapping matrix. After that, for data evaluated as untrustworthy, message passing is carried out:

[0039]

[0040]

[0041]

[0042] in, represents the information from the neighboring nodes of process variable p, is a variable screening coefficient, which controls the proportion of each part of information in the process of data enhancement of untrusted variables, q i Represents the neighbor node of the process variable p. Through the above process untrusted variables, the trusted information of related process variables is selectively introduced to achieve data enhancement.

[0043] Step 6 uses multi-channel gated sequential convolution based on the enhanced data to complete dynamic feature extraction and finally predict silicon content. The prediction results of the method are uploaded to the data platform for operators to refer to and interact with other services.

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

[0045] The overall framework of the fault-tolerant prediction method for silicon content in the blast furnace digital twin system based on graph networks is as follows: Figure 1 As shown, the specific steps are as follows:

[0046] Step 1: Obtain data of 10 process variables, including furnace gas volume, blast kinetic energy, coal injection volume, oxygen enrichment flow rate, cold air pressure, top temperature, top pressure, silicon content at the previous moment, theoretical combustion temperature, and cold air flow rate, in the data platform, and use the time sliding window to divide the original data into time series X = {x 1 ,x 2 ,…,x T},in P = 10, indicating the dimension of the process variable, T indicating the size of the time window, and X being a two-dimensional matrix of P × T. The predicted value is the silicon content in the molten iron at time T.

[0047] Step 2: Conduct point-sequential evaluation to adaptively evaluate the credibility of each variable at each moment. The purpose of point-level credibility calculation is to evaluate The credibility of (the measured value of a single process variable at a single moment). Set the credibility coefficient matrix A pnt , whose shape is exactly the same as X, and each element in the matrix and One to one correspondence. Can be reflected The coefficient weights the original input X and obtains in Represents the point-by-point multiplication of matrix elements. The credibility coefficient matrix is ​​calculated based on the attention mechanism, and the specific calculation method is:

[0048]

[0049]

[0050] Where W A , U A , K A are all trainable weight matrices. When the credibility coefficient is within the range of η1 to η2, the data is considered credible, otherwise it is not credible. η1 and η2 are both hyperparameters. The above evaluation results are recorded in the point credibility mask matrix M pnt middle, represent Untrustworthy, It represents credibility.

[0051] Step 3: Learn the graph structure between variables through the process variable relationship learning module. Each node in the graph structure represents a process variable, the edges between variables represent the relationship between process variables, and the node features are the sequence values ​​of the corresponding variables. This method uses the improved Gaussian radial basis distance to measure this relationship:

[0052]

[0053] where α pq represents the weight of the edge between variable p and variable q in the graph structure, and z is a trainable matrix. To ensure the sparsity of the graph structure, the first g edges with the largest weights are retained, and the rest are set to 0.

[0054] Step 4: The task of sequence-level credibility assessment is to quantify the entire time series x of a single process variable p The graph structure is learned based on the features of the entire sequence, reflecting the dependencies between variables in the entire sequence. If the sensor of a variable produces a long-term drift error, the dependency between the variable and other variables will be abnormal. By comparing the similarities and differences between the graph structure learned based on data and the graph structure in the standard graph library, the untrustworthy sequence and process variables are located, and the sequence trust mask matrix M is obtained. srs The matrix is ​​a one-dimensional vector with dimension P, containing the evaluation results of the entire sequence of each variable. srs,p =1, the entire sequence corresponding to process P is unreliable.

[0055] The standard library in this method is also available through Figure 1 The network structure shown in the figure is obtained by training. Since the proportion of untrusted data in the entire data set is relatively low, the probability of abnormal graph structure is also relatively low. Therefore, the graph structure corresponding to each sequence in the training set is retained, and the number of occurrences of each graph structure is counted. The graphs that appear frequently are obtained by converting trusted data, and these graphs are stored in the standard graph library Ω as a comparison standard. Due to the lack of standard graph libraries at this stage, M srs It cannot be obtained, so during the graph structure training, M srs Randomly generate. Then, the results of point-level evaluation and sequence-level evaluation are combined as follows:

[0056] M ps =M pnt |M′ srs

[0057] Among them, M ps is the evaluation result of the terminal-sequence combination, M′ srs It is M srs The result is replicated T times along the time axis so that it is the same dimension as M. pnt to make a match.

[0058] Step 5: Carry out targeted data quality improvement based on the above evaluation results. When the data of a variable is unreliable, the message propagation in the graph neural network is used to transmit reliable information to the unreliable variable under the guidance of the adjacency matrix. The structure of the message propagation operator is as follows: Figure 2 As shown in Figure 2, through information interaction, the enhancement of untrusted data is completed. The corresponding features of each node are In order to enhance the representation ability of the feature, this feature is mapped to a high-dimensional space to obtain a high-dimensional feature representation Where l represents the number of layers of the graph neural network, R p is a learnable mapping matrix. After that, for data evaluated as untrustworthy, message passing is carried out:

[0059]

[0060]

[0061]

[0062] in, represents the information from the neighboring nodes of process variable p, is a variable screening coefficient, which controls the proportion of each part of information in the process of data enhancement of untrusted variables, q i Represents the neighbor node of the process variable p. Through the above process untrusted variables, the trusted information of related process variables is selectively introduced to achieve data enhancement.

[0063] Step 6 uses multi-channel gated temporal convolution to extract dynamic features based on the enhanced data. The basic unit of multi-channel gated temporal convolution is the dilated inception layer (DIL), which is Figure 3 This layer contains s c The convolution kernels of each channel are different. In this way, when the depth of the convolution layer is deepened, the receptive field of various scales can be realized, and the temporal information of different scales can be fully mined. Each layer of multi-channel gated temporal convolution layer consists of two parallel DILs. One DIL is connected in series with the activation function to realize the nonlinearity of the input, and the other DIL is connected in series with the sigmoid function to control the information. The specific details are as follows: Figure 4 As shown. The calculation method of one-dimensional convolution is:

[0064]

[0065] Where c is a one-dimensional time series, representing the information of each process variable, * represents the convolution operation, and F 1×s (t) represents the convolution kernel with a kernel size of 1×s. After the time series aggregation of multi-channel gated time series convolution, the prediction of silicon content is finally completed. This method uses the mean-square error (MSE) constrained training process, and the loss function is defined as:

[0066]

[0067] in, represents the predicted value of sequence i, y i represents the true value of sequence i, and K is the number of sequences in the entire training set.

[0068] This method is used to predict the silicon content of molten iron in the actual blast furnace production process. The prediction results are as follows: Figure 5 As shown in the figure, when there is unreliable input data, the prediction results of the model can still maintain the accuracy that meets the actual production needs, thus proving the effectiveness of the method. The prediction results of the method are also uploaded to the data platform in the data twin system for reference by operators and can also be used for interaction with other services, facilitating the application of the method in actual industrial scenarios.

[0069] The embodiments described above can be further combined or replaced, and the embodiments are only descriptions of preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various changes and improvements made by ordinary technicians in this field to the technical solution of the present invention belong to the protection scope of the present invention. The protection scope of the present invention is given by the attached claims and any equivalent technical solutions thereof.

Claims

1. A fault-tolerant prediction method for silicon content in a blast furnace digital twin system based on a graph network, characterized in that: Here are the steps: Step 1: Based on the established blast furnace digital twin system, obtain the blast furnace ironmaking process variable data from the system's data platform, and use the time window to divide the original sequence into time series as the input of the prediction method; Step 2: Aiming at the point deviation and sequence drift that may occur during the sensor measurement process, point-sequence evaluation is carried out to adaptively identify point deviation and sequence deviation. When a sensor makes an inaccurate measurement, the credibility of the measurement value decreases. Point deviation identification refers to the use of point-by-point analysis based on the attention mechanism to quantify the difference between sample points in the variable, and then evaluate the credibility of the point. Sequence deviation identification is achieved by comparing the graph structure. Step 3: Use the process variable relationship learning module to obtain the dependency relationship between variables and express it as a process variable graph structure; Step 4: Graph structure comparison compares the differences between variables and completes the sequence credibility assessment; then, the point-sequence assessment results are merged to enable the method to fully identify the sensor's measurement noise and drift; Step 5: Carry out targeted data quality improvement based on the point-sequence evaluation results. When the data of a variable is unreliable, use the message propagation in the graph neural network to transmit reliable information to the unreliable variable under the guidance of the adjacency matrix. Under the representation of the message propagation operator, the unreliable data is enhanced. Step 6: Based on the enhanced data, multi-channel gated sequential convolution is used on a time scale to complete dynamic feature extraction and ultimately predict silicon content; the predicted silicon content data is uploaded to the digital twin platform; Step 5: Carry out targeted data quality improvement based on the point-sequence evaluation results. When the data of a variable is unreliable, use the message propagation in the graph neural network to transmit reliable information for the unreliable variable under the guidance of the adjacency matrix. Under the representation of the message propagation operator, the enhancement of untrusted data is completed, and the corresponding characteristics of each node are In order to enhance the representation ability of the feature, this feature is mapped to a high-dimensional space to obtain a high-dimensional feature representation; After that, for the data evaluated as untrustworthy, message passing is carried out to control the proportion of each part of information in the data enhancement process of the untrustworthy variable. Through the message passing process, the untrustworthy variable selectively introduces the trusted information of the related process variables to achieve data enhancement.

2. The method according to claim 1, characterized in that: Step 1: Obtain data of 10 process variables, including furnace gas volume, blast kinetic energy, coal injection volume, oxygen enrichment flow rate, cold air pressure, top temperature, top pressure, silicon content at the last moment, theoretical combustion temperature, and cold air flow rate, in the data platform, and use the time sliding window to divide the original data into time series X = {x 1 ,x 2 ,…,x T },in represents the dimension of the process variable, T represents the size of the time window, X is a two-dimensional matrix of P×T, and the predicted value is the silicon content in the molten iron at time T.

3. The method according to claim 1, characterized in that: Step 2: Conduct point-to-sequential evaluation to adaptively evaluate the credibility of each variable at each moment. The purpose of point-level credibility is to evaluate The credibility of the measured value of a single process variable at a single moment is set by the credibility coefficient matrix A pnt , the shape is exactly the same as X, and each element in the matrix and One to one correspondence, reflect The coefficient weights the original input X and obtains in Represents the point-by-point multiplication of matrix elements; the credibility coefficient matrix is ​​calculated based on the attention mechanism.

4. The method according to claim 1, characterized in that: Step 3: The graph structure between variables is learned through the process variable learning module. Each node in the graph structure represents a process variable, and the edges between variables represent the relationship between process variables. This method uses the improved Gaussian radial basis distance to measure this relationship: where α pq Represents the weight of the edge between variable p and variable q in the graph structure. z is a trainable matrix. To ensure the sparsity of the graph structure, the first g edges with the largest weights are retained, and the rest are set to 0.

5. The method according to claim 1, characterized in that: The task of sequence-level credibility assessment in step 4 is to quantify the entire time series x of a single process variable p The credibility of the graph structure is learned based on the features of the entire sequence, reflecting the dependencies between variables in the entire sequence. Afterwards, the results of the point-level evaluation and the sequence-level evaluation are taken as the union to obtain the point-sequence combined evaluation result.

Images