A blast furnace fault diagnosis method based on trace distance partition decision tree

Abstract

The present invention provides a blast furnace fault diagnosis method based on a trace distance partition decision tree, and relates to the technical field of blast furnace fault diagnosis. The present invention obtains a required data set from a blast furnace production process data acquisition system, obtains historical production process data from an enterprise server, and uses the data to establish a blast furnace fault diagnosis model based on a decision tree of trace distance partition; then, the real-time sampling value at each sampling point is obtained from the blast furnace production process data acquisition system to obtain the blast furnace online production process data, and the online data is passed to the blast furnace fault diagnosis model, thereby outputting the online diagnosis result of the current blast furnace condition to the on-site operator; in addition, the blast furnace production process data acquisition system will store all newly sampled samples in the enterprise server, and when the number of new samples meets the requirement, the blast furnace fault diagnosis model will be retrained in combination with the historical data and the new data to ensure that the model can track and diagnose the latest working condition of the blast furnace.

Description

Technical Field

[0001] The invention relates to the technical field of blast furnace fault diagnosis, and in particular to a blast furnace fault diagnosis method based on a trace distance partition decision tree. Background Art

[0002] As the main process of steel manufacturing, the basic task of blast furnace ironmaking is to extract metallic iron from iron ore. This is a continuous and complex production process, and a large number of parameters need to be monitored, such as furnace top pressure, molten iron temperature, molten iron content, etc. However, unlike other industrial production processes, the production mechanism model of the blast furnace itself is very complex, involving the coexistence of gas, solid and liquid; the monitoring environment conditions are harsh, and it needs to be operated in a closed and high-dust environment; the state of the ironmaking process is non-equilibrium, non-uniform, non-stable, non-linear, and the amount of data is generally large. Therefore, it is necessary to develop a simple and efficient blast furnace fault diagnosis method, which is of great significance to the sustainable development of the steel industry.

[0003] With the development of industrial automation and intelligence, it is becoming increasingly important to use machine learning technology to diagnose and predict blast furnace faults. Decision tree models have been widely used in the field of fault diagnosis because of their intuitive and easy-to-interpret characteristics. The patent "A multi-scale binary tree blast furnace fault diagnosis method based on sample segmentation" (Wang Anna, Sha Mo, application publication number: CN 105550426 A) uses an improved generalized eigenvalue support vector machine algorithm to construct a binary tree classifier. However, the existing decision tree model may produce an overly complex tree structure during the construction process, which not only increases the computational cost of the model, but also may cause the model to overfit the training data, thereby affecting its diagnostic performance in the blast furnace fault diagnosis task, which is not conducive to the rapid offline iteration and real-time online prediction of the fault diagnosis model, and the fault diagnosis effect is difficult to meet the on-site monitoring and diagnosis needs. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a blast furnace fault diagnosis method based on trace distance partition decision tree, which can make full use of detection data, quickly build a fault diagnosis model, and meet the needs of online real-time diagnosis of blast furnaces.

[0005] The technical solution of the present invention is:

[0006] A blast furnace fault diagnosis method based on a trace distance partition decision tree comprises the following steps:

[0007] Step 1: Construct the initial data set for blast furnace fault diagnosis. The initial data set D = {x s ,y s |x s ∈R m ,s∈{1,2,...,n}}, where x sis the feature vector composed of all collected items of the sth sample, y s is the fault label of the sth sample, n is the number of samples, m is the sample dimension, R m Represents m-dimensional real number space.

[0008] Step 2: Preprocess the initial data set and split it to obtain the preprocessed training set D train and the test set D test ;

[0009] Step 2.1: Use box plots to handle outliers: For each feature vector x corresponding to the initial data set D s , calculate the eigenvector x s The upper quartile Q3, median Q2, and lower quartile Q1 of all values ​​in the box plot are calculated. The upper and lower limits of the box plot are calculated as follows: Q3+1.5(Q3-Q1), and the lower limit is calculated as follows: Q1-1.5(Q3-Q1). s All points whose values ​​are greater than the upper limit and less than the lower limit are replaced by the mean of the eigenvector;

[0010] Step 2.2: Randomly select 75% of the samples from the outlier-processed data set as the training set D train , the remaining 25% of the samples are used as the test set D test ;

[0011] Step 3: Construct a division standard based on the trace distance function and calculate the trace distance f value of the current division;

[0012] Step 3.1: Set the training set D train As the current data set, the number of samples in the data set is represented by n, the sample dimension is represented by m, and the number of fault label categories is represented by k. Calculate the matrix Q;

[0013] Step 3.1.1: Based on the training set D train Existing fault labels, calculate the matrix Where k represents the number of fault label categories in the current training set. For matrix A, when i≠j, A ij =0, that is, the non-diagonal elements are 0, where i,j=1,2,...,k, respectively representing the i-th and j-th fault labels; A ii represents the number of samples of the i-th type fault label, and ∑ i A ii =n, where n represents the number of samples;

[0014] Step 3.1.2: Based on the matrix A obtained in step 3.1.1, further calculate the matrix in Qii It represents the proportion of samples of each type of label in the number of samples in the current data set, then the sum of the probabilities is 1, that is, the sum of the diagonal elements is 1;

[0015] Step 3.2: Set the training set D train As the current data set, the number of samples in the data set is represented by n, the sample dimension is represented by m, and the number of fault label categories is represented by k. The matrix P is calculated as:

[0016] Step 3.2.1: Based on the training set D train The sample features in D train The tth feature X in t The sth element of in n represents the number of samples in the training set, and m represents the sample dimension. The size of the current data set is divided into two sub-data sets, that is, X t Medium to large The samples corresponding to the elements of are divided into the left nodes, and X t Less than or equal to The samples corresponding to the elements of are divided into the right node, and the label with the highest frequency in the data set under the child node is used as the fault diagnosis label of the child node. The fault diagnosis labels of the two child nodes after the division are compared with the real fault labels, and the matrix is ​​calculated. Among them B ij represents the number of samples with the i-th fault label classified into the j-th fault, and ∑ ij B ij =n, i, j = 1, 2, ..., k;

[0017] Step 3.2.2: Based on the matrix B obtained in step 3.2.1, further calculate the matrix in P ij Indicates B ij The proportion of samples in the current data set, and ∑ ij P ij =1, then the sum of the elements of the matrix P is 1, that is, the sum of the probabilities is equal to 1;

[0018] Step 3.3: Calculate the current divided trace distance f value according to the following formula:

[0019]

[0020] Where Tr(·) represents the trace of the matrix. When the classification is completely correct, that is, when P = Q, When the classification is not completely correct, that is, when P≠Q,

[0021] Step 4: Select the feature with the smallest trace distance f value and its corresponding segmentation point as the optimal feature and optimal segmentation point;

[0022] Step 4.1: According to the training set, D train Traverse each feature in and get the minimum trace distance f value for each feature:

[0023] Step 4.1.1: D train The tth feature in Traverse, where t = 1, .., m, m represents the sample dimension, n represents the number of samples in the current data set, and calculates the feature X t The trace distance f value under different elements

[0024] Step 4.1.2: Calculate the feature X obtained in step 4.1.1 t The minimum trace distance f value under different elements, that is,

[0025] Step 4.2: Based on the training set D train , for D train Traverse each feature in and obtain the minimum trace distance f value under all features, that is, f = min(f 1 ,f 2 ,…,f m ), where m represents the sample dimension of the current data set, and the feature corresponding to the trace distance f value is selected as the optimal feature, and the corresponding segmentation point is selected as the optimal segmentation point;

[0026] Step 5: Generate a decision tree model based on trace distance partitioning;

[0027] Step 5.1: Generate two child nodes from the root node according to the optimal features and optimal segmentation points, and distribute the samples in the current data set to the left and right child nodes;

[0028] Step 5.2: Determine whether the left and right child nodes meet the stopping condition. If not, recursively call steps 3 and 4 for the two child nodes until the stopping condition is met. If the stopping condition is met, output the decision tree model based on trace distance division:

[0029] The condition for stopping the calculation is that the number of samples in the current node is less than the preset threshold δ, or the trace distance f value of the data set under the current node is less than the preset threshold ε;

[0030] Step 6: Collect blast furnace ironmaking production process data in real time and input it into the decision tree fault diagnosis model based on trace distance partitioning obtained in step 5. Use the diagnosis model to diagnose the current blast furnace fault. If a fault is detected, the model will push the fault category to the on-site operator to adjust the blast furnace production process in time.

[0031] The beneficial effects of adopting the above technical solution are:

[0032] The present invention provides a blast furnace fault diagnosis method based on a trace distance partition decision tree. When the nodes of the decision tree are split, the partition standard based on the trace distance is used to replace the partition standard based on the Gini index in the traditional CART tree. The method can evaluate the partition effect of a node more quickly on the one hand, reduce the training time of the decision tree model, and on the other hand, it can also reduce the complexity of the decision tree model, and obtain higher diagnostic accuracy. Through experimental testing of actual blast furnace production data, compared with traditional methods, the method of the present invention can obtain the best blast furnace fault diagnosis effect in a shorter training time, can help field operators predict possible blast furnace failures in advance, and timely adjust the blast furnace production process to avoid serious blast furnace production accidents, thereby improving the production efficiency of steel enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the principle of a blast furnace fault diagnosis method based on a trace distance partition decision tree in an embodiment of the present invention;

[0034] Figure 2 It is a flow chart of a blast furnace fault diagnosis method based on a trace distance partition decision tree in an embodiment of the present invention;

[0035] Figure 3 This is a comparison chart of model training time between the method of the embodiment of the present invention and the CART tree model under different numbers of features;

[0036] Figure 4 This is a comparison chart of the model training time of 30 repeated experiments between the method of the embodiment of the present invention and the CART tree model;

[0037] Figure 5 It is a comparison chart of ROC curves of the method of the embodiment of the present invention and the CART tree model, SVM, and DNN algorithm. DETAILED DESCRIPTION

[0038] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0039] As the first step in steel production, blast furnace ironmaking is a continuous and complex production process. Figure 1 As shown, raw materials such as iron ore and coke are continuously injected into the blast furnace from the top of the blast furnace through the distribution system, and a mixed gas containing high-temperature hot air and fuels such as coal powder and natural gas enters the blast furnace from the tuyere. The coke burns with the oxygen in the high-temperature hot air to generate blast furnace gas containing carbon monoxide and hydrogen. The blast furnace gas will produce a reduction reaction with the descending iron ore and other furnace materials during the rising process, thereby removing oxygen from the iron ore to generate molten iron, which is discharged from the iron outlet of the blast furnace hearth; while the unreduced impurities in the iron ore, the ash in the coke and coal combine with fluxes such as limestone to produce slag, which is discharged from the slag outlet of the blast furnace hearth. In view of the possible faults in the blast furnace ironmaking process, the present invention provides a blast furnace fault diagnosis method based on a decision tree based on trace distance division, which can quickly diagnose the blast furnace condition. The working principle of this method is as follows: Figure 1 As shown, the flow chart is as follows Figure 2 As shown. The following steps are included:

[0040] Step 1: During the blast furnace ironmaking process, based on the data collected by the blast furnace's charge distribution system, gas system, cooling system, air supply system, injection system, slag iron system, and various sensors distributed throughout the blast furnace body, an initial data set for blast furnace fault diagnosis is constructed. The initial data set D = {x s ,y s |x s ∈R m ,s∈{1,2,...,n}}, where x s is the feature vector composed of all collected items of the sth sample, y s is the fault label of the sth sample (normal or faulty in this embodiment), n is the number of samples, m is the sample dimension, R m Represents m-dimensional real number space.

[0041] In this embodiment, actual production data of a blast furnace in a steel plant in China from December 31, 2019 to October 14, 2023 are collected. Among them, the sampling interval of the data is 1 hour. Since the fault label is recorded in the blast furnace condition record in units of days, and the inspection and testing indicators of molten iron and slag in the data need to be obtained by offline testing after the blast furnace is tapped, there is a mismatch in sampling time between the various collection items of the blast furnace. In order to solve this problem, the mean of each collection item in each day is used as the sample value of the day, and together with the blast furnace fault label of the day, it constitutes a training sample.

[0042] In this embodiment, the number of samples in the initial data set is 1380, and the sample dimension is 68. The collection items of blast furnace fault data include: coke ratio, furnace top pressure, air volume, coke ad, (MgO), molten iron output, oxygen enrichment pressure, furnace center temperature, air pressure, oxygen enrichment, air temperature, standard wind speed, coke st, pellets, furnace bottom temperature extreme difference, blast kinetic energy, etc., and the detailed information is shown in Table 1:

[0043] Table 1 Collection item information of blast furnace ironmaking production process data

[0044]

[0045] Step 2: Preprocess the initial data set and split it to obtain the preprocessed training set D train and the test set D test ;

[0046] Step 2.1: Use box plots to handle outliers: For each feature vector x corresponding to the initial data set D s , calculate the eigenvector x s The upper quartile Q3, median Q2, and lower quartile Q1 of all values ​​in the box plot are calculated. The upper and lower limits of the box plot are calculated as follows: Q3+1.5(Q3-Q1), and the lower limit is calculated as follows: Q1-1.5(Q3-Q1). s All points whose values ​​are greater than the upper limit and less than the lower limit are replaced by the mean of the eigenvector;

[0047] Step 2.2: Randomly select 75% of the samples (a total of 1035 in this example) from the data set after outlier processing as the training set D train The remaining 25% of the samples (345 in total in this example) are used as the test set D test ;

[0048] Step 3: Construct a division standard based on the trace distance function and calculate the trace distance f value of the current division;

[0049] Step 3.1: Set the training set D train As the current data set, the number of samples in the data set is represented by n, the sample dimension is represented by m, and the number of fault label categories is represented by k. Calculate the matrix Q;

[0050] Step 3.1.1: Based on the training set D train Existing fault labels, calculate the matrix Where k represents the number of fault label categories in the current training set. For matrix A, when i≠j, A ij=0, that is, the non-diagonal elements are 0, where i,j=1,2,...,k, respectively representing the i-th and j-th fault labels; A ii represents the number of samples of the i-th type fault label, and ∑ i A ii =n, where n represents the number of samples;

[0051] In this embodiment, n=1035. Since there are only two types of labels, normal and faulty, in this embodiment, that is, k=2, then Among them A 11 Indicates the number of samples with normal labels, A 22 represents the number of samples of fault labels, and A 11 +A 22 =1035;

[0052] Step 3.1.2: Based on the matrix A obtained in step 3.1.1, further calculate the matrix in Q ii It represents the proportion of samples of each type of label in the number of samples in the current data set, then the sum of the probabilities is 1, that is, the sum of the diagonal elements is 1;

[0053] Step 3.2: Set the training set D train As the current data set, the number of samples in the data set is represented by n, the sample dimension is represented by m, and the number of fault label categories is represented by k. The matrix P is calculated as:

[0054] Step 3.2.1: Based on the training set D train The sample features in D train The tth feature X in t The sth element of in n represents the number of samples in the training set, and m represents the sample dimension. The size of the current data set is divided into two sub-data sets, that is, X t Medium or larger The samples corresponding to the elements of are divided into the left nodes, and X t Less than or equal to The samples corresponding to the elements of are divided into the right node, and the label with the highest frequency in the data set under the child node is used as the fault diagnosis label of the child node. The fault diagnosis labels of the two child nodes after the division are compared with the real fault labels, and the matrix is ​​calculated. Among them B ij represents the number of samples with the i-th fault label classified into the j-th fault, and ∑ ij B ij =n, i, j = 1, 2, ..., k;

[0055] Step 3.2.2: Based on the matrix B obtained in step 3.2.1, further calculate the matrix in P ij Indicates B ij The proportion of samples in the current data set, and ∑ ij P ij =1, then the sum of the elements of the matrix P is 1, that is, the sum of the probabilities is equal to 1;

[0056] Step 3.3: Calculate the current divided trace distance f value according to the following formula:

[0057]

[0058] Where Tr(·) represents the trace of the matrix. When the classification is completely correct, that is, when P = Q, When the classification is not completely correct, that is, when P≠Q,

[0059] Step 4: Select the feature with the smallest trace distance f value and its corresponding segmentation point as the optimal feature and optimal segmentation point;

[0060] Step 4.1: According to the training set, D train Traverse each feature in and get the minimum trace distance f value for each feature:

[0061] Step 4.1.1: D train The tth feature in Traverse, where t = 1, .., m, m represents the sample dimension, n represents the number of samples in the current data set, and calculates the feature X t The trace distance f value under different elements:

[0062] Step 4.1.2: Calculate the feature X obtained in step 4.1.1 t The minimum trace distance f value under different elements, that is,

[0063] Step 4.2: Based on the training set D train , for D train Traverse each feature in and obtain the minimum trace distance f value under all features, that is, f = min(f 1 ,f 2 ,…,f m ), where m represents the sample dimension of the current data set, and the feature corresponding to the trace distance f value is selected as the optimal feature, and the corresponding segmentation point is selected as the optimal segmentation point;

[0064] Step 5: Generate a decision tree model based on trace distance partitioning;

[0065] Step 5.1: Generate two child nodes from the root node according to the optimal features and optimal segmentation points, and distribute the samples in the current data set to the left and right child nodes;

[0066] Step 5.2: Determine whether the left and right child nodes meet the stopping condition. If not, recursively call steps 3 and 4 for the two child nodes until the stopping condition is met. If the stopping condition is met, output the decision tree model based on trace distance division:

[0067] The condition for stopping the calculation is that the number of samples in the current node is less than a predetermined threshold value δ, or the trace distance f value of the data set under the current node is less than a predetermined threshold value ε. In this embodiment, δ = 1, ε = 0.01;

[0068] Step 6: Collect blast furnace ironmaking production process data in real time and input it into the decision tree fault diagnosis model based on trace distance partitioning obtained in step 5. Use the diagnosis model to diagnose the current blast furnace fault. If a fault is detected, the model will push the fault category to the on-site operator to adjust the blast furnace production process in time.

[0069] In this embodiment, a fault diagnosis method based on the trace distance partitioning decision tree of the present invention is compared with the model training time of the traditional CART decision tree. First, a comparative experiment is conducted on the model training time of the two methods under data sets composed of different numbers of features. The number of features is set to gradually increase from 10 to 68, and the training time change curves of the two methods are compared. Figure 3 As shown. It can be seen that compared with the CART algorithm, the training time of the method of the present invention is shorter under all feature numbers. When all 68 features are selected to build the fault diagnosis model, both methods are run independently 30 times to eliminate the influence of experimental errors. The training time of the two methods in 30 experiments is recorded, and the experimental results are compared using box plots. Figure 4 As shown, it can be seen that the training time of the method of the present invention is significantly shorter. The above results show that the method of the present invention can more efficiently construct a decision tree fault diagnosis model.

[0070] In this embodiment, the fault diagnosis method based on the trace distance partition decision tree of the present invention is compared with the traditional CART decision tree model, SVM (support vector machine), and DNN (deep neural network) classification method, wherein the support vector machine uses radial basis function as the kernel function, and the deep neural network is a four-layer network structure (including 1 input layer, 1 output layer, and 2 hidden layers), and the number of nodes in each layer is 68-70-35-2. Accuracy, precision, recall rate, and F1-Score are used as evaluation indicators of the model, and the experimental results of each algorithm are shown in Table 2. It can be seen from Table 2 that the method of the present invention has the best diagnostic results in blast furnace fault diagnosis. The ROC curve comparison diagram of blast furnace fault diagnosis of an embodiment of the present invention and CART tree model, SVM, and RNN is as shown in the figure. Figure 5 shown.

[0071] Table 2 Comparison of experimental results of various algorithms

[0072]

[0073] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with the technical features with similar functions disclosed in the embodiments of the present disclosure (but not limited to) to form a technical solution.

Claims

1. A blast furnace fault diagnosis method based on trace distance partition decision tree, characterized in that: The following steps are involved: Step 1: Construct the initial data set for blast furnace fault diagnosis; Step 2: Preprocess the initial data set and split it to obtain the preprocessed training set D train and the test set D test ; Step 3: Construct a division standard based on the trace distance function and calculate the trace distance f value of the current division; Step 3.1: Set the training set D train As the current data set, the number of samples in the data set is represented by n, the sample dimension is represented by m, and the number of fault label categories is represented by k. Calculate the matrix Q; Step 3.1.1: Based on the training set D train Existing fault labels, calculate the matrix Where k represents the number of fault label categories in the current training set. For matrix A, when i≠j, A ij =0, that is, the non-diagonal elements are 0, where i,j=1,2,...,k, respectively representing the i-th and j-th fault labels; A ii represents the number of samples of the i-th type fault label, and ∑ i A ii =n, where n represents the number of samples; Step 3.1.2: Based on the matrix A obtained in step 3.1.1, further calculate the matrix in Q ii It represents the proportion of samples of each type of label in the number of samples in the current data set, then the sum of the probabilities is 1, that is, the sum of the diagonal elements is 1; Step 3.2: Set the training set D train As the current data set, the number of samples in the data set is represented by n, the sample dimension is represented by m, and the number of fault label categories is represented by k. The matrix P is calculated as: Step 3.2.1: Based on the training set D train The sample features in D train The tth feature X in t The sth element of in n represents the number of samples in the training set, and m represents the sample dimension; according to The size of the current data set is divided into two sub-data sets, that is, X t Medium to large The samples corresponding to the elements of are divided into the left nodes, and X t Less than or equal to The samples corresponding to the elements of are divided into the right node, and the label with the highest frequency in the data set under the child node is used as the fault diagnosis label of the child node. The fault diagnosis labels of the two child nodes after the division are compared with the real fault labels, and the matrix is ​​calculated. Among them B ij represents the number of samples with the i-th fault label classified into the j-th fault, and ∑ ij B ij =n, i, j = 1, 2, ..., k; Step 3.2.2: Based on the matrix B obtained in step 3.2.1, further calculate the matrix in P ij Indicates B ij The proportion of samples in the current data set, and ∑ ij P ij =1, then the sum of the elements of the matrix P is 1, that is, the sum of the probabilities is equal to 1; Step 3.3: Calculate the current divided trace distance f value according to the following formula: Where Tr(·) represents the trace of the matrix. When the classification is completely correct, that is, when P = Q, When the classification is not completely correct, that is, when P≠Q, Step 4: Select the feature with the smallest trace distance f value and its corresponding segmentation point as the optimal feature and optimal segmentation point; Step 5: Generate a decision tree model based on trace distance partitioning; Step 6: Collect blast furnace ironmaking production process data in real time and input it into the decision tree fault diagnosis model based on trace distance partitioning obtained in step 5. Use the diagnosis model to diagnose the current blast furnace fault. If a fault is detected, the model will push the fault category to the on-site operator to adjust the blast furnace production process in time.

2. A blast furnace fault diagnosis method based on trace distance partition decision tree according to claim 1, characterized in that: The initial data set D in step 1 is s ,y s |x s ∈R m ,s∈{1,2,...,n}}, where x s is the feature vector composed of all collected items of the sth sample, y s is the fault label of the sth sample, n is the number of samples, m is the sample dimension, R m Represents m-dimensional real number space.

3. A blast furnace fault diagnosis method based on trace distance partition decision tree according to claim 1, characterized in that: The step 2 specifically includes: Step 2.1: Use box plots to handle outliers: For each feature vector x corresponding to the initial data set D s , calculate the eigenvector x s The upper quartile Q3, median Q2, and lower quartile Q1 of all values ​​in the box plot are calculated. The upper and lower limits of the box plot are calculated as follows: Q3+1.5(Q3-Q1), and the lower limit is calculated as follows: Q1-1.5(Q3-Q1). s All points whose values ​​are greater than the upper limit and less than the lower limit are replaced by the mean of the eigenvector; Step 2.2: Randomly select 75% of the samples from the outlier-processed data set as the training set D train , the remaining 25% of the samples are used as the test set D test .

4. A blast furnace fault diagnosis method based on trace distance partition decision tree according to claim 1, characterized in that: The step 4 specifically comprises the following steps: Step 4.1: Based on the training set, train Traverse each feature in and get the minimum trace distance f value for each feature: Step 4.1.1: D train The tth feature in Traverse, where t = 1, .., m, m represents the sample dimension, n represents the number of samples in the current data set, and calculates the feature X t The trace distance f value under different elements Step 4.1.2: Calculate the feature X obtained in step 4.1.1 t The minimum trace distance f value under different elements, that is, Step 4.2: Based on the training set D train , for D train Traverse each feature in and obtain the minimum trace distance f value under all features, that is, f = min(f 1 ,f 2 ,…,f m ), where m represents the sample dimension of the current data set, and the feature corresponding to the trace distance f value is selected as the optimal feature, and the corresponding segmentation point is selected as the optimal segmentation point.

5. The blast furnace fault diagnosis method based on trace distance partition decision tree according to claim 1 is characterized in that: The step 5 specifically comprises the following steps: Step 5.1: Generate two child nodes from the root node according to the optimal features and optimal segmentation points, and distribute the samples in the current data set to the left and right child nodes; Step 5.2: Determine whether the left and right child nodes meet the stopping condition. If not, recursively call steps 3 and 4 for the two child nodes until the stopping condition is met. If the stopping condition is met, output the decision tree model based on trace distance partitioning.

6. A blast furnace fault diagnosis method based on trace distance partition decision tree according to claim 5, characterized in that: The stopping condition in step 5.2 is that the number of samples in the current node is less than a predetermined threshold δ, or the trace distance f value of the data set under the current node is less than a predetermined threshold ε.

Images