Knowledge-based timing diagram convolutional neural network blast furnace fault diagnosis method

Abstract

The application discloses a knowledge-based temporal convolution neural network (KB-TGCN) blast furnace fault diagnosis method. The method can solve the space-time dependence and sample imbalance problem of the blast furnace ironmaking process. First, the calculated variables that can further reflect the process state are calculated by using the variables directly detected by the sensor. Then, the knowledge-based graph structure is constructed according to the spatial position and calculation relationship of the variables. Subsequently, a one-dimensional time information extraction module is embedded in the graph convolutional neural network. Therefore, the TGCN can capture time information while maintaining the original spatial relationship. By using the knowledge-based graph structure, the KB-TGCN can fully mine the spatial information of different blast furnace positions. In addition, the method uses a focal loss function with an adaptive balance factor instead of a traditional cross-entropy loss function to overcome the sample imbalance problem.

Description

Technical Field

[0001] This invention belongs to the field of industrial process diagnosis, specifically relating to a knowledge-based time-series graph convolutional neural network-based method for blast furnace fault diagnosis. Background Technology

[0002] The steel industry is a vital pillar of my country's modern industry and national economy, and an integral part of the national economy. Blast furnace ironmaking is the core of the entire steel industry and a crucial process for energy and mass conversion in steel manufacturing, accounting for up to 70% of total energy consumption in steel production. Ensuring the safe operation of the blast furnace ironmaking process is of paramount importance for the steel industry's deep energy conservation, emission reduction, and quality and efficiency improvement. An accurate fault diagnosis model can promptly help operators identify faults, adjust furnace conditions in a timely manner, and prevent dangerous situations from occurring.

[0003] In blast furnace ironmaking, abnormal furnace conditions frequently occur due to fluctuations in raw material quality, human error, and equipment malfunctions. These abnormalities often lead to increased fuel ratios, longer shutdown and maintenance times, and substandard molten iron quality, causing significant resource and equipment losses, reducing furnace lifespan, and potentially even triggering accidents resulting in injuries or fatalities. Therefore, improving the safety of the blast furnace ironmaking process, especially reducing the frequency of accidents and abnormal conditions, is crucial for reducing energy consumption, ensuring equipment and personnel safety, and improving the economic efficiency of steel production. However, the complex blast furnace production process and reactions result in coupled spatiotemporal relationships between variables, which are difficult to analyze using existing fault diagnosis methods. Furthermore, the lack of fault samples and labels further exacerbates the diagnostic challenges. With the continuous improvement of industrial informatization, data-driven process diagnosis has become a key technology for improving the safety, quality, and operational efficiency of process industries. How to diagnose blast furnace faults under complex operating conditions has become a cutting-edge and challenging research area in metallurgical technology worldwide. Summary of the Invention

[0004] This invention addresses the existing problems in blast furnace fault diagnosis by proposing a knowledge-based temporal graph convolutional neural network (KB-TGCN) method for blast furnace fault diagnosis. The method constructs a KB-TGCN to mine the complex spatiotemporal dependencies in the blast furnace ironmaking process and further solves the sample imbalance problem by proposing an adaptive balance factor FL loss function.

[0005] A knowledge-based temporal graph convolutional neural network (TGCN) method for blast furnace fault diagnosis is proposed. This method constructs a knowledge-based graph structure based on the spatial location and computational relationships of variables. Subsequently, a one-dimensional temporal information extraction (TIE) module is embedded into the graph convolutional neural network. Therefore, TGCN can capture temporal information while preserving the original spatial relationships. Utilizing the knowledge-based graph structure, KB-TGCN can fully mine the spatiotemporal information of different blast furnace locations. Furthermore, a focal loss function (FL) with an adaptive balance factor is used instead of the traditional cross-entropy loss function to overcome the sample imbalance problem.

[0006] The knowledge-based graph structure constructed based on the spatial location and computational relationships of variables is as follows:

[0007] The process variables of the blast furnace are divided into two types: sensor variables and calculated variables. Sensor variables are variables directly measured by sensors, while calculated variables are variables calculated based on multiple sensor variables according to calculation formulas.

[0008] When constructing a graph structure, the existence of edges between sensor variables is determined by the distance between their actual locations. For example, sensors simultaneously located on the furnace top are considered to have an edge relationship, as are sensors on the same blast furnace bottom path. Variables corresponding to sensors that are far apart are considered to have no edge relationship. For computational variables, the existence of edges is determined by computational relationships; each computational variable has an edge relationship with every variable directly used in its computation. Based on this principle, a knowledge-based graph structure can be constructed.

[0009] The module for embedding one-dimensional temporal information extraction (TIE) into the graph convolutional neural network is as follows:

[0010] In TGCN, each node represents a different variable, and for each node, there exists a separate one-dimensional time information extraction module. This module structure can be composed of a recurrent neural network (RNN). Assume the node variable is x = {x...} t-T ,x t-T+1 ,…,x t Considering a time step length of T+1, using an RNN to extract time information from this one-dimensional variable, we can obtain the corresponding latent variable {z}. t-T ,z t-T+1 ,…,z t}. Then, by using a fully connected neural network to perform a weighted summation of the latent variables at different time scales, we obtain Z = w·[z]. t-T ,z t-T+1 ,…,z t]+b, the final node's state is Z o =Z+x t .

[0011] The TGCN described above can capture temporal information while maintaining the original spatial relationships.

[0012] The time information extractor extracts time information at the nodes, thus not disrupting the spatial relationships of the original variables. Subsequently, it is only necessary to process the node state Z after the time information has been extracted. o Based on the constructed graph structure, spatial information is aggregated using a k-layer GCN, and finally, new node states can be obtained.

[0013] The focal loss (FL) with adaptive balancing factor is as follows:

[0014] An adaptive balancing factor was designed based on the traditional FL loss function. The FL loss function is as follows:

[0015] FL(p i )=∑-α*h*(1-p i ) γ log(p i )

[0016] Where p i Let α be the probability that the i-th sample belongs to a different class, α be the weight of the different classes, h be the one-hot encoding of the label, and γ be adjusted according to the classification difficulty of the sample. The adaptive balancing factor mentioned here is the adaptive determination of α in the loss function, as shown in the following formula.

[0017]

[0018] Where α i It is the weight of the i-th category, N i R is the number of samples in the i-th category, N is the total number of samples in all categories, and r i This represents the proportion of samples in the i-th category.

[0019] The beneficial effects of this invention are as follows:

[0020] 1. This is an end-to-end spatiotemporal fault diagnosis method that can solve the complex spatiotemporal dependencies and sample imbalance problems in the blast furnace ironmaking process.

[0021] 2. Knowledge-based graph structures allow information about variables to flow more accurately at different locations within the blast furnace, thus enabling better diagnosis of fault types.

[0022] 3. In the KB-TGCN network, time information extraction and spatial information extraction are separated and will not interfere with each other.

[0023] 4. The adaptive balancing factor of the FL loss function can adaptively determine the class weights without requiring manual selection. Attached Figure Description

[0024] Figure 1 A graph structure for sensor variables;

[0025] In this context, a line connecting variables represents an edge relationship.

[0026] Figure 2 A graph structure for all variables;

[0027] In this context, a line connecting variables represents an edge relationship.

[0028] Figure 3 This is a diagram of the overall framework of the algorithm.

[0029] Figure 4 The confusion matrix represents the classification results of the proposed method. Detailed Implementation

[0030] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0031] A knowledge-based temporal graph convolutional neural network (TGCN) method for blast furnace fault diagnosis is proposed. This method constructs a knowledge-based graph structure based on the spatial location and computational relationships of variables. Subsequently, a one-dimensional temporal information extraction (TIE) module is embedded into the graph convolutional neural network. Therefore, TGCN can capture temporal information while preserving the original spatial relationships. Utilizing the knowledge-based graph structure, KB-TGCN can fully mine the spatiotemporal information of different blast furnace locations. Furthermore, a focal loss function (FL) with an adaptive balance factor is used instead of the traditional cross-entropy loss function to overcome the sample imbalance problem.

[0032] The knowledge-based graph structure constructed based on the spatial location and computational relationships of variables is as follows:

[0033] The process variables of the blast furnace are divided into two types: sensor variables and calculated variables. Sensor variables are variables directly measured by sensors, while calculated variables are variables calculated based on multiple sensor variables according to calculation formulas.

[0034] When constructing a graph structure, the existence of edges between sensor variables is determined by the distance between their actual locations. For example, sensors simultaneously located on the furnace top are considered to have an edge relationship, as are sensors on the same blast furnace bottom path. Variables corresponding to sensors that are far apart are considered to have no edge relationship. For computational variables, the existence of edges is determined by computational relationships; each computational variable has an edge relationship with every variable directly used in its computation. Based on this principle, a knowledge-based graph structure can be constructed.

[0035] The module for embedding one-dimensional temporal information extraction (TIE) into the graph convolutional neural network is as follows:

[0036] In TGCN, each node represents a different variable, and for each node, there exists a separate one-dimensional time information extraction module. This module structure can be composed of a recurrent neural network (RNN). Assume the node variable is x = {x...} t-T ,x t-T+1 ,…,x t Considering a time step length of T+1, using an RNN to extract time information from this one-dimensional variable, we can obtain the corresponding latent variable {z}. t-T ,z t-T+1 ,…,z t}. Then, by using a fully connected neural network to perform a weighted summation of the latent variables at different time scales, we obtain Z = w·[z]. t-T ,z t-T+1 ,…,z t ]+b, the final node's state is Z o =Z+x t .

[0037] The TGCN described above can capture temporal information while maintaining the original spatial relationships.

[0038] The time information extractor extracts time information at the nodes, thus not disrupting the spatial relationships of the original variables. Subsequently, it is only necessary to process the node state Z after the time information has been extracted. o Based on the constructed graph structure, spatial information is aggregated using a k-layer GCN, and finally, new node states can be obtained.

[0039] The focal loss (FL) with adaptive balancing factor is as follows:

[0040] An adaptive balancing factor was designed based on the traditional FL loss function. The FL loss function is as follows:

[0041] FL(p i )=∑-α*h*(1-pi ) γ log(p i )

[0042] Where p i Let α be the probability that the i-th sample belongs to a different class, α be the weight of the different classes, h be the one-hot encoding of the label, and γ be adjusted according to the classification difficulty of the sample. The adaptive balancing factor mentioned here is the adaptive determination of α in the loss function, as shown in the following formula.

[0043]

[0044] Where α i It is the weight of the i-th category, N i R is the number of samples in the i-th category, N is the total number of samples in all categories, and r i This represents the proportion of samples in the i-th category.

[0045] Example

[0046] 1. Introduction to blast furnace process variables

[0047] This experiment focused on a 2650 cubic meter blast furnace of a steel group. Data from the furnace body were sampled every 10 seconds on average, containing 13 variables, as shown in Table 1.

[0048] Table 1. List of variables in the dataset

[0049]

[0050] 2. Knowledge-based graph structure construction

[0051] Sensor variables:

[0052] Oxygen-enriched flow rate, oxygen-enriched pressure, blast humidity, and hot blast pressure are measured by different sensors on the same link. Therefore, they are physically close together and are considered interconnected, thus having an edge relationship. On the other hand, top temperature and top pressure are measured at the top of the blast furnace, so there is also an edge between them. This results in a graph structure of the sensor variables, as follows: Figure 1 As shown.

[0053] The calculated variables are derived from different sensor data. Therefore, the calculated variables typically follow a specific formula, although this may vary slightly between different steel mills. The formulas for the six calculated variables are as follows.

[0054]

[0055] TPD = HAP-TP

[0056]

[0057]

[0058]

[0059]

[0060]

[0061] In the formula V B V represents the blast furnace air volume, a represents the oxygen content in the oxygen, and EOR represents the oxygen enrichment rate. Table 1 provides the abbreviations for the variables, which will be used directly thereafter. In the formula, V... BG P is the amount of gas in the furnace belly. C denoted as pulverized coal injection rate, H as hydrogen content in the pulverized coal, and d as blast furnace hearth diameter. F represents the total blast furnace tuyeres area, n as the number of tuyeres, t as air temperature, and p0 as atmospheric pressure.

[0062] Furthermore, the formula for calculating TFT is complex, but TFT is only directly related to EOR and EOF. Therefore, based on the above equations, we can obtain the relationships between the variables and obtain the final graph structure, such as... Figure 2 As shown.

[0063] 3. Algorithm Framework

[0064] The framework of the entire algorithm is as follows Figure 3 As shown, firstly, blast furnace sensors collect process variables. Then, based on empirical formulas from on-site experts, computational variables that further reflect the blast furnace's operating status are calculated. Based on the sensor and computational variables, a knowledge-based graph structure is constructed using spatial location and computational relationships. The KB-TGCN architecture is then used to mine the spatiotemporal dependencies of the variables, ultimately obtaining updated node attributes. Finally, the node attributes are used for classification training through fully connected layers and Focal loss to obtain a classification model.

[0065] 4. Final system output:

[0066] The training and testing data are shown in Table 2.

[0067] Table 2. Dataset Overview

[0068]

[0069] This embodiment uses four different metrics—accuracy, precision, recall, and F1-score—to evaluate the method. The calculation methods and meanings of these four metrics are shown in Table 3.

[0070] Table 3. Introduction to Evaluation Indicators

[0071]

[0072] The classification accuracy of the final method is shown in Table 4.

[0073] Table 4 Classification Results

[0074] method Accuracy Precision Recall F1 score SVM 0.9002 0.8935 0.9152 0.8950 MLP 0.8840 0.9086 0.8547 0.8598 TCN 0.9078 0.9049 0.8779 0.8855 GCN 0.9153 0.9190 0.8843 0.8927 KB-TGCN 0.9783 0.9789 0.9610 0.9689

[0075] As shown in the figure, the proposed method achieves higher classification accuracy and better classification results compared to ordinary methods. The classification confusion matrix of the proposed method is as follows: Figure 4 As shown, it can be seen that high classification accuracy was achieved in the classification results of different categories.

[0076] The embodiments described above can be further combined or replaced, and these embodiments are merely descriptions of preferred embodiments of the present invention, not limitations on the concept and scope of the present invention. Various changes and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the inventive concept are all within the protection scope of the present invention. The protection scope of the present invention is given by the appended claims and any equivalents.

Claims

1. A knowledge-based temporal graph convolutional neural network-based method for blast furnace fault diagnosis, characterized in that, Based on the spatial location and computational relationship of variables, a knowledge-based graph structure is constructed; then, a one-dimensional temporal information extraction module is embedded in the graph convolutional neural network; while preserving the original spatial relationship, temporal information is captured; using the knowledge-based graph structure, KB-TGCN mines the spatiotemporal information of different blast furnace locations; and a focus loss function FL with an adaptive balance factor is used to overcome the sample imbalance problem. The knowledge-based graph structure constructed based on the spatial location and computational relationships of variables is as follows: The process variables of the blast furnace are divided into two types: sensor variables and calculated variables. Sensor variables are variables directly measured by sensors, while calculated variables are variables calculated based on multiple sensor variables according to calculation formulas. When constructing the graph structure, sensor variables are determined by the distance between actual sensor locations to determine whether there are edges between variables, while computational variables are determined by computational relationships to determine whether there are edges between them. Computational variables have edges with each variable directly used in their computation; thus, a knowledge-based graph structure is constructed. The module for extracting one-dimensional time information embedded in the graph convolutional neural network is as follows: In TGCN, each node represents a different variable, and for each node, there exists a separate one-dimensional time information extraction module. This module is structured as a recurrent neural network (RNN). Let the node variable be... The time step length under consideration is By using an RNN to extract time information from this one-dimensional variable, the corresponding latent variables can be obtained. Subsequently, a fully connected neural network was used to weight and sum the latent variables at different time scales to obtain... The final node's state is ; The focus loss function with adaptive balance factor is: First, the FL loss function is: ； in For the first The probability that a sample belongs to a different category. These are the weights of different categories. It is the one-hot encoding of the tag. The weights are adjusted based on the classification difficulty of the samples; the adaptive balancing factor is a component of the loss function. The adaptive determination is shown in the following formula. ； in It is the first The weights of each category, It is the first Number of samples in each category It is the total number of samples across all categories. Then it is the first The percentage of samples in each category.

2. The method according to claim 1, characterized in that, TGCN captures temporal information while preserving the original spatial relationships: The time information extractor extracts time information at the nodes and then updates the node status after the time information has been extracted. Based on the graph structure described above, using The layer's GCN aggregates spatial information and finally obtains the new node state.

Images