A method and system for setting optimal material level of blast furnace

Through the three-dimensional reconstruction and noise reduction automatic coding machine model, combined with the improved particle swarm algorithm, the problem of low optimal material surface setting accuracy of blast furnace is solved, and multi-region feature extraction of material surface shape and optimization of production indicators is realized, which has improved the Gas utilization and stability of blast furnaces.

CN114048671BActive Publication Date: 2024-12-06CENT SOUTH UNIV
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202111227244.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-10-21
Publication Date
2024-12-06
Estimated Expiration
2041-10-21

AI Technical Summary

Technical Problem

The existing blast furnace optimal material surface setting method has low accuracy, and it is impossible to make accurate decisions quickly and under abnormal working conditions in the furnace, and ignores the multi-regional characteristics of the material surface shape and the inherent connection between production indicators.

Method used

By performing three-dimensional reconstruction of blast furnace surface image, three-dimensional shape features are extracted, and a blast furnace surface optimization process model is established based on a noise reduction automatic encoder and an improved particle swarm algorithm. Combining the attention mechanism and multi-objective optimization algorithm, we can solve the optimal material. Face feature parameters.

Benefits of technology

It realizes the precise setting of the optimal material surface of the blast furnace, improves gas utilization and production stability, and ensures the accurate characterization of multi-region characteristics of the material surface shape and the optimization of production indicators.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114048671B_ABST
    Figure CN114048671B_ABST
Patent Text Reader

Abstract

The invention discloses a method and system for setting an optimal charge surface of a blast furnace. The method comprises the following steps: performing three-dimensional reconstruction on an acquired charge surface image to obtain a three-dimensional charge surface shape of the blast furnace, performing feature extraction on the three-dimensional charge surface shape of the blast furnace to obtain charge surface shape features, establishing a charge surface optimization process model of the blast furnace and solving an optimal solution set of the charge surface optimization process model based on the charge surface shape features, a furnace condition state parameter of the blast furnace and a blast furnace production index, selecting a charge surface shape feature corresponding to an optimal solution suitable for the current furnace condition from the optimal solution set as a decision variable, thereby completing the setting of the optimal charge surface of the blast furnace, solving the technical problem of low precision of the existing method for setting the optimal charge surface of the blast furnace, converting the setting problem of the optimal charge surface of the blast furnace into a production index optimization problem based on the intrinsic connection between the blast furnace production index and the charge surface shape of the blast furnace and the furnace condition state parameter of the blast furnace from a data perspective, and solving the decision variables characterizing the optimal charge surface of the blast furnace by establishing a charge surface optimization process model of the blast furnace, thereby accurately realizing the setting of the optimal charge surface of the blast furnace, and providing theoretical guidance for subsequent blast furnace charging.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention mainly relates to the field of blast furnace charge surface depth detection, and in particular to a blast furnace optimal charge surface setting method and system. Background Art

[0002] Blast furnace ironmaking is the upstream process with the highest energy consumption in the steel production process, and its charge surface shape has always been the focus and difficulty of blast furnace research. The ideal charge surface shape can make the gas flow distribution more reasonable and the production more efficient. In actual production, the blast furnace charging operation parameters are adjusted according to manual experience to form an ideal charge surface shape and a reasonable gas flow distribution. Although the blast furnace charge surface setting method based on experience can keep the blast furnace in a relatively stable production state, when abnormal conditions such as hanging material, collapsed material, inclined material, and pipeline travel occur in the furnace, this method cannot make accurate decisions quickly, and it is difficult to ensure that the production indicators are within the target range. With the development of distributed control systems, a large amount of production data is recorded and stored. Therefore, based on the large amount of production data collected, the intrinsic relationship between state parameters and production indicators is mined, and from the perspective of data-driven, the optimal charge surface setting method of the blast furnace is studied, which is of great significance to improving the utilization rate of coal gas in the furnace, promoting the stable and smooth operation of the blast furnace, improving product quality, and achieving energy conservation and consumption reduction.

[0003] The research object of the present invention is the charge surface of a blast furnace. Due to the harsh environmental characteristics of the blast furnace, such as high temperature, high pressure, dust, and dim light, there are problems such as difficulty in obtaining charge surface information, limited charge surface detection methods, and low accuracy of charge surface information. The present invention uses a high-temperature industrial endoscope as a detection device to perform three-dimensional reconstruction of the captured charge surface image, and extracts the characteristics of the charge surface shape based on the three-dimensional information of the blast furnace charge surface; based on the basis for adjusting the charge surface shape at the blast furnace production site, the problem of setting the optimal charge surface is converted into a multi-objective optimization problem; and the inherent relationship between the charge surface characteristic parameters and state parameters and the production indicators is fully considered, and a SA-DAE method is proposed to establish a process model for charge surface optimization. Finally, an improved AWPSO-NSGA-II algorithm is proposed to solve the model, and the optimal solution obtained is the charge surface characteristic parameter value, thereby realizing the setting of the optimal charge surface of the blast furnace.

[0004] The invention with application publication number CN112853014A proposes a method for adjusting the shape of the material surface of a blast furnace. The detected material surface shape is presented on the screen using a laser device, and the current material surface structure is judged and adjusted based on the inclination angle of the material surface in the graph. When the inclination angle is in different set intervals, the material surface is corrected by adjusting the material distribution angle. This method has a good adjustment effect in actual production, but still has the following shortcomings. (1) Only the inclination angle from the platform to the central area is used as the material surface shape feature, ignoring the characteristics of the platform area and the edge area, making the material surface shape description inaccurate; (2) Only the relationship between the material surface inclination angle and the daily wind volume deviation value is considered, ignoring the potential relationship between the production indicators and state parameters and the material surface inclination angle.

[0005] The invention with application publication number CN106682688A proposes a stacked denoising autoencoder network bearing fault diagnosis method based on particle swarm optimization. By preprocessing the original input signal and using the improved particle swarm algorithm to tune the hyperparameters of the SDAE network, the deep feature representation of the fault state can be obtained, and then the soft-max classifier is used to realize the classification and identification of the fault. Although this method improves the accuracy of fault classification, it still has the following problems. (1) It only changes the inertia weight of the particles, ignoring the importance of the individual learning factor and social learning factor of the particles in finding the global optimal solution, and does not consider maintaining the diversity of the population, resulting in premature convergence of particles; (2) The deep features extracted by SDAE are only continuously strengthened by stacking each DAE, without considering the relationship between each hidden layer feature and the fault state, resulting in poor extraction of deep features. Summary of the invention

[0006] The method and system for setting the optimal material level of a blast furnace provided by the present invention solve the technical problem of low precision of the existing method for setting the optimal material level of a blast furnace.

[0007] In order to solve the above technical problems, the method for setting the optimal charge level of a blast furnace proposed by the present invention comprises:

[0008] Performing three-dimensional reconstruction on the acquired material surface image to obtain the three-dimensional blast furnace material surface shape, and performing feature extraction on the three-dimensional blast furnace material surface shape to obtain material surface shape features;

[0009] Based on the relationship between the shape characteristics of the charge surface and the state parameters of the blast furnace and the blast furnace production indicators, a blast furnace charge surface optimization process model is established.

[0010] The optimal solution set of the blast furnace charge surface optimization process model is solved, and the charge surface shape characteristics corresponding to the optimal solution suitable for the current furnace conditions are selected from the optimal solution set as decision variables, thereby completing the setting of the blast furnace's optimal charge surface.

[0011] Furthermore, based on the relationship between the charge surface shape characteristics and the blast furnace condition parameters and the blast furnace production indicators, a blast furnace charge surface optimization process model is established, including:

[0012] Determine the material surface shape characteristics and blast furnace condition parameters as the original input vectors of the blast furnace material surface optimization process model;

[0013] Determine the blast furnace production index as the target output vector of the blast furnace charge level optimization process model;

[0014] Based on the original input vector and the target output vector, a blast furnace charge level optimization process model was established using a noise reduction autoencoder.

[0015] Furthermore, based on the original input vector and the target output vector, a blast furnace charge level optimization process model is established by using a noise reduction autoencoder, including:

[0016] An attention module is introduced between the encoding and decoding of the denoising autoencoder to obtain a denoising autoencoder model based on the target attention mechanism.

[0017] The original input vector is used as the input vector of the denoising autoencoder model based on the target attention mechanism, and the hidden layer feature vector and the target output vector of the denoising autoencoder model based on the target attention mechanism are input into the attention module together, and the correlation between each feature in the hidden layer feature vector and each variable in the target output vector is calculated. The specific calculation formula is:

[0018]

[0019]

[0020]

[0021] Where g(·) represents the activation function from the hidden layer to the attention module of the denoising autoencoder model based on the target attention mechanism, h represents the feature vector of the hidden layer, y represents the output vector of the denoising autoencoder model based on the target attention mechanism, and W h represents the weight matrix of the eigenvector h, W y represents the weight matrix of the output vector y, b 0 represents the bias vector from the hidden layer to the attention module, r is the attention weight vector before normalization, softmax is the normalized exponential function, β represents the normalized attention weight vector, d h represents the dimension of the feature vector h, r i represents the attention weight coefficient corresponding to the i-th feature in the feature vector h, represents the Hadamard product, h y is the feature vector based on target attention;

[0022] The feature vector h based on the target attention y Map to the output layer to obtain the reconstructed output vector;

[0023] With the goal of minimizing the reconstruction loss, the BP algorithm is used to adjust the model parameters θ of the denoising autoencoder based on the target attention mechanism. e = {W, b, W′, b′} and θ a = {W h ,W y ,b 0} for training, where W and b represent the weight matrix and bias vector from the input layer to the hidden layer of the denoising autoencoder model based on the target attention mechanism, respectively, and W′ and b′ represent the weight matrix and bias vector from the hidden layer to the output layer of the denoising autoencoder model based on the target attention mechanism, respectively;

[0024] Stacking multiple trained denoising autoencoder models based on the target attention mechanism in sequence to obtain a stacked denoising autoencoder model based on the target attention mechanism;

[0025] An output layer is added after the last hidden layer of the stacked denoising autoencoder model based on the target attention mechanism to obtain a complete model of the blast furnace charge level optimization process.

[0026] Furthermore, the three-dimensional blast furnace material surface shape is subjected to feature extraction, and the material surface shape features obtained include:

[0027] Based on the right-hand rule, establish a three-dimensional material surface coordinate system;

[0028] At a first preset sampling interval, the charge is sampled successively from the furnace wall to the furnace core at the furnace throat in a circular shape to obtain a circular sampling set;

[0029] At a second preset sampling interval, sampling points on each circular ring in the circular ring sampling set is performed successively to obtain sampling points corresponding to the circular ring;

[0030] Calculate the mean of the vertical distances between the sampling points corresponding to the circular ring and the zero material line, and perform mean processing on the sampling points corresponding to the circular ring according to the mean, so as to obtain a symmetrical blast furnace charge surface that is completely symmetrical about the center line of the blast furnace;

[0031] According to the slope of the radial material line between adjacent points in the radial direction of the symmetrical blast furnace material surface, the radial material line is divided into three areas, which are the edge area, the platform area and the funnel area;

[0032] According to the three regions of radial material line, the feature extraction of three-dimensional blast furnace material surface is carried out to obtain the material surface shape characteristics.

[0033] Furthermore, according to the three regions of the radial material line, the three-dimensional blast furnace material surface shape is extracted, and the material surface shape features obtained include:

[0034] Establish a radial material line coordinate system, with the throat zero material line position as the x-axis, the furnace center as the coordinate origin, and the y-axis vertically downward;

[0035] Based on the radial material line coordinate system and the three areas of the radial material line, the material surface shape characteristics are extracted. The material surface shape characteristics include the vertical distance from the zero material line to the platform, the inclination of the furnace wall, the platform width, the funnel width, the funnel depth, the funnel inclination and the funnel center angle.

[0036] Furthermore, the optimal solution set of the blast furnace charge level optimization process model includes:

[0037] Step 1: Assume that the size of the population P is N and the number of iterations is t max , randomly initialize the particle position within the set value range, V max is the maximum allowed speed, initialize the individual optimal position and the global optimal position, that is, p best =g best =0, current iteration number t=0;

[0038] Step 2: Calculation, specifically calculating the target values ​​of all particles in the current population P;

[0039] Step 3: Update the optimal particle. Specifically, let t = t + 1 and use the dynamic weighted method to calculate the dynamic fitness of each particle. The calculation formula is as follows:

[0040]

[0041] In the formula, p is the current particle fitness, M is the number of targets to be optimized, and f j (x) is the function value of the particle on the jth objective function, ω j is the weight assigned to the particle on the jth target, and

[0042] The particle with the largest dynamic fitness is selected as the global optimal particle, and the individual optimal particle is selected by comparing the current particle position with the individual historical optimal position, thus completing g best and p best Updates;

[0043] Step 4: Generate new particles, specifically by copying P to form a new population Q, calculating the updated speed and position of each particle in population Q, and calculating each target value of each particle in Q, combining P and Q into population R, with a particle number of 2N. The calculation formula is as follows:

[0044]

[0045] In the formula, v i (t) is the flying speed of the i-th particle at time t, ω is the inertia weight of the particle, g best is the position of the global optimal particle, p best is the position of the individual optimal particle, v i (t-1) is the flying speed of the i-th particle at time t-1, x i (t) is the position of the ith particle at time t, x i (t-1) is the position of the i-th particle at time t-1, r 1 and r 2 is a random number uniformly distributed in [0,1], ω is the inertia weight, and ω=ω end +ρ(ω start -ω end ), where ω start is the initial inertia weight, ω end is the convergence inertia weight, ω end =[0,0.5],ω start >ω end and are all constants, ρ is a random number distributed in [0,1], c 1 and c 2 They are individual learning factor and social learning factor, and:

[0046]

[0047] In the formula, c 11 and c 12 They are c 1 The initial value and convergence value of c 21 and c 22 They are c 2 The initial value and convergence value of , t is the current iteration number, t max is the total number of iterations, and c 11 =c 21 =2.5, c 12 =c 22 =0.5;

[0048] Step 5: Sorting, specifically using the non-dominated sorting idea of ​​NSGA-II to perform non-dominated sorting and crowding distance sorting on the population R;

[0049] Step 6: Select the next generation of particles, specifically select N particles from the population R in sequence to the population P;

[0050] Step 7: Mutation, specifically, for the particles in the population P, determine whether their overall average speed is less than V low, if true, perform mutation operation;

[0051] Step 8: If t <t max , go to step 2;

[0052] Step 9: Population P is the final Pareto optimal solution set.

[0053] Furthermore, performing the mutation operation specifically includes:

[0054] Copy population P to population P 0 ;

[0055] Calculate the population P 0 The position of the mutant particles in is:

[0056]

[0057] In the formula, v mut is the variation value, the variation coefficient α∈[0,1], V max is the maximum velocity of the specified particle, x i (t) is the position of the ith particle at time t, x i变异 (t) is the population P 0 The position of the i-th particle after the mutation operation;

[0058] Calculate the population P 0 The target values ​​of the particles in the population P are determined to determine whether they are non-dominated solutions. If so, the mutant particles are used to replace the corresponding original particles in the population P. Furthermore, the blast furnace state parameters are specifically the air volume, air temperature, top pressure, top temperature, actual wind speed, total pressure difference and actual coal injection volume in the last hour, and the blast furnace production indicators are specifically the gas utilization rate, blast furnace utilization coefficient, permeability index and molten iron silicon content.

[0059] The blast furnace optimal material level setting system provided by the present invention comprises:

[0060] A memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the method for setting the optimal charge level of a blast furnace provided by the present invention are implemented.

[0061] The method and system for setting the optimal charge surface of a blast furnace provided by the present invention perform three-dimensional reconstruction on an acquired charge surface image to obtain a three-dimensional charge surface shape of the blast furnace, perform feature extraction on the three-dimensional charge surface shape of the blast furnace to obtain charge surface shape features, establish a charge surface optimization process model for the blast furnace based on the relationship between the charge surface shape features and the state parameters of the blast furnace and the production index of the blast furnace, and solve the optimal solution set of the charge surface optimization process model for the blast furnace, and select the charge surface shape features corresponding to the optimal solution suitable for the current furnace condition from the optimal solution set as decision variables, thereby completing the setting of the optimal charge surface of the blast furnace, solving the technical problem of low precision of the existing method for setting the optimal charge surface of the blast furnace, and converting the setting problem of the optimal charge surface of the blast furnace into a production index optimization problem according to the intrinsic relationship between the production index of the blast furnace and the charge surface shape of the blast furnace and the state parameters of the blast furnace from the data perspective, and solve the decision variables related to the setting of the optimal charge surface of the blast furnace by establishing the charge surface optimization process model for the blast furnace, thereby accurately realizing the setting of the optimal charge surface of the blast furnace, and providing theoretical guidance for subsequent blast furnace charging.

[0062] Purpose of the present invention:

[0063] In view of the shortcomings in the current research on the optimal charge surface setting of blast furnaces, the purpose of the present invention is to perform three-dimensional reconstruction of the charge surface images taken by an industrial endoscope. On the basis of obtaining the three-dimensional shape of the blast furnace charge surface, according to the characteristics of the annular distribution of the charge at the throat, the blast furnace charge surface shape is divided into three areas: "edge", "platform" and "funnel". By analyzing the distribution law of the charge in each area and combining the expert experience, the multi-dimensional feature extraction of the charge surface shape is performed, so as to better characterize the blast furnace charge surface shape.

[0064] The purpose of the present invention is to convert the setting problem of the optimal charge surface of the blast furnace into a production index optimization problem based on the intrinsic connection between the blast furnace production index and the charge surface shape of the blast furnace from the data perspective. Since the indicators are both interrelated and different, it is difficult to ensure that the set charge surface shape is in the optimal state if only the optimization of a single indicator is considered. Therefore, on the basis of discretizing the charge surface and normalizing each indicator, the setting problem of the optimal charge surface of the blast furnace is converted into a multi-objective optimization problem.

[0065] The purpose of the present invention is to fully consider the potential relationship between the material surface characteristic parameters and state parameters and the production indicators, and propose a method of stacked denoising autoencoder (SA-DAE) based on the target attention mechanism to establish a material surface optimization process model. In addition, an improved AWPSO-NSGA-II algorithm is proposed for solving the model, which is convenient for obtaining a more comprehensive Pareto optimal solution set. The operator can select the optimal solution suitable for the current furnace condition from the solution set according to the real-time production demand, thereby realizing the setting of the optimal material surface of the blast furnace under the current production conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flow chart of a method for setting an optimal charge level for a blast furnace according to a second embodiment of the present invention;

[0067] Figure 2 The shape of the blast furnace charge surface after being processed by the blast furnace optimal charge surface setting method of the second embodiment of the present invention;

[0068] Figure 3 The material surface feature distribution model of the second embodiment of the present invention;

[0069] Figure 4 The specific structure of the A-DAE model of the second embodiment of the present invention;

[0070] Figure 5 The overall structure of the SA-DAE model of the second embodiment of the present invention;

[0071] Figure 6 This is a flow chart of the AWPSO-NSGA-II algorithm of the second embodiment of the present invention;

[0072] Figure 7 It is a structural block diagram of a system for setting an optimal material level for a blast furnace according to an embodiment of the present invention.

[0073] Reference numerals:

[0074] 10. Memory; 20. Processor. DETAILED DESCRIPTION

[0075] In order to facilitate the understanding of the present invention, the present invention will be described more comprehensively and meticulously below in conjunction with the accompanying drawings and preferred embodiments of the present invention, but the protection scope of the present invention is not limited to the following specific embodiments.

[0076] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0077] Embodiment 1

[0078] The method for setting the optimal charge level of a blast furnace provided in the first embodiment of the present invention comprises:

[0079] Step S101, performing three-dimensional reconstruction on the acquired material surface image to obtain a three-dimensional blast furnace material surface shape, and performing feature extraction on the three-dimensional blast furnace material surface shape to obtain material surface shape features;

[0080] Step S102, establishing a blast furnace charge surface optimization process model based on the charge surface shape characteristics and the relationship between the blast furnace condition parameters and the blast furnace production index;

[0081] Step S103, solving the optimal solution set of the blast furnace charge level optimization process model, and selecting the charge level shape characteristics corresponding to the optimal solution suitable for the current furnace condition from the optimal solution set as decision variables, thereby completing the setting of the blast furnace optimal charge level.

[0082] The method for setting the optimal charge surface of a blast furnace provided by an embodiment of the present invention performs three-dimensional reconstruction on an acquired charge surface image to obtain a three-dimensional charge surface shape of the blast furnace, performs feature extraction on the three-dimensional charge surface shape of the blast furnace to obtain charge surface shape features, establishes a blast furnace charge surface optimization process model and solves the optimal solution set of the blast furnace charge surface optimization process model based on the charge surface shape features and the relationship between the blast furnace condition state parameters and the blast furnace production index, and selects the charge surface shape features corresponding to the optimal solution suitable for the current furnace condition from the optimal solution set as decision variables, thereby completing the setting of the optimal charge surface of the blast furnace, solving the technical problem of low precision of the existing method for setting the optimal charge surface of the blast furnace, and converting the setting problem of the optimal charge surface of the blast furnace into a production index optimization problem according to the intrinsic connection between the blast furnace production index and the blast furnace charge surface shape and the blast furnace condition state parameters from a data perspective, and solves the decision variables related to the setting of the optimal charge surface of the blast furnace by establishing the blast furnace charge surface optimization process model, thereby accurately realizing the setting of the optimal charge surface of the blast furnace, and providing theoretical guidance for subsequent blast furnace charging.

[0083] Embodiment 2

[0084] Figure 1 As shown, the method for setting the optimal charge level of a blast furnace according to the second embodiment of the present invention comprises the following steps:

[0085] (1) Material surface feature extraction. The three-dimensional shape of the blast furnace material surface is obtained by performing three-dimensional reconstruction based on the image of the blast furnace material surface taken by the industrial endoscope. According to the characteristics of the multi-ring charging method, the charge is distributed in a ring shape at the furnace throat and can be divided into three areas: "edge", "platform" and "funnel". By analyzing the charge distribution in these three areas and combining expert experience, multi-dimensional feature extraction of the material surface shape is performed.

[0086] (2) Mathematical description of the optimal material surface setting problem. Blast furnace production indicators can characterize the operation of the blast furnace and are the external manifestation of the furnace status. In addition, actual production requires setting the corresponding material surface shape according to the production target. Therefore, the problem of setting the optimal material surface of the blast furnace can be transformed into a multi-objective optimization problem. The present invention takes the gas utilization rate and the blast furnace utilization coefficient as the optimization targets of the multi-objective optimization problem; takes the blast furnace permeability index and the silicon content of molten iron as constraints; and takes the material surface characteristic parameters as decision variables.

[0087] (3) Establishment of blast furnace charge surface optimization model. As a deep learning algorithm, denoising autoencoder has powerful feature extraction capabilities and good prediction results when performing multivariate nonlinear regression analysis. Considering that the blast furnace charge surface optimization model is a multi-input and multi-output model, a SA-DAE (Stacked Attention-based Denoising Autoencoder) method is proposed to establish a charge surface optimization process model. By introducing the attention mechanism to improve the original structure of DAE (Denoising Autoencoder), the model can better capture features that are highly correlated with the target output, thereby improving the accuracy and generalization ability of the model.

[0088] (4) Solving multi-objective optimization problems. For solving the material surface optimization problem, an adaptive weighted particle swarm optimization algorithm and a fast non-dominated sorting genetic algorithm are combined to obtain the Pareto optimal solution set. This algorithm combines the advantages of PSO and NSGA-II, ensuring the global search capability of each particle while maintaining the diversity of the population. Operators can select the optimal solution suitable for the current furnace conditions from the optimal solution set according to the actual needs of the production site, thereby completing the setting of the optimal material surface of the blast furnace.

[0089] Specifically, the specific implementation of each step in the second implementation of the present invention is as follows:

[0090] (1) Material surface shape feature extraction

[0091] Affected by the internal environment of the blast furnace, assuming that the material surface is symmetrically distributed about the center line of the blast furnace, the commonly used material surface shape description is to detect points with different radial distances relative to the furnace core, and then use an effective algorithm to fit the shape of the radial material line. However, this material surface shape description has many defects: (1) It only considers the shape of the material line in a single radial direction, ignoring the differences between the shapes of the material lines in each radial direction within the circumference; (2) Two points that are symmetrically distributed about the furnace core in the radial direction are very likely to be asymmetric in the actual material surface, that is, they are at different heights from the zero material line.

[0092] Taking into full consideration the deficiencies in the above-mentioned material surface shape representation, the specific steps of material surface shape feature extraction in the embodiment of the present invention are as follows:

[0093] 1) Perform 3D reconstruction on the material surface image taken by the industrial endoscope to obtain the 3D shape of the material surface, establish a right-handed coordinate system, and stipulate that the coordinate system Z is vertically downward from the center line of the blast furnace. w In the positive direction of the axis, the intersection of the blast furnace center line and the zero material line is the coordinate origin O, and the zero material line horizontally to the right is X w The positive direction of the axis, perpendicular to the paper and outward is Y w The coordinates of any point on the material surface are expressed as (x, y, z). It is the radial distance from any point on the material surface to the center line of the blast furnace, and z is the vertical distance from any point on the material surface to the zero material line.

[0094] 2) The charge is distributed in a ring shape from the furnace wall to the furnace core at the furnace throat. Let the radius of the furnace throat be R. Take any point on the ring with radius r = R as the starting point. Let the sampling interval be rad, you can collect n points on the current ring, and then is the sampling interval, and the m rings are sampled successively, and finally the information of m*n points can be collected on the material surface. Considering that the actual material surface shape is not completely symmetrical about the center line of the blast furnace, and there are abnormalities such as material surface collapse in some areas, the material surface is processed as follows. First, the box plot method is used to successively count the vertical distances of n points on i (i=m,m-1,...,1) rings from the zero material line, and the abnormal points are removed; then the remaining n points on the i-th ring are counted. i The z of the points ij The values ​​are averaged:

[0095]

[0096] In the formula, z ij is the vertical distance from the jth point on the ith ring to the zero material line, z i Represents the vertical distance between the ith ring and the zero material line, n i It indicates the number of normal values ​​after removing outliers in the i-th ring.

[0097] The z of each point on the material surface ij After averaging, a completely symmetrical blast furnace charge surface can be obtained with respect to the center line of the blast furnace, and the shapes of the charge lines in each radial direction are exactly the same, such as Figure 2 shown.

[0098] 3) According to the radial material line shape, the present invention divides the material surface into three areas: "edge", "platform" and "funnel". First, determine the inflection point position of each area, and determine whether it is in the slope range specified by the current area by calculating the slope between adjacent points in the radial direction. Once the slope of the adjacent points exceeds the specified range, it is considered that one of the adjacent points must be the inflection point of the adjacent area; when the inflection points of each area are determined, the feature extraction of the material surface shape is performed. Figure 3The definition method of the radial two-dimensional material line of the blast furnace is described. The setting rules of the coordinate axis are as follows: the position of the zero material line of the furnace throat is the x-axis, the furnace center is the coordinate origin, and the y-axis is vertically downward; finally, the 7-dimensional material surface features are extracted, which are: the vertical distance h from the zero material line to the platform 1 , furnace wall inclination angle α, platform width l 2 、Funnel width l 1 , funnel depth h 2 , funnel inclination angle β, funnel center angle θ. The calculation formula for the above 7-dimensional material surface shape characteristics is as follows:

[0099]

[0100] In the formula, R l is the radius of the arc where the "funnel" area is located.

[0101] It can be seen that when extracting the shape features of the material surface, the embodiment of the present invention fully considers the differences between the material line shapes in each radial direction within the circumference, and divides the blast furnace material surface shape into three areas of "edge", "platform" and "funnel" according to the characteristic of the annular distribution of the charge at the throat. By analyzing the distribution law of the charge in each area and combining the expert experience, the multi-dimensional feature extraction of the material surface shape is performed, so as to obtain the feature vector that accurately characterizes the shape of the blast furnace material surface, laying the foundation for the subsequent establishment of a blast furnace material surface optimization process model.

[0102] (2) Mathematical description of the optimal material surface setting problem

[0103] The blast furnace production index can characterize the operation of the blast furnace and is the external manifestation of the furnace condition. In addition, the actual production needs to comprehensively consider the various production indicators of the blast furnace to set the corresponding material surface shape, that is, the optimal material surface setting problem is essentially a multi-objective optimization problem. According to the characteristics of the blast furnace ironmaking process and combined with the smelting experience, the present invention determines the optimization objectives and constraints in the material surface optimization problem, and then summarizes the mathematical model of the blast furnace optimal material surface setting problem.

[0104] 1) Gas utilization rate refers to the conversion of CO into CO in the blast furnace 2 The ratio is an important indicator to characterize the operation status and energy consumption of the blast furnace, and is also the indicator that the furnace manager is most concerned about. The calculation formula is as follows:

[0105]

[0106] In the formula, CO 2 (%) indicates CO in coal gas 2 The percentage of CO content, CO (%) indicates the percentage of CO content in coal gas. The higher the coal gas utilization rate, the more complete the chemical reaction in the furnace and the more stable the furnace condition;

[0107] The blast furnace utilization coefficient is the average number of tons of qualified pig iron produced per cubic meter of blast furnace effective volume per day and night. It is an economic indicator used to comprehensively reflect blast furnace production technology, fuel consumption, etc. The calculation formula is as follows:

[0108]

[0109] In the formula, P is the pig iron output of the blast furnace in one day and night, and V is the effective volume of the blast furnace. The larger the blast furnace utilization coefficient, the higher the blast furnace productivity and the better the economic benefits.

[0110] The present invention comprehensively considers energy consumption index and economic index, selects the highest gas utilization rate and the maximum blast furnace utilization coefficient as the optimization target, and the objective function is expressed as follows:

[0111]

[0112] In the formula, f 1 (·),f 2 (·) respectively represent the gas utilization rate y 1 , blast furnace utilization coefficient y 2 The unknown nonlinear relationship with the material surface characteristic parameter X and state parameter Z.

[0113] 2) Stable and smooth operation of the blast furnace is both the most basic and the most important guarantee in blast furnace production, and smelting qualified molten iron is the ultimate goal of blast furnace ironmaking. Therefore, under the premise of stable and smooth operation of the furnace and guaranteed molten iron quality, further adjusting the blast furnace to optimize production is of practical significance.

[0114] The permeability index is an important indicator to characterize the smooth operation of the furnace. Too high or too low permeability will have an adverse effect on production. Therefore, in actual production, the permeability index needs to be controlled within a certain range.

[0115] The silicon content of molten iron refers to the percentage of silicon element in molten iron. It is an important indicator for evaluating the quality of pig iron and furnace conditions. In actual production, it is necessary to ensure that the quality of molten iron is maintained within an appropriate range.

[0116] The present invention selects the permeability index and the silicon content of molten iron as the constraint conditions of the multi-objective optimization problem, and its mathematical description is as follows:

[0117]

[0118]

[0119] In the formula, f 3 (·),f 4 (·) respectively represent the air permeability index y 3 、Silicon content in molten iron 4 The unknown nonlinear relationship between the material surface characteristic parameter X and the state parameter Z, They are the pre-set upper and lower thresholds for the permeability index and the silicon content in molten iron.

[0120] 3) After determining the optimization target and state parameters, appropriate surface feature parameters X and state parameters Z should be selected as the input of the model. The surface feature parameter X is the 7-dimensional surface shape feature extracted in (1), and it is used as the decision variable in the multi-objective optimization problem.

[0121] There are nearly a hundred parameters that can characterize the furnace condition in a blast furnace. If all parameters are input into the model, it will inevitably lead to problems such as high model complexity and low model accuracy due to parameter redundancy. Therefore, it is necessary to analyze the correlation between each state parameter and gas utilization rate, blast furnace utilization coefficient, permeability index and molten iron silicon content, and select state parameters that have a strong correlation with each production index. The mutual information method is used here for correlation analysis, and the calculation formula is as follows:

[0122]

[0123] Where I(Z,Y) is the mutual information between the state parameter Z and the production index Y, p(z,y) is the joint probability distribution of Z and Y, and p(z) and p(y) are the marginal probability distributions of Z and Y, respectively.

[0124] Specifically, in this embodiment, seven state parameters, namely, air volume, air temperature, top pressure, top temperature, actual wind speed, total pressure difference, and actual coal injection amount in the previous hour, are input into the material surface optimization model.

[0125] 4) Finally, the multi-objective optimization problem of blast furnace charge level can be described as:

[0126]

[0127] In the formula, and is the value range of the material surface characteristic parameters; y k =f k (X,Z) k=1,2,3,4 They are nonlinear models between four production indicators and material surface characteristic parameters X and state parameters Z respectively.

[0128] (3) Establishment of blast furnace charge level optimization model

[0129] The premise of blast furnace charge level optimization is to establish an accurate process model. However, due to the complexity of the blast furnace ironmaking process, its mechanism model is difficult to establish. With the rapid development of detection technology and computing power, a large amount of data in the blast furnace production process is recorded and stored. In recent years, data-driven deep learning models have been effective in dealing with nonlinear problems and have been widely used in modeling complex industrial processes such as blast furnace ironmaking. The present invention proposes a method of stacked denoising autoencoder (SA-DAE) based on a target attention mechanism to establish an optimal charge level setting model, and reasonably constructs the intrinsic connection between charge level feature parameters and state parameters and production indicators.

[0130] 1) Denoising Autoencoder (DAE) is a typical unsupervised neural network model. A DAE unit contains an input layer, a hidden layer, and an output layer, and consists of two processes: encoding and decoding. In order to enable the DAE model to extract implicit features that are highly correlated with the target output and determine the closeness between each feature and the output variable, an attention module is introduced between encoding and decoding, namely, the Attention-based Denoising Autoencoder (A-DAE) based on the target attention mechanism. Figure 4 The specific structure of A-ADE is given, and the establishment process of the model is as follows:

[0131] Step 1: Add random noise to the original input vector x to convert it into a noisy vector right Encode, generate hidden layer feature vector h, and The encoding function of the encoding is specifically:

[0132]

[0133] Among them, S(·) is the activation function, W and b are the weight matrix and bias vector from the input layer to the hidden layer respectively;

[0134] Step 2: Input the hidden layer feature vector h and the target output vector y into the attention module together, and calculate the correlation between each feature in h and each variable in y. The specific calculation formula is:

[0135]

[0136]

[0137]

[0138] Among them, θa = {W h ,W y ,b 0} is the parameter to be optimized in the attention module in the denoising autoencoder, where g(·) represents the activation function from the hidden layer to the attention module of the denoising autoencoder model based on the target attention mechanism, h represents the feature vector of the hidden layer, y represents the output vector of the denoising autoencoder model based on the target attention mechanism, and W h represents the weight matrix of the eigenvector h, W y represents the weight matrix of the output vector y, b 0 represents the bias vector from the hidden layer to the attention module, r is the attention weight vector before normalization, softmax is the normalized exponential function, β represents the normalized attention weight vector, d h represents the dimension of the feature vector h, r i represents the attention weight coefficient corresponding to the i-th feature in the feature vector h, represents the Hadamard product, h y is the feature vector based on target attention;

[0139] Step 3: The feature vector h based on the target attention y Mapped to the output layer, the reconstructed output vector is obtained The decoding function is:

[0140]

[0141] Among them, W′ and b′ are the weight matrix and bias vector from the attention module to the output layer respectively;

[0142] step4: In order to The BP algorithm is used to adjust the A-DAE model parameters θ e = {W, b, W′, b′} and θ a = {W h ,W y ,b a}Continuously optimize to minimize the reconstruction loss. The loss function is as follows:

[0143]

[0144] Where N represents the total number of training samples;

[0145] 2) After the first A-DAE training is completed, the parameters of the encoding process {W 1 ,b 1} and the attention weight vector α 1 Save it and put h″ 1As the input vector of the second A-DAE. At this time, repeat the training steps in 1) to complete the training of the second A-DAE, and set the parameters of the encoding process {W 2 ,b 2} and the attention weight vector α 2 Save it. And so on, until the kth A-DAE training is completed. At this point, all A-DAEs are stacked in sequence to form SA-DAE. Figure 5 The overall structure of SA-DAE is given.

[0146] 3) After completing the layer-by-layer unsupervised pre-training of SA-DAE, an output layer is added after the hidden layer of the kth A-DAE, and the output function is:

[0147]

[0148] Where W o ,b o are the weight matrix and bias vector from the Kth A-DAE hidden layer to the output layer, respectively. is the final prediction output of the SA-DAE model, represents the hidden feature vector extracted by the Kth denoising autoencoder based on the target attention mechanism.

[0149] In order to enhance the generalization performance of the model, we first o = {W o ,b o} initialization, and then use the BP algorithm to train the parameters θ of the SA-DAE model i = {W i ,b i} i=1,2,...,K and θ o Fine-tune to minimize the prediction error. The error function is as follows:

[0150]

[0151] In the formula, y i and are the true value and predicted value of the i-th training sample respectively.

[0152] So far, the blast furnace charge level optimization model based on SA-DAE has been established.

[0153] It is not difficult to see that the embodiment of the present invention introduces an attention module between the encoding and decoding of the denoising autoencoder, and inputs the hidden layer feature vector and the target output vector of the denoising autoencoder based on the target attention mechanism into the attention module together, which can enhance the correlation between the extracted material surface shape features and the output vector, and can also distinguish the importance between the features, thereby establishing a high-precision blast furnace material surface optimization process model, and laying the foundation for the subsequent precise setting of the optimal blast furnace material surface, and for the subsequent use of optimization algorithms to solve the model, and find the optimal material surface feature parameters. A reliable foundation has been laid. Specifically, compared to shallow neural networks such as denoising autoencoders, stacked denoising autoencoders belong to multi-layer neural networks, which can well learn the complex relationship between input and output, and continuously extract deeper features from the previous hidden layer through the current hidden layer, so as to more accurately represent the input. However, further analysis of the stacked denoising autoencoder revealed the following problems: (1) The ultimate goal of using this model is to use the extracted hidden layer features to achieve prediction tasks such as regression or classification, while the typical stacked denoising autoencoder continuously extracts features from the original input and does not include any features related to the output; (2) For each hidden layer feature vector, each feature has a different degree of correlation with the target output, and each feature is assigned the same weight, which will cause important features to be diluted in the layer-by-layer extraction process, and unimportant features are continuously strengthened, resulting in unsatisfactory final prediction results of the model. Therefore, in view of the above problems, it is necessary to introduce output information when the model extracts features. The embodiment of the present invention introduces an attention module between the encoding and decoding of the denoising autoencoder, and inputs the hidden layer feature vector and the target output vector of the denoising autoencoder based on the target attention mechanism into the attention module together, which can enhance the correlation between the input vector and the output vector, and can also distinguish the importance of each feature in the feature vector, thereby establishing a high-precision blast furnace material surface optimization process model.

[0154] (4) Solving multi-objective optimization problems

[0155] Particle swarm optimization (PSO) is a widely used heuristic search method in recent years. It can obtain a satisfactory optimal solution with a relatively low computational cost and is generally applicable to the field of single-objective optimization. The second-generation non-dominated sorting genetic algorithm (NSGA-II) introduces the idea of ​​fast non-dominated sorting and elite strategy, which not only ensures the uniform distribution of the optimal solution on the Pareto optimal frontier and the diversity of the population, but also improves the overall quality of the population and the convergence speed of the algorithm. It is very popular in solving multi-objective optimization problems. Therefore, an AWPSO-NSGA-II algorithm combining an adaptive weighted particle swarm algorithm and a fast non-dominated sorting genetic algorithm is proposed to solve the material surface optimization problem.

[0156] Based on the SA-DAE model training, the present invention converts the material surface feature vector x=[h 1 ,h 2 ,l 1 ,l 2 ,α,θ,β] T ∈R 7×1 As a particle, that is, the decision variable of the optimization problem; then the objective function F(x,z) is continuously optimized through the AWPSO-NSGA-II algorithm. At this time, the state parameter y takes the current detection value and does not participate in the optimization process; finally, a Parteo optimal solution set is obtained to maximize the value of the objective function. The operator can select the optimal solution suitable for the current production needs from the solution set as the initial value for setting the optimal material surface.

[0157] The overall process of the AWPSO-NSGA-II algorithm is as follows Figure 6 As shown, the specific calculation steps are as follows:

[0158] 1) Assume that the size of the population P is N and the number of iterations is t max , randomly initialize the particle position within the set value range, V max is the maximum allowed speed, initialize the individual optimal position and the global optimal position, that is, p best =g best =0, current iteration number t=0;

[0159] 2) Calculation: Calculate the target values ​​of all particles in the current population P;

[0160] 3) Update the optimal particle; let t = t + 1, and use the dynamic weighted method to calculate the fitness of each particle. The calculation formula is as follows:

[0161]

[0162] In the formula, p is the current particle fitness, M is the number of targets to be optimized, and f j (x) is the function value of the particle on the jth objective function, ω j is the weight assigned to the particle on the jth target, and

[0163] The particle with the largest dynamic fitness is selected as the global optimal particle, and the individual optimal particle is selected by comparing the current particle position with the individual historical optimal position, thus completing g best and p best Updates;

[0164] 4) Generate new particles; copy P to form a new population Q, calculate the updated speed and position of each particle in population Q according to equations (19)-(21), and calculate the target value of each particle in Q, combine P and Q into population R (the number of particles is 2N), and the calculation formula is as follows:

[0165]

[0166] In the formula, v i (t) is the flying speed of the ith particle at time t, ω is the inertia weight of the particle, g best is the position of the global optimal particle, p best is the position of the individual optimal particle, v i (t-1) is the flying speed of the i-th particle at time t-1, x i (t) is the position of the ith particle at time t, x i (t-1) is the position of the i-th particle at time t-1, r 1 and r 2 is a random number uniformly distributed in [0,1], ω is the inertia weight, ρ is a random number distributed in [0,1], c 1 and c 2 They are individual learning factor and social learning factor.

[0167] In order to overcome the inherent defects of the PSO algorithm that it is easy to fall into local optimal solutions and premature convergence, this algorithm introduces a dynamically changing inertia weight:

[0168] ω=ω end +ρ(ω start -ω end ) (20)

[0169] where ω start is the initial inertia weight, ω end is the convergence inertia weight, ω end =[0,0.5],ω start >ω end And they are all constants, and ρ is a random number distributed in [0,1].

[0170] In addition, in order to improve the local search capability of the particle swarm in the early stage and the global search capability in the later stage, this algorithm makes the following changes to the learning factor:

[0171]

[0172] In the formula, c 11 and c 12 They are c 1 The initial value and convergence value of c 21 and c 22 They are c2 The initial value and convergence value of , t is the current iteration number, t max is the total number of iterations, and c 11 =c 21 =2.5, c 12 =c 22 =0.5.

[0173] 5) Sorting: Use the non-dominated sorting idea of ​​NSGA-II to perform non-dominated sorting and crowding distance sorting on the population R;

[0174] 6) Select the next generation of particles; select N particles from the population R in sequence to the population P;

[0175] 7) Mutation: For particles in population P, determine whether their overall average velocity is less than V low , if true, perform the following mutation operation:

[0176] (a) Copy population P to population P 0 ;

[0177] (b) Calculate the population P 0 The position of the mutant particles in is:

[0178]

[0179] In the formula, v mut is the variation value, the variation coefficient α∈[0,1], V max is the maximum velocity of the specified particle, x i (t) is the position of the ith particle at time t, x i变异 (t) is the population P 0 The position of the i-th particle after the mutation operation;

[0180] (c) Calculate P 0 The target values ​​of the particles in the population P are determined to determine whether they are non-dominated solutions. If so, the corresponding original particles in the population P are replaced by mutant particles;

[0181] 8): If t <t max , go to step 2;

[0182] 9): Population P is the final Pareto optimal solution set.

[0183] It is not difficult to see that the basic idea of ​​the particle swarm algorithm is to randomly initialize a group of particles, and update the individual optimal position and the global optimal position of the particles through continuous iteration, so as to find the optimal solution to the problem. The algorithm has memory, fewer parameters, simple and efficient, but still has the following shortcomings: (1) The algorithm lacks dynamic adjustment of particle speed, and it is easy to fall into the local optimal solution, resulting in poor final convergence effect; (2) For different types of problems, how to choose a set of appropriate particle parameters to maximize the optimization effect still depends on subjective experience to a large extent, and there is no unified parameter adjustment standard. In response to the above problems, the present invention adjusts the inertia weight ω, the learning factor c 1 and c 2 The three parameters are adaptively adjusted, where ω is used to balance the global and local search performance. The larger ω is, the stronger the global search ability of the particle is; the smaller ω is, the stronger the local search ability of the particle is. By adaptively adjusting ω, the local and global search capabilities of the particle can be improved; the learning factor c 1 and c 2 Adjust the particles to g best and p best The maximum step size in the direction determines the flight trajectory of the particles. In the initial stage, the learning factor is made larger to keep the particles at a high search capability. As the number of iterations increases, the adjustment step size of the particles is continuously reduced to make the particles tend to the global optimal position as much as possible. Although the improved particle swarm algorithm can find an optimal solution faster and better, a solution usually has great limitations for solving multi-objective optimization problems such as optimal material surface setting. Therefore, the embodiment of the present invention combines the ideas of non-dominated sorting and crowding distance sorting of the NSGA-II algorithm with the particle swarm algorithm to obtain a Pareto optimal solution set, and improves the mutation conditions and specific operations in order to maintain the diversity of the population. At this point, a comprehensive Pareto optimal solution set is obtained, and the operator can select the optimal solution suitable for the current furnace conditions from the solution set according to the current production needs, that is, the material surface shape characteristic value as the decision variable.

[0184] The proposed method for setting the optimal material surface of a blast furnace based on SA-DAE optimized by AWPSO-NSGA-II reconstructs the material surface image taken by an industrial endoscope in three dimensions, and extracts multidimensional features of the three-dimensional shape of the blast furnace material surface based on the distribution law of the charge. Then, based on the potential relationship between the blast furnace production index and the blast furnace material surface shape, the optimal material surface setting problem is transformed into a multi-objective optimization problem, and the material surface optimization problem is fully described mathematically. Next, a SA-DAE method is proposed to establish a material surface optimization process model. Finally, an improved AWPSO-NSGA-II algorithm is used to obtain the Pareto optimal solution set. The operator can select the optimal solution suitable for the current furnace condition from the solution set according to the real-time production demand, and select the material surface shape characteristics corresponding to the optimal solution suitable for the current furnace condition from the optimal solution set as the decision variable, thereby completing the setting of the optimal material surface of the blast furnace.

[0185] Embodiment 3

[0186] The optimal material level setting method proposed in this invention is based on the material level of a blast furnace, which can be applied to a 2650m 3 The blast furnace charge distribution operation is as follows: First, a blast furnace charge level optimization model based on SA-DAE is established, then an improved particle swarm algorithm is used to solve the model, and then the optimal solution suitable for the current furnace condition is selected from the Pareto optimal solution set as the initial value of the optimal charge level.

[0187] The steps of the method for setting the optimal charge level of a blast furnace according to the third embodiment of the present invention are as follows:

[0188] 1. The blast furnace material surface image taken by the industrial endoscope is reconstructed in three dimensions, the three-dimensional material surface shape is processed, and the accurate and reliable radial material line shape is extracted. The material surface shape is divided into three areas: "edge", "platform" and "funnel", and the multi-dimensional material surface shape features are extracted;

[0189] 2. Convert the optimal material level setting problem into a multi-objective optimization problem, and summarize the mathematical model of the optimal material level setting problem for blast furnaces;

[0190] 3. A SA-DAE method is proposed to establish the optimal material surface setting model, and the intrinsic relationship between the material surface characteristic parameters and state parameters and production indicators is reasonably constructed;

[0191] 4. An AWPSO-NSGA-II algorithm is used to solve the multi-objective optimization problem and obtain a Pareto optimal solution set;

[0192] 5. Based on real-time production needs, the operator selects the optimal solution suitable for the current furnace conditions from the optimal solution set as the set value of the initial material level.

[0193] At this point, the setting of the optimal material level for the blast furnace has been completed.

[0194] Reference Figure 7 The blast furnace optimal material level setting system proposed in the embodiment of the present invention comprises:

[0195] A memory 10, a processor 20, and a computer program stored in the memory 10 and executable on the processor 20, wherein the processor 20 implements the steps of the method for setting the optimal material level of a blast furnace proposed in this embodiment when executing the computer program.

[0196] The specific working process and working principle of the blast furnace optimal material level setting system in this embodiment can refer to the working process and working principle of the blast furnace optimal material level setting method in this embodiment.

[0197] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for setting the optimal charge level of a blast furnace, characterized in that: The method comprises: Performing three-dimensional reconstruction on the acquired material surface image to obtain the three-dimensional blast furnace material surface shape, and performing feature extraction on the three-dimensional blast furnace material surface shape to obtain material surface shape features; Based on the relationship between the material surface shape characteristics and the blast furnace condition parameters and the blast furnace production index, a blast furnace material surface optimization process model is established, wherein based on the relationship between the material surface shape characteristics and the blast furnace condition parameters and the blast furnace production index, the blast furnace material surface optimization process model is established including: Determine the material surface shape characteristics and blast furnace condition parameters as the original input vectors of the blast furnace material surface optimization process model; Determine the blast furnace production index as the target output vector of the blast furnace charge level optimization process model; Based on the original input vector and the target output vector, a noise reduction automatic encoder is used to establish a blast furnace charge level optimization process model, wherein based on the original input vector and the target output vector, a noise reduction automatic encoder is used to establish a blast furnace charge level optimization process model including: An attention module is introduced between the encoding and decoding of the denoising autoencoder to obtain a denoising autoencoder model based on the target attention mechanism. The original input vector is used as the input vector of the denoising autoencoder model based on the target attention mechanism, and the hidden layer feature vector and the target output vector of the denoising autoencoder model based on the target attention mechanism are input into the attention module together, and the correlation between each feature in the hidden layer feature vector and each variable in the target output vector is calculated. The specific calculation formula is: Where g(·) represents the activation function from the hidden layer to the attention module of the denoising autoencoder model based on the target attention mechanism, h represents the feature vector of the hidden layer, y represents the output vector of the denoising autoencoder model based on the target attention mechanism, and W h represents the weight matrix of the eigenvector h, W y represents the weight matrix of the output vector y, b0 represents the bias vector from the hidden layer to the attention module, r is the attention weight vector before normalization, softmax is the normalized exponential function, β represents the normalized attention weight vector, d h represents the dimension of the feature vector h, r i represents the attention weight coefficient corresponding to the i-th feature in the feature vector h, represents the Hadamard product, h y is the feature vector based on target attention; The feature vector h based on the target attention y Map to the output layer to obtain the reconstructed output vector; With the goal of minimizing the reconstruction loss, the BP algorithm is used to adjust the model parameters θ of the denoising autoencoder based on the target attention mechanism. e = {W, b, W′, b′} and θ a = {W h ,W y ,b0} for training, where W and b represent the weight matrix and bias vector from the input layer to the hidden layer of the denoising autoencoder model based on the target attention mechanism, respectively, and W′ and b′ represent the weight matrix and bias vector from the hidden layer to the output layer of the denoising autoencoder model based on the target attention mechanism, respectively; Stacking multiple trained denoising autoencoder models based on the target attention mechanism in sequence to obtain a stacked denoising autoencoder model based on the target attention mechanism; Add an output layer after the last hidden layer of the stacked denoising autoencoder model based on the target attention mechanism to obtain a complete model of the blast furnace charge level optimization process; The optimal solution set of the blast furnace charge surface optimization process model is solved, and the charge surface shape characteristics corresponding to the optimal solution suitable for the current furnace conditions are selected from the optimal solution set as decision variables, thereby completing the setting of the blast furnace's optimal charge surface.

2. The method for setting the optimal charge level of a blast furnace according to claim 1, characterized in that: The feature extraction of the three-dimensional blast furnace material surface shape is carried out, and the material surface shape features obtained include: Based on the right-hand rule, establish a three-dimensional material surface coordinate system; At a first preset sampling interval, the charge is sampled successively from the furnace wall to the furnace core at the furnace throat in a circular shape to obtain a circular sampling set; At a second preset sampling interval, sampling points on each circular ring in the circular ring sampling set is performed successively to obtain sampling points corresponding to the circular ring; Calculate the mean of the vertical distances between the sampling points corresponding to the circular ring and the zero material line, and perform mean processing on the sampling points corresponding to the circular ring according to the mean, so as to obtain a symmetrical blast furnace charge surface that is completely symmetrical about the center line of the blast furnace; According to the slope of the radial material line of the symmetrical blast furnace material surface between adjacent points in the radial direction, the radial material line is divided into three areas, and the three areas are specifically an edge area, a platform area and a funnel area; According to the three regions of the radial material line, the feature extraction of the three-dimensional blast furnace material surface shape is performed to obtain the material surface shape feature.

3. The method for setting the optimal charge level of a blast furnace according to claim 2, characterized in that: According to the three regions of the radial material line, the three-dimensional blast furnace material surface shape is extracted, and the material surface shape features obtained include: Establish a radial material line coordinate system, wherein the radial material line coordinate system takes the position of the furnace throat zero material line as the x-axis, the furnace center as the coordinate origin, and the y-axis vertically downward; Based on the radial material line coordinate system and the three regions of the radial material line, the material surface shape features are extracted. The material surface shape features specifically include the vertical distance from the zero material line to the platform, the furnace wall inclination, the platform width, the funnel width, the funnel depth, the funnel inclination and the funnel center angle.

4. The method for setting the optimal charge level of a blast furnace according to any one of claims 1 to 3, characterized in that: The optimal solution set for solving the blast furnace charge level optimization process model includes: Step 1: Assume that the size of the population P is N and the number of iterations is t max , randomly initialize the particle position within the set value range, V max is the maximum allowed speed, initialize the individual optimal position and the global optimal position, that is, p best =g best =0, current iteration number t=0; Step 2: Calculation, specifically calculating the target values ​​of all particles in the current population P; Step 3: Update the optimal particle. Specifically, let t = t + 1 and use the dynamic weighted method to calculate the dynamic fitness of each particle. The calculation formula is as follows: In the formula, p is the current particle fitness, M is the number of targets to be optimized, and f j (x) is the function value of the particle on the jth objective function, ω j is the weight assigned to the particle on the jth target, and The particle with the largest dynamic fitness is selected as the global optimal particle, and the individual optimal particle is selected by comparing the current particle position with the individual historical optimal position, thus completing g best and p best Updates; Step 4: Generate new particles, specifically by copying P to form a new population Q, calculating the updated speed and position of each particle in population Q, and calculating each target value of each particle in Q, combining P and Q into population R, with a particle number of 2N. The calculation formula is as follows: In the formula, v i (t) is the flying speed of the i-th particle at time t, ω is the inertia weight of the particle, g best is the position of the global optimal particle, p best is the position of the individual optimal particle, v i (t-1) is the flying speed of the i-th particle at time t-1, x i (t) is the position of the ith particle at time t, x i (t-1) is the position of the i-th particle at time t-1, r1 and r2 are random numbers uniformly distributed in [0,1], ω is the inertia weight, and ω=ω end +ρ(ω start -ω end ), where ω start is the initial inertia weight, ω end is the convergence inertia weight, ω end =[0,0.5],ω start >ω end and are all constants, ρ is a random number distributed in [0,1], c1 and c2 are individual learning factors and social learning factors respectively, and: In the formula, c 11 and c 12 are the initial value and convergence value of c1, c 21 and c 22 are the initial value and convergence value of c2, t is the current iteration number, t max is the total number of iterations, and c 11 =c 21 =2.5, c 12 =c 22 =0.5; Step 5: Sorting, specifically using the non-dominated sorting idea of ​​NSGA-II to perform non-dominated sorting and crowding distance sorting on the population R; Step 6: Select the next generation of particles, specifically select N particles from the population R in sequence to the population P; Step 7: Mutation, specifically, for the particles in the population P, determine whether their overall average speed is less than V low , if true, perform mutation operation; Step 8: If t<t max , go to Step 2; Step 9: Population P is the final Pareto optimal solution set.

5. The method for setting the optimal charge level of a blast furnace according to claim 4, characterized in that: The mutation operations include: Copy population P to population P0; Calculate the position of the mutant particles in the population P0, specifically: In the formula, v mut is the variation value, the variation coefficient α∈[0,1], V max is the maximum velocity of the specified particle, x i (t) is the position of the ith particle at time t, x i变异 (t) is the position of the i-th particle in the population P0 after the mutation operation; Calculate the target values ​​of the particles in population P0 and determine whether it is a non-dominated solution. If so, replace the corresponding original particles in population P with mutant particles.

6. The method for setting the optimal charge level of a blast furnace according to claim 5, characterized in that: The specific parameters of blast furnace condition are air volume, wind temperature, top pressure, top temperature, actual wind speed, and total pressure difference; the specific blast furnace production indicators are gas utilization rate, blast furnace utilization coefficient, permeability index, and silicon content in molten iron.

7. A blast furnace optimal charge level setting system, the system comprising: A memory (10), a processor (20), and a computer program stored in the memory (10) and executable on the processor (20), characterized in that the processor (20) implements the steps of any of the methods described in claims 1 to 6 when executing the computer program.

Citation Information

Patent Citations

  • Pile-up noise reduction own coding network bearing fault diagnosis method based on particle swarm optimization

    CN106682688A

  • Method for adjusting shape of material distribution surface of blast furnace

    CN112853014A

  • Multi-information fusion modeling method for shapes of burden surfaces in burden distribution process of blast furnace

    CN105950807A

  • Optimization method of burden distribution matrix of blast furnace burden distribution process having integer programming problem

    CN107119159A

  • Blast furnace charge level optimization method based on data and knowledge dual drive

    CN112861276A