MDGCN-based blast furnace ironmaking process fault diagnosis method

Abstract

The invention relates to the field of blast furnace ironmaking process monitoring and fault diagnosis, and aims to provide a blast furnace ironmaking process fault diagnosis method based on MDGCN. In the off-line training stage, using sample data to construct a sample matrix, and considering that abstract nodes without actual observation values exist to construct a data moment; calculating an adjacent matrix according to the production process topological structure construction diagram; inputting the feature matrix and the constructed graphs into an MDGCN model for training, and learning features of different levels from each graph; sample data of the blast furnace ironmaking system are collected in real time in the online monitoring stage, and then a new data matrix is reconstructed by referring to the offline training stage; and updating the data matrix by using the graph and the adjacent matrix, inputting the MDGCN model, and predicting whether a fault exists or not and the type of the fault. According to the method, dynamic features of process data can be extracted, and features of different levels are learned by constructing a multi-channel GCN model; and fault diagnosis is carried out by adopting the multi-layer perceptron classifier, so that the accuracy of fault diagnosis is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of blast furnace ironmaking process monitoring and fault diagnosis, and in particular to a blast furnace ironmaking process fault diagnosis method based on MDGCN. Background Art

[0002] The blast furnace ironmaking process is one of the most complex industrial processes currently, and mainly includes five major subsystems: blast furnace, feeding, hot air, coal powder injection, and gas treatment to generate molten iron. In the blast furnace ironmaking process, raw materials and fuel, hot air, and coal powder are respectively fed into the blast furnace from the top and tuyere, and complex physical and chemical reactions occur under high temperature and high pressure. Due to the harsh working conditions, the blast furnace ironmaking process often encounters various abnormal conditions such as difficult furnace conditions, collapsed materials, suspended materials, and furnace temperature heating / cooling. If it is not discovered and diagnosed in a timely and accurate manner, it will not only lead to a decline in the quality of molten iron, but may even threaten the safety of equipment and personnel. Therefore, fault diagnosis is crucial to ensure the safe operation of the blast furnace ironmaking process.

[0003] In recent years, with the increasing requirements for the quality of steel products, R&D personnel have proposed a variety of solutions for fault diagnosis in the blast furnace ironmaking process, trying to detect relevant abnormal conditions in a timely manner through fault monitoring and diagnosis, so as to make timely adjustments to ensure the quality of molten iron and equipment safety. For example, the industry commonly uses fault diagnosis methods such as support vector machine (SVM), random forest (RF), deep convolutional neural network (DCNN), knowledge graph convolutional network (k-GCN), parallel time convolutional network (PTCN), etc., but they all have low accuracy due to their own defects.

[0004] Some researchers have proposed to use graph convolutional neural networks (GCN) for fault diagnosis research. For example, the public documents "Knowledge-based Time-series Graph Convolutional Neural Network Blast Furnace Fault Diagnosis Method" (CN118245937A) and "Knowledge-based Time-series Graph Convolutional Neural Network Blast Furnace Fault Diagnosis Method" (CN118245937A) all involve this. However, the use of graph convolutional neural networks in these technical solutions ignores system structure and process knowledge, which makes GCN still a black box model and has no interpretability. Secondly, GCN has the effect of low-pass filtering. Inappropriately increasing the number of GCN layers will aggravate the network's smoothing effect on the input. And the shallower network structure cannot learn deep feature representations. Finally, although the conventional GCN method aggregates information in the time dimension. But from the perspective of dynamics, it is obviously unreasonable to have the same feature weights at different time nodes. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the deficiencies in the prior art and provide a blast furnace ironmaking process fault diagnosis method based on MDGCN.

[0006] To solve the technical problem, the solution of the present invention is:

[0007] A new fault diagnosis method for blast furnace ironmaking process based on MDGCN is provided, which consists of an offline training phase and an online monitoring phase;

[0008] The offline training phase includes:

[0009] (1.1) Collect sufficient sample data of main process variables for the blast furnace ironmaking system under normal operation and fault conditions;

[0010] (1.2) Standardize the data and construct the sample matrix X, where the sample matrix at time t is represented as X t ; According to the dynamic characteristics of the blast furnace ironmaking process, different sample matrices X t Assign dynamic weights to each and get the data matrix X t，d ; Considering the existence of abstract nodes without actual observation values, reconstruct the data matrix X t，d And mark the true label to get the feature matrix X used as input f，t ;

[0011] (1.3) Based on the topological structure of the blast furnace ironmaking process, calculate the adjacency matrix of each graph;

[0012] (1.4) Substitute the feature matrix X in step (1.2) f，t The graph constructed in (1.3) is used as input to the MDGCN model, and trained using the Adam optimizer and dropout strategy; the MDGCN model has multiple isomorphic graph channels, and learns different levels of features from each graph;

[0013] The online monitoring phase includes:

[0014] (2.1) Collecting sample data of the blast furnace ironmaking system in real time during operation, and then processing the sample data according to the operation in step (1.2) in the offline training phase to reconstruct a new data matrix;

[0015] (2.2) Use the graph and adjacency matrix in step (1.3) to update the data matrix in step (2.1) to obtain the feature matrix

[0016] (2.3) The feature matrix obtained in step (2.2) The MDGCN model is input for prediction. During this process, the multi-layer perceptron classifier MLP is used for processing and the results are input into the Sofimax layer. Finally, the output of the Softmax layer is used to determine whether there is a fault in the blast furnace ironmaking production process and the type of fault.

[0017] As a preferred solution of the present invention, in step (1.2), different sample matrices X are obtained based on Euclidean distance calculation. t The dynamic weights are as follows:

[0018] (a) Assume that the sample matrix X = [x1, x2…x N ] T The data are collected from N time points and normalized; the normalization formula is as follows:

[0019]

[0020] Where, X normalized refers to the normalized data matrix; X healthy It is a data matrix of samples collected during normal operation of the blast furnace ironmaking system; and S healthy They are X healthy The mean and variance matrices of ;

[0021] (b) Assuming that each set of x observations consists of m process variables, the sample matrix at time t is expressed as but:

[0022] X t =[x t-s+1 , x t- s +2 …x t ] T

[0023] Where s is the time delay; T represents the transpose;

[0024] (c) To maintain universality, the calculated Euclidean distance is normalized using the Softmax function; in the sample matrix X t In, x i and x j The Euclidean distance between is defined as:

[0025] d ij =||x i -x j ||2

[0026] In the formula, x i and x j Point to X t The i-th and j-th vectors in , i and j represent different time lags respectively;

[0027] Then, the weight matrix is defined as:

[0028]

[0029] In the formula, Yes ij The average value of Represents a distance metric based on mean normalization.

[0030] As a preferred solution of the present invention, in step (1.2), different sample matrices X t The dynamic relationship between them is determined by the weight matrix M, on which the local dynamic characteristics of the process data are extracted; after calculating the weight matrix M for each group of sampled data, the original sample matrix X t is updated to the data matrix X t，d ,

[0031] X t，d =MX t

[0032] Where, X t represents the sample matrix obtained after preliminary preprocessing of the sampled data; M represents the weight matrix; X t，d Represents the data matrix obtained after calculating the dynamic weights.

[0033] As a preferred embodiment of the present invention, in step (1.2), there are abstract nodes without actual observation values ​​in the topological structure of the blast furnace ironmaking production process, and unit nodes will also be generated in the graph after the graph is constructed; define the zero matrix As the data matrix of all abstract nodes, the observation values ​​of all abstract nodes are initialized to zero, where p represents the number of abstract nodes; then the data matrix X containing dynamic features t，d Horizontally concatenate the feature matrix O representing the abstract node to obtain the feature matrix X f，t ; The specific formula is as follows:

[0034] X t，f =[X t，d |O] T

[0035] Where T represents transpose.

[0036] As a preferred solution of the present invention, in step (1.2), the fault diagnosis of the MDGCN model is implemented through a graph-level multi-classification task. To meet the requirements of supervised learning, for each feature matrix X f，t Assign the corresponding true label at time t y t The value is an integer from 0 to c, which is used to represent the different operating states of the blast furnace ironmaking system.

[0037] As a preferred solution of the present invention, when constructing the graph in step (1.3), the edges between the nodes in the graph are determined based on the process flow content to represent the relationship between the variables; each sample matrix corresponds to a graph, and all graphs will be used for graph-level classification tasks during the operation of the MDGCN model.

[0038] As a preferred embodiment of the present invention, the graph constructed in step (1.3) is an undirected graph, and its adjacency matrix A is a symmetric matrix; each data matrix in step (1.2) corresponds to a graph, and the adjacency matrix of each graph is Sum degree matrix They are shown as follows:

[0039]

[0040] Where A represents the adjacency matrix (n×n) in step (1.3); I n represents an n-order identity matrix; i and j represent two node indices; n represents the total number of nodes.

[0041] As a preferred solution of the present invention, in step (1.4), the MDGCN model is used to realize the learning of the multi-channel graph neural network, which specifically includes:

[0042] (a) Initialize the L convolutional layers and q filters in the model, and then calculate each layer as follows:

[0043]

[0044] Wherein, ReLU(·) represents the rectified linear unit (ReLU); represents the degree matrix with self-directed edges; represents the adjacency matrix after adding self-directed edges; 1 = 1…, L, and They are the feature matrices X f，t At time t, the input features and output features of the jth filter corresponding to the lth graph convolutional layer; is the weight feature that the model needs to learn, s represents the time lag, and b is the feature dimension that sets the embedding;

[0045] (b) After all graphs have completed the convolutional layer calculation, the deep representation h of the data learned by all filters is extracted j ;

[0046]

[0047] Among them, flat(·) represents the flatten operation, that is, flattening the high-dimensional feature matrix into a one-dimensional vector; q represents the number of filters;

[0048] Stack each independent and flattened feature into H:

[0049] H=[h1,h2,...,h q ] T

[0050] Where T represents transpose;

[0051] (c) The stacked H is used as the input of the k-layer multi-layer perceptron (MLP) classifier to perform calculations on the MLP fully connected layer:

[0052] H k =ReLU(w k H k-1 +b k )

[0053] z K =[z0,z1,...,z c ] T

[0054] Among them, H k represents the activation value of the kth layer; w k represents the weight matrix of the kth layer; b k represents the bias vector of the kth layer; k = 1, ..., K, K represents the number of fully connected layers; z k is the output of the kth layer in the MLP fully connected layer; c represents the number of categories in the classification task;

[0055] (d) Fault diagnosis is implemented as a classification task; for this purpose, a Softmax layer is used after the kth fully connected layer:

[0056] p=[p0,p1,…,p c ] T =softmax(z K )

[0057] The output of the Softmax layer is a vector p, which is the probability distribution for all categories.

[0058] As a preferred solution of the present invention, in step (1.4), when the MDGCN model is used to learn the multi-channel graph neural network, the cross entropy loss function is used for calculation in the final forward propagation process of the neural network:

[0059]

[0060] Among them, W l , w k , b k They represent the weight matrix of the convolutional layer, the weight matrix of the kth layer in the MLP, and the bias vector of the kth layer in the MLP respectively; yt represents the true label of the sample; pt represents the model's predicted positive class probability; and d represents the regularization factor.

[0061] The present invention further provides a computer device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor executes the aforementioned new fault diagnosis method for blast furnace ironmaking process based on MDGCN.

[0062] The present invention also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable the computer to execute the aforementioned new fault diagnosis method for blast furnace ironmaking process based on MDGCN.

[0063] Description of the invention principle:

[0064] During the operation of blast furnace ironmaking, various abnormal conditions may occur, such as difficult furnace operation, material collapse, hanging material, and furnace temperature heating / cooling. The existing technology of using graph convolutional network (GCN) to diagnose faults is usually limited to summarizing the rules of the data itself, which is out of the ironmaking production process, resulting in unsatisfactory fault diagnosis accuracy.

[0065] In order to solve the above problems, the applicant proposed an innovative solution: using the Multichannel Dynamic Graph Convolutional Network (MDGCN) model to construct a graph through system structure and process knowledge. The Euclidean distance between process data at different times is calculated and normalized by softmax. The calculated Euclidean distance is used as the weight of data at different times to further extract the dynamic features of process data. The MDGCN model integrates multiple isomorphic graph channels into the graph neural network and learns features at different levels from the constructed graph. Finally, a multi-layer perceptron classifier is used for fault diagnosis.

[0066] Compared with the MDGCN model reported previously, the MDGCN diagnostic method proposed in the present invention organically combines the mechanism and data of the ironmaking process, and can effectively improve the accuracy of process monitoring and fault diagnosis. It can be seen that the present invention breaks through the inherent technical thinking habits of those skilled in the art and achieves unexpected technical effects.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] 1. The MDGCN-based fault diagnosis method for blast furnace ironmaking process proposed in the present invention uses system structure and process knowledge to construct a graph, assigns different weights to time nodes to extract dynamic features of process data; adopts a multi-isomorphic graph channel strategy to construct a multi-channel GCN model to learn features at different levels; and uses a multi-layer perceptron classifier for fault diagnosis in the online monitoring stage, which effectively improves the accuracy of fault diagnosis.

[0069] 2. The blast furnace ironmaking process has a large amount of data and many data features. In order to perform fault diagnosis more accurately, the present invention combines process topology knowledge and assigns different weights to construct a graph to extract more useful information. Information is transmitted through the edge nodes of the central node, and the features of adjacent nodes are summarized to the central node to update the node status. In this way, the fault diagnosis pre-system process topology is organically combined, and features at different levels can be learned from the constructed graph, thereby improving the accuracy.

[0070] 3. By calculating the dynamic weights between time nodes according to the j of the data, the present invention uses Euclidean distance and Softmax normalization to accurately capture the dynamic relationship between process data at different times. Compared with the static processing method of time nodes in traditional fault diagnosis methods, this dynamic weight assignment mechanism can effectively adapt to the rapidly changing working conditions in the blast furnace ironmaking process, thereby improving the sensitivity to abnormal conditions and diagnostic accuracy.

[0071] 4. The present invention adopts the multi-channel learning mechanism of the MDGCN model to extract features at different levels from multiple isomorphic graphs. This strategy enables the model to better capture the deep dependencies between multidimensional variables in the complex system of blast furnace ironmaking, significantly improving the feature learning ability. Compared with the traditional single-graph or single-channel GCN method, the multi-channel strategy can achieve more efficient feature representation capabilities at the same network depth, further improving the diagnostic accuracy.

[0072] 5. In the online monitoring stage, the present invention uses a multi-layer perceptron (MLP) to classify the extracted features. Compared with the traditional linear classifier, MLP can better handle high-dimensional and nonlinear distribution data features. Combined with the deep features learned by MDGCN, the model achieves significant performance improvement in multi-classification tasks, especially in the diagnosis of complex fault types such as furnace conditions and material collapse. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a schematic diagram of the process of the method of the present invention. DETAILED DESCRIPTION

[0074] The implementation scheme of the present invention is described in detail below in conjunction with the contents of the specific embodiments.

[0075] Part I Implementation of the Invention

[0076] like Figure 1 As shown, the novel fault diagnosis method for blast furnace ironmaking process based on MDGCN of the present invention consists of an offline training stage and an online monitoring stage. In the offline training stage, the historical sample data in the blast furnace ironmaking process is first collected and standardized, the industrial process graph is constructed according to the topological structure of the ironmaking process, and the corresponding adjacency relationship matrix is ​​calculated; by assigning different weights at different time nodes, the dynamic characteristics of the process data are further extracted, and then the dynamic weight matrix is ​​calculated according to the dynamic characteristics of the ironmaking process, and the original data is adjusted to obtain the processed data matrix. On this basis, a new data matrix is ​​constructed and the real label is marked for it, and finally the MDGCN is trained by using the Adam optimizer and the dropout strategy. In the online monitoring stage, the real-time operation data in the blast furnace ironmaking process is first collected and standardized, and the new data matrix is ​​reconstructed according to the data processing method during offline modeling, and the data is updated according to the industrial process graph and the adjacency matrix. Finally, the updated data matrix is ​​input into the trained MDGCN model for prediction, and the output result of the model is used to determine whether there is a fault and the type of fault.

[0077] The specific implementation method is described as follows:

[0078] (I) Offline training phase includes:

[0079] 1. For the blast furnace ironmaking system under normal operation and fault conditions, sufficient sample data of 10 main process variables were collected, including CO (%), H2 (%), CO2 (%), oxygen enrichment flow (m 3 / h), blast furnace top pressure (MPa), cold air pressure (MPa), total pressure difference (MPa), hot air pressure (MPa), resistance coefficient, actual coal injection amount (Ton / h).

[0080] 2. Standardize the data and construct a sample matrix X, where the sample matrix at time t is represented as X t ; According to the dynamic characteristics of the blast furnace ironmaking process, different sample matrices X t Assign dynamic weights to each and get the data matrix X t，d ; Considering the existence of abstract nodes without actual observation values, reconstruct the data matrix X t，d And mark the true label to get the feature matrix X used as input f，t ;

[0081] In the present invention, the sample matrix refers to the matrix X=[x1 x2 …xN ] T , which is composed of sampling data from multiple time points, and each row represents a feature set of a sampling; the data matrix refers to the matrix after normalization; the feature matrix refers to the model input generated after extracting dynamic features and adding abstract nodes based on the data matrix, which is used for classification or prediction.

[0082] (2.1) Based on the Euclidean distance calculation, we can get different sample matrices X t The dynamic weights are as follows:

[0083] (a) Assume that the sample matrix X = [x1, x2…x N ] T The data are collected from N time points and normalized; the normalization formula is as follows:

[0084]

[0085] Where, X normalized refers to the normalized data matrix; X healthy It is a data matrix of samples collected during normal operation of the blast furnace ironmaking system; and S healthy They are X healthy The mean and variance matrices of ;

[0086] (b) Assuming that each set of x observations consists of m process variables, the sample matrix at time t is expressed as but:

[0087] X t =[x t-s+1 , x t-s+2 …x t ] T

[0088] Where s is the time delay; T represents the transpose;

[0089] (c) To maintain universality, the calculated Euclidean distance is normalized using the Softmax function; in the sample matrix X t In, x i and x j The Euclidean distance between is defined as:

[0090] d ij =||x i -x j ||2

[0091] In the formula, x i and x j Point to X tThe i-th and j-th vectors in , i and j represent different time lags respectively;

[0092] Then, the weight matrix is defined as:

[0093]

[0094] In the formula, Yes ij The average value of Represents a distance metric based on mean normalization.

[0095] (2.2) Different sample matrices X t The dynamic relationship between them is determined by the weight matrix M, on which the local dynamic characteristics of the process data are extracted; after calculating the weight matrix M for each group of sampled data, the original sample matrix X t is updated to the data matrix X t，d ,

[0096] X t，d =MX t

[0097] Where, X t represents the sample matrix obtained after preliminary preprocessing of the sampled data; M represents the weight matrix; X t，d Represents the data matrix obtained after calculating the dynamic weights.

[0098] (2.3) Since there are abstract nodes without actual observation values ​​in the topological structure of the blast furnace ironmaking production process, unit nodes will also be generated in the graph after the graph is constructed; define the zero matrix As the data matrix of all abstract nodes, the observation values ​​of all abstract nodes are initialized to zero, where p represents the number of abstract nodes; then the data matrix X containing dynamic features t，d Horizontally concatenate the feature matrix O representing the abstract node to obtain the feature matrix X f，t ; The specific formula is as follows:

[0099] X t，f =[X t，d |O] T

[0100] Where T represents transpose.

[0101] (2.4) The fault diagnosis of the MDGCN model is implemented through graph-level multi-classification tasks. To meet the requirements of supervised learning, for each feature matrix X f，t Assign the corresponding true label at time t The value of yt is an integer from 0 to c, which is used to represent different operating states of the blast furnace ironmaking system.

[0102] 3. Based on the topological structure diagram of the blast furnace ironmaking production process, calculate the adjacency matrix of each diagram;

[0103] (3.1) When constructing a graph, the edges between the nodes in the graph are determined based on the process flow content to represent the relationship between the variables; each sample matrix corresponds to a graph, and all graphs will be used for graph-level classification tasks during the operation of the MDGCN model.

[0104] (3.2) The constructed graph is an undirected graph. The graph constructed in step 3 is an undirected graph, and its adjacency matrix A is a symmetric matrix. In step 2, each data matrix corresponds to a graph, and the adjacency matrix of each graph is Sum degree matrix They are shown as follows:

[0105]

[0106] Where A represents the adjacency matrix (n×n) in step 3; I n represents an n-order identity matrix; i and , represent two node indices; and n represents the total number of nodes.

[0107] 4. The feature matrix X in step 2 f，t The graph constructed in step 3 is input into the MDGCN model and trained using the Adam optimizer and dropout strategy; the MDGCN model has multiple isomorphic graph channels and learns different levels of features from each graph;

[0108] The principle of multi-channel dynamic GCN is: adopting a multi-isomorphic graph channel strategy can improve the representation learning ability without adding layers. The information of each node in the graph is propagated along the topological connection between different units in a complex industrial process. Multiple isomorphic graph channels can enhance the model's ability to learn different deep features. The iterative learning of model parameters uses the Adam optimizer, and the use of the dropout strategy can avoid overfitting problems.

[0109] (4.1) Using the MDGCN model to implement the learning of multi-channel graph neural network, specifically including:

[0110] (a) Initialize the L convolutional layers and q filters in the model, and then calculate each layer as follows:

[0111]

[0112] Wherein, ReLU(·) represents the rectified linear unit (ReLU); represents the degree matrix with self-directed edges; represents the adjacency matrix after adding self-directed edges (the adjacency matrix with self-directed edges corresponding to the sample matrix); 1 = 1…, L, and They are the feature matrices X f，t At time t, the input features and output features of the jth filter corresponding to the lth graph convolutional layer; is the weight feature that the model needs to learn, s represents the time lag, and b is the feature dimension that sets the embedding;

[0113] (b) After all graphs have completed the convolutional layer calculation, the deep representation h of the data learned by all filters is extracted j ;

[0114]

[0115] Among them, flat(·) represents the flatten operation, that is, flattening the high-dimensional feature matrix into a one-dimensional vector; q represents the number of filters;

[0116] Stack each independent and flattened feature into H:

[0117] H=[h1,h2,...,h q ] T

[0118] Where T represents transpose;

[0119] (c) The stacked H is used as the input of the k-layer multi-layer perceptron (MLP) classifier to perform calculations on the MLP fully connected layer:

[0120] H k =ReLU(w k H k-1 +b k )

[0121] z K =[z0,z1,...,z c ] T

[0122] Among them, H k represents the activation value of the kth layer; w k represents the weight matrix of the kth layer; b k represents the bias vector of the kth layer; k = 1, ..., K, K represents the number of fully connected layers; z k is the output of the kth layer in the MLP fully connected layer; c represents the number of categories in the classification task;

[0123] (d) Fault diagnosis is implemented as a classification task; for this purpose, a Softmax layer is used after the kth fully connected layer:

[0124] p=[p0,p1,…,p c ] T =softmax(z K )

[0125] The output of the Softmax layer is a vector p, which is the probability distribution for all categories.

[0126] (4.2) When using the MDGCN model to learn a multi-channel graph neural network, the cross entropy loss function is used for calculation in the final forward propagation process of the neural network:

[0127]

[0128] Among them, W l , w k ,b k They represent the weight matrix of the convolutional layer, the weight matrix of the kth layer in the MLP, and the bias vector of the kth layer in the MLP respectively; yt represents the true label of the sample; pt represents the model's predicted positive class probability; and d represents the regularization factor.

[0129] (II) The online monitoring stage includes:

[0130] 1. Collect sample data of the blast furnace ironmaking system in real time during operation, and then process the sample data according to the operation in step 2 of the offline training phase to reconstruct a new data matrix X f，t ;

[0131] 2. Use the graph and adjacency matrix from step 3 of the offline training phase to update the feature matrix X from the previous step f，t , and obtain the feature matrix

[0132] 3. The feature matrix in the previous step The MDGCN model is input for prediction, and the multi-layer perceptron classifier MLP is used to process the classification task. According to the output diagnosis results, it is determined whether there is a fault in the blast furnace ironmaking production process and the type of fault.

[0133] Part II Contents of a Verification Experiment

[0134] The following simulation experiment is designed based on a real blast furnace ironmaking process (BFIP) data set to verify the effectiveness of the new fault diagnosis method for the blast furnace ironmaking process based on MDGCN. The blast furnace ironmaking process (BFIP) in this simulation experiment is divided into four main units: the blast furnace body, the blast furnace top, the pulverized coal injection system, and the hot air system. The data set comes from an ironmaking enterprise in Zhejiang Province.

[0135] For the MDGCN model, its main structure is to improve the feature learning ability through multi-channel design, and combine the topological feature modeling ability of the graph convolutional network; MDGCN consists of 5 channels, each of which is set to 4 convolution layers, 2 MLP fully connected layers, ReLU is selected as the activation function, the cross entropy function is used for training, and Adam is selected for optimization.

[0136] Table 1 Monitoring variables for BFIP fault diagnosis

[0137]

[0138] In this embodiment, 10 main process measurements are selected to identify the BFIP working status. Table 1 describes the selected 10 process variables.

[0139] Specifically, 4667 samples were collected, including 125 collapse samples, 216 furnace temperature rise samples, 1057 crossflow samples, 245 furnace temperature drop samples, and 3018 normal samples. Commonly used methods in the industry, such as support vector machine (SVM), random forest (RF), deep convolutional neural network (DCNN), knowledge graph convolutional network (k-GCN), parallel time convolutional network (PTCN), etc., were used for comparative study. In order to make a fair comparison, the number of convolutional layers of DCNN, k-GCN, PTCN and MDGCN are all 4. The number of neighbors of k-GCN is set to 2. For MDGCN, the time lag is set to 10.

[0140] The fault diagnosis accuracy (ACR) of various methods is shown in Table 2.

[0141] Table 2 BFIP comparison experimental results

[0142]

[0143] It can be seen from Table 2 that the method used in the present invention has the highest detection accuracy for several common faults in the blast furnace ironmaking process compared with other methods. Therefore, from the perspective of actual verification, it can also be explained that the method can be used in the blast furnace ironmaking production system and effectively monitor the fault conditions in the production process.

Claims

1. A new fault diagnosis method for blast furnace ironmaking process based on MDGCN, characterized in that: It consists of an offline training phase and an online monitoring phase; The offline training phase includes: (1.1) Collect sufficient sample data of main process variables for the blast furnace ironmaking system under normal operation and fault conditions; (1.2) Standardize the data and construct the sample matrix X, where the sample matrix at time t is represented as X t ; According to the dynamic characteristics of the blast furnace ironmaking process, different sample matrices X t Assign dynamic weights to each and get the data matrix X t,d ; Considering the existence of abstract nodes without actual observation values, reconstruct the data matrix X t,d And mark the true label to get the feature matrix X used as input f,t ; (1.3) Based on the topological structure of the blast furnace ironmaking process, calculate the adjacency matrix of each graph; (1.4) Substitute the feature matrix X in step (1.2) f,t The graph constructed in (1.3) is used as input to the MDGCN model, and trained using the Adam optimizer and dropout strategy; the MDGCN model has multiple isomorphic graph channels, and learns different levels of features from each graph; The online monitoring phase includes: (2.1) Collecting sample data of the blast furnace ironmaking system in real time during operation, and then processing the sample data according to the operation in step (1.2) in the offline training phase to reconstruct a new data matrix; (2.2) Using the graph and adjacency matrix in step (1.3), update the data matrix in step (2.1) to obtain the feature matrix (2.3) The feature matrix obtained in step (2.2) The MDGCN model is input for prediction. During this process, the multi-layer perceptron classifier MLP is used for processing and the results are input into the Softmax layer. Finally, the output of the Softmax layer is used to determine whether there is a fault in the blast furnace ironmaking production process and the type of fault.

2. The method according to claim 1, characterized in that In step (1.2), the different sample matrices X are obtained based on the Euclidean distance calculation t The dynamic weights are as follows: (a) Assume that the sample matrix X = [x1, x2…x N ] T The data are collected from N time points and normalized; The normalization formula is as follows: Where, X normalized refers to the normalized data matrix; X healthy It is a data matrix of samples collected during normal operation of the blast furnace ironmaking system; and S healthy They are X healthy The mean and variance matrices of ; (b) Assuming that each set of x observations consists of m process variables, the sample matrix at time t is expressed as but: X t =[x t-s+1 ,x t-s+2 …x t ] T Where s is the time delay; T represents the transpose; (c) To maintain universality, the calculated Euclidean distance is normalized using the Softmax function; in the sample matrix X t In, x i and x j The Euclidean distance between is defined as: d ij =∥∥x i -x j ∥∥2 In the formula, x i and x j Point to X t The i-th and j-th vectors in , i and j represent different time lags respectively; Then, the weight matrix is defined as: In the formula, Yes ij The average value of Represents a distance metric based on mean normalization.

3. The method according to claim 1, characterized in that In step (1.2), different sample matrices X t The dynamic relationship between them is determined by the weight matrix M, on which the local dynamic characteristics of the process data are extracted; after calculating the weight matrix M for each group of sampled data, the original sample matrix X t is updated to the data matrix X t,d , X t,d =MX t Where, X t represents the sample matrix obtained after preliminary preprocessing of the sampled data; M represents the weight matrix; X t,d Represents the data matrix obtained after calculating the dynamic weights.

4. The method according to claim 1, characterized in that: In step (1.2), there are abstract nodes without actual observation values ​​in the topological structure of the blast furnace ironmaking production process. After the graph is constructed, unit nodes will also be generated in the graph; define the zero matrix As the data matrix of all abstract nodes, the observation values ​​of all abstract nodes are initialized to zero, where p represents the number of abstract nodes; then the data matrix X containing dynamic features t,d Horizontally concatenate the feature matrix O representing the abstract node to obtain the feature matrix X f,t ; The specific formula is as follows: Where T represents transpose.

5. The method according to claim 1, characterized in that In step (1.2), the fault diagnosis of the MDGCN model is implemented through a graph-level multi-classification task. To meet the requirements of supervised learning, for each feature matrix X f,t Assign the corresponding true label at time t y t The value is an integer from 0 to c, which is used to represent the different operating states of the blast furnace ironmaking system.

6. The method according to claim 1, characterized in that When constructing the graph in step (1.3), the edges between the nodes in the graph are determined based on the process flow content to represent the relationship between the variables; each sample matrix corresponds to a graph, and all graphs will be used for graph-level classification tasks during the operation of the MDGCN model.

7. The method according to claim 1, characterized in that The graph constructed in step (1.3) is an undirected graph, and its adjacency matrix A is a symmetric matrix; each data matrix in step (1.2) corresponds to a graph, and the adjacency matrix of each graph is Sum degree matrix They are shown as follows: Where A represents the adjacency matrix (n×n) in step (1.3); I n represents an n-order identity matrix; i and j represent two node indices; n represents the total number of nodes.

8. The method according to claim 1, characterized in that In step (1.4), the MDGCN model is used to implement the learning of multi-channel graph neural network, including: (a) Initialize the L convolutional layers and q filters in the model, and then calculate each layer as follows: Among them, ReLU(·) represents the linear rectification function (Rectified Linear Unit, ReLU); represents the degree matrix with self-directed edges; represents the adjacency matrix after adding self-directed edges; l = 1…, L, and They are the feature matrices X f,t At time t, the input features and output features of the jth filter corresponding to the lth graph convolutional layer; is the weight feature that the model needs to learn, s represents the time lag, and b is the feature dimension that sets the embedding; (b) After all graphs have completed the convolutional layer calculation, the deep representation h of the data learned by all filters is extracted j ; Among them, flat(·) represents the flatten operation, that is, flattening the high-dimensional feature matrix into a one-dimensional vector; q represents the number of filters; Stack each independent and flattened feature into H: H=[h1,h2,…,h q ] T Where T represents transpose; (c) The stacked H is used as the input of the k-layer multi-layer perceptron (MLP) classifier to perform calculations on the MLP fully connected layer: A k =ReLU(w k A k-1 +b k ) With K =[z0,z1,…,z c ] T Among them, H k represents the activation value of the kth layer; w k represents the weight matrix of the kth layer; b k represents the bias vector of the kth layer; k = 1, ..., K, K represents the number of fully connected layers; z k is the output of the kth layer in the MLP fully connected layer; c represents the number of categories in the classification task; (d) Fault diagnosis is implemented as a classification task; for this purpose, a Softmax layer is used after the kth fully connected layer: p=[p0,p1,…,p c ] T =softmax(from K ) The output of the Softmax layer is a vector p, which is the probability distribution for all categories.

9. The method according to claim 1, characterized in that: In step (1.4), when using the MDGCN model to learn the multi-channel graph neural network, the cross entropy loss function is used for calculation in the final forward propagation process of the neural network: Among them, W l ,w k ,b k Respectively represent the weight matrix of the convolutional layer, the weight matrix of the k-th layer in the MLP, and the bias vector of the k-th layer in the MLP; y t represents the true label of the sample; p t represents the probability of the model predicting the positive class; d represents the regularization factor.

10. A computer device, characterized in that: include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor executes the new fault diagnosis method for blast furnace ironmaking process based on MDGCN as described in any one of claims 1 to 9.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the new fault diagnosis method for blast furnace ironmaking process based on MDGCN as described in any one of claims 1 to 9.

Images