A machine learning-based design method and system for a smelting cyclone dry powder lance

Abstract

The present application relates to the technical field of submerged top-blowing lance, and discloses a smelting cyclone dry powder lance design method and system based on machine learning, which comprises the following steps: optimizing the structural parameters of the cyclone chaotic dry powder lance by combining CFD-DEM and ANN technology; simulating and calculating the cyclone chaotic dry powder lance under different structural parameter configurations and extracting feature data; training an artificial neural network model according to the feature data; predicting target parameters by the artificial neural network model, and selecting, crossing and mutating the parameters by a genetic algorithm (GA) according to the difference between the predicted value and the target parameter value to obtain optimal parameters in terms of fitness; and simulating and analogizing according to the optimal parameters in terms of fitness. The present application clarifies the influence of structural parameters such as the lance blade height, blade rotation angle, blade installation height, CH4 pipe diameter and mixing zone length on the cyclone intensity and turbulent intensity, and finds the optimal parameter combination on this basis, thereby improving the performance and service life of the lance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of submerged top-submerged lance, and particularly relates to a smelting cyclone dry powder lance design method and system based on machine learning. BACKGROUND

[0002] The top-submerged lance (TSL) process is an immersion injection method by which a high flow of jet or jet-particle is injected into a high-temperature molten pool through a specially designed vertically installed lance. The jet-particle coaxial cyclone is concerned in various industrial processes due to its low load and good momentum / heat transfer efficiency. In the smelting process, the top-submerged lance with a cyclone converts the axial jet into chaotic cyclone and then injects the particles into the high-temperature molten pool. The particles in the reactor present helical motion under the disturbance of the cyclone, which enhances the momentum and mass transfer performance between phases and improves the melting and reaction rate of the particles. However, in the actual application process, the tangential entrainment of the gas phase to the solid particles and the four-way coupling between phases have strong anisotropic behavior. The chaotic nonlinear characteristics of the helical motion of the gas-solid two-phase flow lead to the fact that the design of the industrial equipment and the optimization of the process largely depend on the actual experience. Therefore, a comprehensive understanding of the motion mechanism, flow pattern, gas phase cyclone intensity, particle dispersion and turbulent flow structure of the coaxial gas-solid cyclone is conducive to improving the momentum, heat and mass transfer efficiency between the gas phase and the solid phase and optimizing the reactor device.

[0003] At present, the coaxial jet-particle cyclone is mainly studied by experimental measurement to explore its flow pattern and particle behavior, and analyze the factors affecting the cyclone intensity and particle dispersion characteristics. However, due to the complexity of local and instantaneous details of multiphase information in the industrial scale reactor and the high cost and long cycle of measurement, many scholars begin to use the coupling of computational fluid dynamics (CFD) and discrete element method (DEM) to study the particle-particle or particle-wall interaction. It is found that the cyclone entrained particles accelerate the motion and improve the operation performance of the reactor, and the cyclone intensity directly affects the motion and dispersion characteristics of the particles in the reactor. The nonlinear motion characteristics of the cyclone lead to the multi-scale turbulent motion characteristics in the reactor, and changing the structure of the cyclone will make the turbulent motion more complex. This method provides in-depth and comprehensive insights into the basic physical laws of the gas-solid cyclone system.

[0004] In the foregoing description, although the CFD-DEM coupling technology can simulate the particle-fluid interaction and the particle motion trajectory in detail, it may consume a large amount of computing resources and time in dealing with highly nonlinear and complex multiphase flow problems, which limits the application efficiency of the CFD-DEM simulation in design optimization, and there are still limitations. Therefore, the present application introduces the artificial neural network (ANN) technology in machine learning to combine with CFD-DEM to deeply understand the influence of the geometric structure of the cyclone chaotic dry powder spray gun on the cyclone intensity and the turbulence intensity of the spray gun, and to establish the nonlinear mapping relationship between the two, so as to optimize the structure of the spray gun and improve its performance and service life.

[0005] Prior art one, Chinese patent, application number: CN201910751432.9 discloses a Laval jet submerged top blowing lance, relates to non-ferrous metal smelting technology field, including hollow gun body, the gun body is cylindrical, one end of the gun body is closed end, the other end is open end, characterized in that, the inner cavity of the gun body is divided into mixing section, conveying section, cyclone section and turbulent section from closed end to open end in turn, a fuel pipe is coaxially inserted into the gun body from the center of the closed end and extends to the turbulent section, a pressure guide pipe is provided outside the fuel pipe in the gun body, an end pressure sampler is installed outside the closed end of the gun body, and the end pressure sampler is inserted into the gun body and connected with the pressure guide pipe;At least two fluid interfaces are opened on the side wall of the mixing section of the gun body;At least one Laval jet device is installed in the gun body, and the Laval jet device is located between any two adjacent fluid interfaces;A cyclone is installed in the gun body of the cyclone section. Although the structure is simple, the fluid is mixed uniformly, it is safe and reliable, and the service life is long, but when dealing with highly nonlinear and complex multiphase flow problems, a large amount of computing resources and time are consumed.

[0006] Prior art two, Chinese patent, application number: CN202110223827.9 discloses an immersed top blowing converter, which comprises a top blowing converter body, a material inlet and outlet are arranged at the top of the top blowing converter body, and an immersed top blowing lance is inserted into the material inlet and outlet. The reaction efficiency of the furnace is high, the immersed top blowing bath smelting is used, the material can be continuously added into the furnace, the furnace shutdown time for adding material is reduced, and the smelting operation rate is improved. Since the immersed top blowing lance can be directly inserted into the molten pool, and the immersed top blowing lance is provided with a cyclone, the stirring of the immersed top blowing lance on the molten pool is violent, the material mixing is more uniform, and the reaction efficiency in the furnace is improved. The top blowing converter is provided with a probe, which can take samples from the molten material at any time, reduce the sampling frequency of the furnace shutdown, keep the furnace temperature constant, and prolong the furnace life. The support of the top blowing converter adopts a combination structure of a support ring and a support roller, and adopts gear transmission, which makes the transmission more stable and overcomes the vibration impact of the furnace body, but at the same time, the stirring ability to the liquid phase area is also weakened.

[0007] The prior art three, Chinese patent, application number: CN202010588955.9 discloses a method for modifying a top-blown smelting facility, a continuous copper smelting facility and method, the method for modifying the top-blown smelting facility includes: cancelling the converter or the converter of the top-blown smelting facility; reducing the hearth height of the top-blown smelting furnace of the top-blown smelting facility; cancelling the single immersion type top-blown main spray gun of the top-blown smelting furnace; arranging at least three top-blown smelting spray guns at the top of the top-blown smelting furnace and spacing them apart; adding a copper matte discharge port on the top-blown smelting furnace; cancelling the electrode and electrode matching device of the settling electric furnace of the top-blown smelting facility, arranging at least three top-blown converter spray guns at the top of the settling electric furnace and spacing them apart, so as to transform the settling electric furnace into a top-blown converter; and connecting the copper matte discharge port of the top-blown smelting furnace with the copper matte inlet of the top-blown converter through a flow channel. The method for modifying the top-blown smelting facility can make full use of the existing top-blown smelting facility, is also conducive to producing high-grade copper matte, creates conditions for continuous blowing, and can also reduce the investment cost of continuous copper smelting, but the friction coefficient of the annular pipe wall and the outer wall surface of the center pipe of the spray gun increases rapidly, thereby causing short service life of the spray gun.

[0008] At present, the prior art one, the prior art two and the prior art three have the problems of consuming a large amount of computing resources and time when processing highly nonlinear and complex multiphase flow problems, weakening the stirring ability of the liquid phase area, and causing the friction coefficient of the annular pipe wall and the outer wall surface of the center pipe of the spray gun to increase rapidly, thereby causing short service life of the spray gun. To solve the above problems, the present application provides a smelting cyclone dry powder spray gun design method and system based on machine learning. SUMMARY

[0009] The main purpose of the present application is to provide a smelting cyclone dry powder spray gun design method and system based on machine learning to solve the problems of consuming a large amount of computing resources and time when processing highly nonlinear and complex multiphase flow problems in the prior art, weakening the stirring ability of the liquid phase area, and causing the friction coefficient of the annular pipe wall and the outer wall surface of the center pipe of the spray gun to increase rapidly.

[0010] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0011] A smelting cyclone dry powder spray gun design method based on machine learning, the molten pool smelting cyclone chaotic dry powder spray gun structure parameter optimization method, including:

[0012] Simulate the cyclone chaotic dry powder spray gun under different structure parameter configurations using CFD-DEM coupling software; extract the key characteristic parameters of the flow field in the CFD-DEM coupling software, calculate the ratio of the tangential momentum to the axial momentum of the fluid through the key characteristic parameters, and form the target parameters;

[0013] The flow field characteristic parameters are matched with target parameters, a data set is constructed, the data set is cleaned, the data set is divided into a training set and a verification set, an artificial neural network model is trained according to the training set, and the trained artificial neural network model is evaluated by using the verification set;

[0014] The influence of each hyperparameter in the artificial neural network model on the performance of the model is analyzed, the structural parameter that has the greatest influence on the target parameter is determined, and a parameter combination is set; the target parameter of the parameter combination is calculated, and the individual is selected according to the calculation result to perform cross and mutation, and the parameter combination with the optimal fitness is formed;

[0015] The maximum number of iterations or the convergence of the fitness function is calculated, if the new model combination is simulated and verified, if not, the cross and mutation are continued.

[0016] As a further improvement of the application, the process of forming the target parameter comprises:

[0017] The rotational flow chaotic dry powder spray gun is locally encrypted, and the rotational flow blade boundary and the key positions such as the spray gun outlet are simulated and calculated by using CFD-DEM coupling software under different structural parameter configurations;

[0018] The key characteristic parameters of the flow field are extracted in the CFD-DEM software;

[0019] The key characteristic parameters include the density, coordinates, tangential velocity U tan , axial velocity U axi , inner diameter R of the annular pipe of the spray gun, range from R0 to R1 of the annular zone to the mixing zone, characteristic length L of the spray gun, velocity u of the jet and integral volume V;

[0020] The ratio SN of the tangential momentum flux to the axial momentum flux of the fluid and the turbulent intensity of the jet in the spray gun are calculated.

[0021] As a further improvement of the application, the process of calculating the ratio SN of the tangential momentum flux to the axial momentum flux of the fluid and the turbulent intensity of the jet in the spray gun comprises:

[0022] The key characteristic parameters of the flow field are extracted in the CFD-DEM software;

[0023] The ratio SN of the tangential momentum flux to the axial momentum flux of the fluid is calculated; the equation of SN is expressed as:

[0024]

[0025] In the formula, SN local represents the ratio of the tangential momentum flux to the axial momentum flux, represents the momentum distribution characteristics of the local area, and G angular,localG represents the tangential momentum flux of a local region, represents the momentum transfer of the fluid in the tangential direction, G linear,local G represents the axial momentum flux of a local region, represents the momentum transfer of the fluid in the axial direction, R represents the radius, usually refers to the characteristic radius of the flow field or the radius of a certain region, p represents the density of the fluid, represents the mass of the fluid per unit volume, r represents the radial distance, represents the distance from the center of the flow field to a point, U axi U represents the axial velocity, represents the velocity component of the fluid in the axial direction, U tan U represents the tangential velocity, represents the velocity component of the fluid in the tangential direction, R0 represents the lower limit of integration, usually represents the inner boundary radius of the flow field, R1 represents the upper limit of integration, usually represents the outer boundary radius of the flow field;

[0026] The turbulent Reynolds number Re and turbulent Reynolds TRe of the jet inside the lance are introduced to evaluate the turbulent intensity of the jet inside the lance; wherein, Re is defined as the ratio of inertial force to friction force, and TRe is defined as the ratio of fluid turbulent kinetic energy k g to friction work W, and its expression is listed as:

[0027]

[0028]

[0029] In the formula, Re represents the Reynolds number, represents the ratio of inertial force to viscous force, and is used to evaluate the turbulent intensity of the fluid, p g represents the density of the gas phase fluid, represents the mass per unit volume of the gas phase fluid, u g represents the characteristic velocity of the gas phase fluid, represents the velocity of the gas phase fluid over a certain characteristic length, L represents the characteristic length, represents the characteristic dimension of the flow field, usually the width or height of the flow field, p g represents the dynamic viscosity of the gas phase fluid, represents the viscous property of the gas phase fluid, TRe represents the turbulent Reynolds number, represents the ratio of fluid turbulent kinetic energy to friction work, and is used to evaluate the intensity of turbulence, k g represents the turbulent kinetic energy of the gas phase fluid, represents the energy of turbulent motion in the gas phase fluid, V represents the volume of the fluid, represents the total volume of the fluid in the flow field, represents the average velocity square, represents the mean square value of the fluid velocity, represents the average velocity square, represents the square value of the average velocity of the fluid, W represents the friction work, represents the work done by friction in the fluid, p represents the dynamic viscosity of the fluid, represents the viscous property of the fluid, represents the velocity gradient, represents the rate of change of the fluid velocity in a certain direction, n v represents the normal vector, represents the unit vector in a certain direction in the flow field, <u i ′u jdenotes the Reynolds stress tensor, and denotes the correlation of the velocity fluctuations in the fluid.

[0030] As a further improvement of the application, the equation for evaluating the trained artificial neural network model using the validation set comprises:

[0031] The flow field characteristic parameters are matched with the target parameters, a data set is constructed, and the data set is cleaned and the like; the data set is divided into 80% training set and 20% validation set;

[0032] The grid search method is used to adjust the artificial neural network model, and the artificial neural network model with the minimum loss function is selected for training; the artificial neural network model is trained according to the training set, the network parameters are adjusted through the back propagation algorithm, and the early stopping strategy is adopted to avoid overfitting;

[0033] The trained artificial neural network is evaluated using the validation set, the error index of the artificial neural network model on the validation set is calculated; a threshold is set, if the threshold is exceeded, the artificial neural network model is adjusted again; if the threshold is lower, feature extraction and analysis are performed.

[0034] As a further improvement of the application, the process of constructing the data set comprises:

[0035] The missing values in the data set are checked; if missing values exist, the mean filling is used for processing as follows:

[0036] j'=1

[0037]

[0038] In the formula, x i′ is the value at the position of the missing value, n' represents the total number of non-missing values in the data set, x i′j′ represents the j'th feature value of the i'th sample or data point;

[0039] Box plots of each mechanism parameter and flow characteristic value are drawn; the upper quartile, lower quartile and quartile range of each parameter are calculated, and the upper and lower threshold values are determined; values below the lower limit or above the upper limit are marked as abnormal values; abnormal values are analyzed by replacing or deleting and the like;

[0040] The extracted structure parameters and flow field characteristics are standardized; the standardization formula is as follows:

[0041]

[0042] In the formula, x norm is the standardized data, x is the original data, μ is the mean of the data, and σ is the standard deviation of the data.

[0043] As a further improvement of the application, the process of training the artificial neural network model with the training set comprises:

[0044] The network structure of the artificial neural network model is designed, ReLU is selected as the activation function, and Huner loss is selected as the loss function; the training set is input into the artificial neural network model, and the Adam optimizer is used to adjust the weights and biases of the artificial neural network model;

[0045] The network structure of the artificial neural network model includes an input layer (containing 7 input features: blade length, blade rotation angle, blade installation height, blade tip clearance, blade thickness, CH4 pipe diameter, and mixing zone length), several hidden layers, and an output layer (containing 3 output features: SN, Re, and TRe).

[0046] The value range of the learning rate, the number of hidden layers, the number of nodes, the batch number, and the batch size are determined; all the hyperparameters form a grid; for each combination of hyperparameters in the grid, the training set is used to train the artificial neural network model.

[0047] The mean square error and accuracy of the artificial neural network model after each training are verified using the validation set, a threshold is set, and the artificial neural network model corresponding to the hyperparameter combination with the smallest error compared to the threshold is selected as the final artificial neural network model.

[0048] As a further improvement of the application, the process of forming the optimal fitness parameter combination comprises:

[0049] The SHAP importance evaluation method is used to analyze the influence of each structural parameter in the artificial neural network model on the target parameter; through the analysis of the importance of the target feature, the key structural parameter with the greatest influence on the target parameter is determined, the parameter combination is set, and the combination is input into the trained artificial neural network model to predict the corresponding target parameter.

[0050] The difference between the target parameter predicted by the artificial neural network model and the actual parameter is adjusted, the initial structural parameter population is randomly generated using the genetic algorithm; the optimization is continuously iterated until the maximum iteration number or the function converges, and finally the optimal fitness parameter combination is obtained.

[0051] The optimal fitness parameter combination is selected for CFD-DEM simulation verification, and the performance after verification is tested and corrected.

[0052] As a further improvement of the application, the process of predicting the corresponding target parameter comprises:

[0053] A collection of training artificial neural network data sets, data includes input features and corresponding target parameters; after the model training is completed, the SHAP method is used to calculate the contribution value of each input feature to the target parameter; set the contribution value standard, compare and select the features with contribution value higher than the standard;

[0054] According to the contribution value obtained by comparison, the importance is sorted; the higher the contribution value, the greater the influence of the feature on the target parameter, and the higher the relative importance; according to the sorting result, the key feature ranked first is selected;

[0055] The selected key features are combined and input into the trained artificial neural model to predict the target parameter; by comparing the actual parameter with the predicted parameter, if it is higher than a certain threshold, retraining and prediction are performed; repeat the above steps until the maximum iteration number or reach the preset standard.

[0056] As a further improvement of the present application, the process of randomly generating an initial structure parameter population using a genetic algorithm includes:

[0057] Through the analysis of feature importance, the key structure parameters that have the greatest influence on the target parameters are determined, and the parameter combination is set, and the parameters are input into the trained artificial neural network model to predict the corresponding target parameters;

[0058] The gap between the target parameters predicted by the artificial neural model and the actual target parameters is adjusted to minimize the gap;

[0059] Randomly generate an initial structure parameter population, and predict the SN, Re and TRe of each parameter combination through an artificial intelligence network model; calculate the fitness function of each combination, that is, the gap between the predicted values of SN, Re and TRe and the target SN, Re and TRe; according to the fitness function, select individuals with better performance for crossover and mutation to generate a new parameter population;

[0060] Repeat the selection, crossover and mutation process, and continuously iterate and optimize until the maximum iteration number is reached or the fitness function converges; finally, the parameter combination with the optimal fitness is verified by CFD-DEM simulation; set a verification threshold, if higher than the threshold, continue to select individuals according to the fitness function for crossover and mutation to generate a parameter combination with optimal fitness.

[0061] To achieve the above purpose, the present application also provides the following technical solutions:

[0062] A smelting cyclone dry powder spray gun design system based on machine learning, the molten pool smelting cyclone chaotic dry powder spray gun structure parameter optimization method includes:

[0063] The flow field simulation and data acquisition module is configured to simulate the cyclone chaotic dry powder spray gun under different architecture parameter configurations using CFD-DEM coupling software, and the grid at the boundary of the cyclone blade and the outlet of the spray gun is encrypted before calculation; the characteristic parameters of the flow field are extracted in the CFD-DEM coupling software; the ratio of the tangential momentum flux and the axial momentum flux of the fluid and the turbulent intensity of the jet in the spray gun are calculated based on the density, coordinates, tangential velocity and axial velocity of each data point, and the calculated data is transmitted to the data cleaning and model training module;

[0064] The data cleaning and model training module is configured to match the flow heat parameters with the calculated data to build a data set; check the missing values in the data set, describe the distribution of the data and identify abnormal values using the box plot method of data distribution; and standardize the extracted structure parameters and flow field characteristics; output the processed training set to the artificial neural network model, adjust the model using the grid search method; adjust the network parameters through the back propagation algorithm to reduce the loss function value; and evaluate the trained artificial neural network model using the validation set;

[0065] The model training and verification module is configured to determine the key structure parameters that have the greatest impact on the target parameters through feature importance analysis, and set parameter combinations; input the set parameter combinations into the trained artificial neural network model to predict the corresponding target parameters; compare the target parameters predicted by the artificial neural network model with the actual target parameters to minimize the difference between the structure parameters; generate an initial structure parameter population randomly using a genetic algorithm, and predict the SN, Re and TRe of each parameter combination through the artificial intelligence network model; calculate the fitness function of each combination, i.e., the difference between the predicted values of SN, Re and TRe and the target SN, Re and TRe; set a fitness function threshold, select the individual with the smallest difference from the threshold for crossover and mutation to form a new parameter population; repeat the selection, crossover and mutation process for continuous iteration and optimization until the maximum iteration number or the fitness function converges; perform CFD-DM simulation verification on the most optimal parameter combination with the best fitness; set a verification threshold, and if the threshold is higher, continue to select individuals for crossover and mutation according to the fitness function to generate the most optimal parameter combination.

[0066] The present application can quickly predict the swirl number (SN), turbulent intensity Reynolds number (Re) and turbulent Reynolds number (TRe) of different structure parameter combinations by combining the computational fluid dynamics (CFD) and discrete element method (DEM) coupling technology and artificial neural network (ANN) technology, effectively shortening the design cycle; then the genetic algorithm (GA) is used for iterative optimization of the parameter combination, the target parameter value predicted by the ANN model is combined, the individuals with excellent performance are selected based on the fitness function for crossover and mutation, and the difference between the predicted value and the target value is gradually reduced; not only the optimization efficiency is improved, but also the accuracy and reliability of the results are improved; finally, the optimized parameter combination is verified by CFD-DEM simulation, which proves the stability and efficiency and ensures that the expected performance indicators can be reached. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 A step flow schematic diagram of an embodiment of the smelting cyclone dry powder lance design method based on machine learning of the present application;

[0068] Figure 2 A step flow schematic diagram of forming target parameters of an embodiment of the smelting cyclone dry powder lance design method based on machine learning of the present application;

[0069] Figure 3 A step flow schematic diagram of calculating the ratio of tangential momentum flux to axial momentum flux SN and the turbulent intensity of the jet in the lance of an embodiment of the smelting cyclone dry powder lance design method based on machine learning of the present application;

[0070] Figure 4 A step flow schematic diagram of evaluating the trained artificial neural network model using the validation set of an embodiment of the smelting cyclone dry powder lance design method based on machine learning of the present application;

[0071] Figure 5 A principle diagram of evaluating the trained artificial neural network model using the validation set of an embodiment of the smelting cyclone dry powder lance design method based on machine learning of the present application;

[0072] Figure 6 A step flow schematic diagram of constructing a data set of an embodiment of the smelting cyclone dry powder lance design method based on machine learning of the present application;

[0073] Figure 7 A step flow schematic diagram of training an artificial neural network model using a training set of an embodiment of the smelting cyclone dry powder lance design method based on machine learning of the present application;

[0074] Figure 8 A step flow schematic diagram of forming a fitness optimal parameter combination of an embodiment of the smelting cyclone dry powder lance design method based on machine learning of the present application;

[0075] Figure 9 A schematic diagram of a step of forming a principle diagram of an optimal parameter combination for an embodiment of the design method of a smelting cyclone dry powder spray gun based on machine learning of the application;

[0076] Figure 10 A schematic diagram of a step of predicting a corresponding target parameter for an embodiment of the design method of a smelting cyclone dry powder spray gun based on machine learning of the application;

[0077] Figure 11 A schematic diagram of a step of randomly generating an initial structure parameter population by using a genetic algorithm for an embodiment of the design method of a smelting cyclone dry powder spray gun based on machine learning of the application;

[0078] Figure 12 A schematic diagram of a step of CFD-DEM simulation verification of an optimal parameter for an embodiment of the design method of a smelting cyclone dry powder spray gun based on machine learning of the application;

[0079] Figure 13 A schematic diagram of a step of running CFD-DEM simulation for an embodiment of the design method of a smelting cyclone dry powder spray gun based on machine learning of the application;

[0080] Figure 14 A schematic diagram of a functional module for an embodiment of the design system of a smelting cyclone dry powder spray gun based on machine learning of the application;

[0081] Figure 15 A schematic diagram of a structure for an embodiment of the electronic device of the application;

[0082] Figure 16 A schematic diagram of a structure for an embodiment of the storage medium of the application. DETAILED DESCRIPTION

[0083] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0084] The terms "first", "second", "third", etc. are used only for descriptive purposes and do not connote or imply relative importance or a number of the indicated technical features. Thus, features defined with "first", "second", "third" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, such as two, three, etc., unless otherwise explicitly and specifically limited. All directional references, such as up, down, left, right, front, back, etc., used in the description of the present application are only used for the convenience of description and are relative to the specific orientation of the drawings shown in the application. If the specific orientation changes, the directional references will also change accordingly. In addition, the terms "comprise" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0085] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearance of the phrase in various places in the specification does not necessarily all refer to the same embodiment, nor is it necessarily mutually exclusive of other embodiments. It is expressly understood that the embodiments described herein can be combined with other embodiments.

[0086] As shown in Figure 1 The present embodiment provides an embodiment of a machine learning-based smelting cyclone dry powder lance design method, which specifically includes the following steps in the present embodiment:

[0087] Step S1: Simulate the cyclone chaotic dry powder lance under different structural parameter configurations using CFD-DEM coupling software; extract the key feature parameters of the flow field in the CFD-DEM coupling software, calculate the ratio of tangential momentum to axial momentum of the fluid through the key feature parameters, and form the target parameters;

[0088] Step S2: Match the flow field feature parameters with the target parameters, build a data set, and clean the data set; divide the data set into a training set and a validation set, train an artificial neural network model according to the training set, and evaluate the trained artificial neural network model using the validation set;

[0089] Step S3: analyze the influence of each structural parameter in the artificial neural network model on the target parameter, determine the structural parameter that has the greatest influence on the target parameter, and set the parameter combination; calculate the target parameter of the parameter combination, and select individuals for crossover and mutation according to the calculation result to form a parameter combination with the optimal fitness;

[0090] The calculation reaches the maximum number of iterations or the fitness function converges, if the new model combination is simulated and verified, if not, continue to cross and mutate.

[0091] Preferably, steps S1-S3 of the embodiment can quickly predict the swirl number (SN), turbulent intensity Reynolds number (Re), and turbulent Reynolds number (TRe) of different structural parameter combinations by combining computational fluid dynamics (CFD) and discrete element method (DEM) coupling technology and artificial neural network (ANN) technology, effectively shortening the design cycle; then, the parameter combination is iteratively optimized by genetic algorithm (GA), and the target parameter value predicted by the ANN model is combined to select individuals with excellent performance for crossover and mutation based on the fitness function, gradually reducing the difference between the predicted value and the target value; not only improves the optimization efficiency, but also improves the accuracy and reliability of the results; finally, the optimized parameter combination is simulated and verified by CFD-DEM, proving its stability and efficiency, ensuring that the expected performance indicators can be achieved.

[0092] Therefore, the embodiment significantly improves the design and optimization efficiency of the spray gun, reduces the test cost, and effectively optimizes the performance and durability of the swirl chaotic dry powder spray gun, realizes an efficient and stable smelting process, and has important application value for the smelting industry.

[0093] Further, as shown in Figure 2 , the process of forming the target parameter in step S1 specifically includes the following steps:

[0094] Step S11: locally encrypt the swirl chaotic dry powder spray gun, especially the key positions such as the boundary of the swirl vane and the outlet of the spray length, and use CFD-DEM coupling software to simulate and calculate the swirl chaotic dry powder spray gun under different structural parameter configurations;

[0095] Step S12: extract the key characteristic parameters of the flow field in the CFD-DEM software;

[0096] The key characteristic parameters include the density, coordinates (distance r between the calculation rotation point and the velocity vector), tangential velocity U tan , axial velocity U axi , inner diameter R of the annular pipe of the spray gun, range from R0 to R1 of the annular zone to the mixing zone, characteristic length L of the spray gun, velocity u of the jet, integral volume V, etc.

[0097] Step S13: Calculate the ratio of tangential momentum flux to axial momentum flux SN of the fluid and the turbulent intensity of the jet in the lance by the above parameters.

[0098] Preferably, step S11 of the present embodiment can accurately capture the complex flow behavior of the fluid in these regions by performing local encryption at key positions, thereby improving the accuracy of the simulation. Using CFD-DEM coupling software can simulate the interaction between fluid and particles, which is crucial for understanding the working mechanism of the cyclone chaotic dry powder lance. Step S12 extracts key feature parameters such as density, coordinates, tangential velocity, axial velocity, etc., which can deeply analyze the dynamic characteristics of the flow field and provide basic data for momentum flux calculation and turbulent intensity analysis. Step S13 calculates the ratio of tangential momentum flux to axial momentum flux SN to evaluate the influence of cyclone intensity on fluid motion; while the calculation of turbulent intensity understands the instability of the fluid, which is of great significance for optimizing the design of the lance and improving the efficiency of the spray.

[0099] Therefore, the significance of the present embodiment lies in that by combining CFD-DEM coupling simulation and extraction of key feature parameters, the design of the cyclone chaotic dry powder lance can be effectively analyzed and optimized, thereby improving its performance and efficiency

[0100] Further, as shown in Figure 3 , the process of calculating the ratio of tangential momentum flux to axial momentum flux SN of the fluid and the turbulent intensity of the jet in the lance in step S13 specifically includes the following steps:

[0101] Step S131: Extract the key feature parameters of the flow field in the CFD-DEM software;

[0102] Step S132: Calculate the ratio of tangential momentum flux to axial momentum flux SN of the fluid; the equation of SN can be expressed as:

[0103]

[0104]

[0105] In the formula, SN local represents the ratio of tangential momentum flux to axial momentum flux, and G angular,local represents the momentum distribution characteristics of the local region. G linear,local represents the tangential momentum flux of the local region, represents the momentum transfer of the fluid in the tangential direction, G axidenotes the axial velocity, denotes the velocity component of the fluid in the axial direction, U tan denotes the tangential velocity, denotes the velocity component of the fluid in the tangential direction, R0denotes the lower limit of integration, typically denotes the inner boundary radius of the flow field, R1denotes the upper limit of integration, typically denotes the outer boundary radius of the flow field;

[0106] Step S133: Introducing the turbulent intensity Reynolds number (Re) and the turbulent Reynolds number (TRe) of the gas-phase jet inside the lance for evaluating the turbulent intensity of the jet inside the lance;

[0107] where Re is defined as the ratio of inertial forces to frictional forces, and TRe is defined as the ratio of the turbulent kinetic energy (k g ) of the fluid to the work done by friction (W), which can be expressed as:

[0108]

[0109] where Re denotes the Reynolds number, denotes the ratio of inertial forces to viscous forces, used to evaluate the turbulent intensity of the fluid, p g denotes the density of the gas-phase fluid, denotes the mass per unit volume of the gas-phase fluid, u g denotes the characteristic velocity of the gas-phase fluid, denotes the velocity of the gas-phase fluid over a characteristic length, L denotes the characteristic length, denotes the characteristic dimension of the flow field, typically the width or height of the flow field, m g denotes the dynamic viscosity of the gas-phase fluid, denotes the viscous properties of the gas-phase fluid, TRe denotes the turbulent Reynolds number, denotes the ratio of the turbulent kinetic energy to the work done by friction, used to evaluate the intensity of the turbulence, k g denotes the turbulent kinetic energy of the gas-phase fluid, denotes the energy of turbulent motion in the gas-phase fluid, V denotes the volume of the fluid, denotes the total volume of the fluid in the flow field, denotes the mean velocity squared, denotes the mean square value of the fluid velocity, denotes the mean velocity squared, denotes the square value of the mean velocity of the fluid, W denotes the work done by friction, denotes the work done by friction in the fluid, m denotes the dynamic viscosity of the fluid, denotes the viscous properties of the fluid, denotes the velocity gradient, denotes the rate of change of the fluid velocity in a certain direction, n v denotes the normal vector, denotes the unit vector in a certain direction in the flow field, <u i ′u j ′> denotes the Reynolds stress tensor, denotes the correlation of the velocity fluctuations in the fluid.

[0110] Preferably, the key characteristic parameters of the flow field are extracted in the CFD-DEM software in step S131. The combination of computational fluid dynamics (CFD) and discrete element method (DEM) can more accurately simulate and analyze the behavior of fluid in complex geometric structures; the micro and macro characteristics of the fluid can be captured. In step S132, the ratio SN of the tangential momentum flux of the fluid to the axial momentum flux is calculated. By calculating the momentum flux of the fluid in different directions, the flow characteristics and vortex structure of the fluid can be better understood; it is helpful to evaluate the turbulence intensity and flow stability of the fluid, which is of great significance for optimizing the design of fluid system. In step S133, the turbulent Reynolds number (Re) and turbulent Reynolds number (TRe) of the gas-phase jet in the lance are introduced to evaluate the turbulent intensity of the jet in the lance. By introducing the turbulence model, the turbulent characteristics of the jet in the lance can be accurately predicted and analyzed. The selection and parameter setting of the turbulence model are crucial for the accuracy of the simulation results. Different models show different performance under different conditions.

[0111] Therefore, the significance of the embodiment is that by combining CFD and DEM technology, the behavior of fluid in complex geometric structures can be more comprehensively and accurately analyzed and predicted, thereby providing a scientific basis for engineering design and optimization.

[0112] Further, as shown in Figure 4 The process of verifying the trained artificial neural network model in step S2 includes the following steps:

[0113] Step S21: match the flow field characteristic parameters with the target parameters, build a data set, and perform cleaning and other operations on the data set; divide the data set into 80% training set and 20% verification set;

[0114] Step S22: adjust the artificial neural network model using the grid search method, select the artificial neural network model with the smallest loss function for training; train the artificial neural network model according to the training set, adjust the network parameters through the back propagation algorithm, and use the early stopping strategy to avoid overfitting;

[0115] Step S23: use the verification set to evaluate the trained artificial neural network, calculate the error index of the artificial neural network model on the verification set; set a threshold, if the threshold is exceeded, adjust the artificial neural network model again; if the threshold is lower, perform feature extraction and analysis.

[0116] Preferably, in step S21 of the embodiment, the data set is constructed by matching the flow field characteristic parameters with the target parameters, ensuring the accuracy and consistency of the data. The data cleaning operation can remove noise and outliers, improve data quality, and thus improve the training effect of the model. In step S22, the grid search method is used to adjust the artificial neural network model, which can systematically explore different super parameter combinations, select the model with the smallest loss function for training, and effectively improve the performance of the model. The network parameters are adjusted by the back propagation algorithm, which can optimize the weights and biases of the model to better fit the training data. The early stopping strategy is adopted to avoid overfitting, that is, the training is stopped when the error on the validation set no longer decreases, preventing the model from overfitting on the training set, thereby improving the generalization ability of the model. In step S23, the trained artificial neural network is evaluated using the validation set, and the error index is calculated, which can objectively evaluate the performance of the model and ensure the performance of the model on unseen data. A threshold is set, if the threshold is exceeded, the artificial neural network model is adjusted, if the threshold is below, feature extraction and analysis are performed, and through the dynamic adjustment mechanism, the adaptability and stability of the model on different data sets can be ensured. (For specific principles, refer to the attached Figure 5 ).

[0117] Therefore, the significance of the embodiment lies in ensuring the accuracy and generalization ability of the artificial neural network model in flow field feature prediction through data preprocessing, model optimization, verification and adjustment, etc.

[0118] Further, as shown in Figure 6 , the process of constructing the data set in step S21 specifically includes the following steps:

[0119] Step S211: check the missing values in the data set; if missing, use mean filling for processing as follows:

[0120] j' = 1

[0121]

[0122] In the formula, x i′ is the value at the position of the missing value, n' represents the total number of non-missing values in the data set, and x i′j′ represents the j'th feature value of the i'th sample or data point;

[0123] Step S212: draw a box plot of each mechanism parameter and flow field characteristic value; calculate the upper quartile, lower quartile and interquartile range of each parameter to determine the upper and lower threshold values; values below the lower limit or above the upper limit are marked as outliers; analyze the outliers by replacing or deleting, etc.

[0124] Step S213: standardize the extracted structure parameters and flow field characteristics; the standardization formula is as follows:

[0125]

[0126] where x norm is the standardized data, x is the original data, μ is the mean of the data, and σ is the standard deviation of the data.

[0127] Preferably, the missing values in the data set are processed in step S121 of the embodiment, ensuring the integrity of the data and the accuracy of subsequent analysis. The presence of missing values can affect the accuracy of statistical estimates and the effectiveness of subsequent work. Mean filling is a simple and commonly used method for handling missing values, suitable for cases with low missing rates. Through mean filling, information loss caused by deleting missing tuples can be avoided, while maintaining the size of the data set unchanged. In step S122, outliers in the data are identified and processed, ensuring the quality of the data and the accuracy of the analysis. Box plot is an effective visualization tool that can intuitively show the distribution of data and outliers. By calculating the quartiles and interquartile range, the upper and lower threshold values of the data can be determined, so that the outliers deviating from the normal range can be identified. The processing method of outliers includes replacement or deletion, depending on the nature of the outliers and the impact on the overall data. In step S123, data of different dimensions are converted to the same dimension, which facilitates subsequent comparison and analysis. Standardization can eliminate the influence of different data index dimensions, so that the data can be compared on the same scale, thereby improving the accuracy of data analysis. The standardized data is more suitable for training and evaluation of machine learning models, as many models are sensitive to the scale of data.

[0128] Therefore, the significance of the embodiment lies in playing a key role in the data preprocessing process, ensuring the quality of the data and the effectiveness of the analysis.

[0129] Further, as shown in Figure 7 , the process of training the artificial neural network model in step S22 specifically includes the following steps:

[0130] Step S221: design the network structure of the artificial neural network model, select the activation function and the loss function; input the training set into the artificial neural network model, and use the Adam optimizer to adjust the weights and biases of the artificial neural network model;

[0131] where the network structure of the artificial neural network model includes an input layer (containing 7 input features: blade length, blade rotation angle, blade installation height, blade tip clearance, blade thickness, CH4 pipe diameter, and mixing zone length), several hidden layers, and an output layer (containing 3 output features: SN, Re, and TRe);

[0132] The expression of the activation function is:

[0133]

[0134] where x is the input value, and a is the adaptive leakage coefficient, which is dynamically adjusted according to the gradient during the training process, and the initial value is set to 0.01;

[0135] The expression of the loss function is:

[0136] Weighted Huber Loss multi-scale =∑i=1 n

[0137]

[0138] where n represents the number of samples, y i represents the actual value of the i-th sample, represents the predicted value of the i-th sample, δ represents the threshold value of the Huber loss, which is set to 1, K represents the number of scales of multi-scale regularization, and λ k represents the regularization coefficient of the k-th scale, which controls the influence degree of the regularization term, and w j represents the j-th weight in the model, and m represents the total number of weights in the model;

[0139] Step S222: Determine the value range of hyperparameters such as learning rate, number of hidden layers, number of nodes, batch size, and batch size; form a grid with all hyperparameters; for each combination of hyperparameters in the grid, use the training set to train the artificial neural network model;

[0140] Step S223: Use the validation set to verify the artificial neural network model after each training in terms of mean square error, accuracy, etc., set a threshold, and select the artificial neural network model corresponding to the hyperparameter combination with the smallest error compared to the threshold as the final artificial neural network model.

[0141] Preferably, in step S221 of the embodiment, the network structure of the artificial neural network model is designed, the appropriate number of hidden layers and activation function (ReLU) are selected, and the Huber loss is used as the loss function. This ensures that the network can effectively learn and represent complex nonlinear relationships, and the weights and biases of the network are adjusted by the Adam optimizer to optimize the performance of the model. In step S222, the hyperparameters such as learning rate, number of hidden layers, number of nodes, batch size, and batch size are determined and optimized. By forming a grid of hyperparameters and training each combination of hyperparameters, the optimal combination of hyperparameters can be systematically explored and selected, thereby improving the generalization ability and training efficiency of the model. In step S223, the trained model is verified using the validation set, and the artificial neural network model corresponding to the hyperparameter combination with the smallest error compared to the threshold is selected as the final model by setting a threshold. This step ensures the performance of the model on unseen data, thereby improving the reliability and practicality of the model.

[0142] Therefore, the significance of the present embodiment lies in ensuring that the artificial neural network model can efficiently and accurately learn and predict through a systematic design, optimization and verification process, thereby performing well in practical applications.

[0143] Further, as shown in Figure 8 , the process of forming the optimal fitness parameter combination in step S3 specifically includes the following steps:

[0144] Step S31: using the SHAP importance evaluation method, analyze the influence of each structural parameter in the artificial neural network model on the target parameter; through the analysis of the importance of the target features, determine the key structural parameters that have the greatest influence on the target parameters, set the parameter combination, and input the combination into the trained artificial neural network model to predict the corresponding target parameters;

[0145] Step S32: adjust the parameter combination by the gap between the target parameters predicted by the artificial neural network model and the actual parameters; randomly generate an initial structural parameter population using a genetic algorithm; continuously iterate and optimize until the maximum number of iterations or function convergence is reached;

[0146] Step S33: select the optimal fitness parameters for CFD-DEM simulation verification, and test and correct the performance after verification.

[0147] Preferably, the step S31 of the present embodiment identifies the key structural parameters that have the greatest influence on the target parameters through SHAP importance analysis, which helps to simplify the model input, improve the interpretability and prediction efficiency of the model. By setting the key parameter combination and inputting it into the trained artificial neural network model, the target parameters can be more accurately predicted, thereby optimizing the performance of the model. In step S32, an initial structural parameter population is randomly generated using a genetic algorithm, and continuously iterated and optimized until the maximum number of iterations or function convergence is reached. This can effectively explore the parameter space and find the optimal or near-optimal parameter combination, thereby improving the prediction accuracy and generalization ability of the model. In step S33, CFD-DEM simulation verification and performance test correction are performed to ensure the effectiveness of the optimized parameter combination in practical applications. The present embodiment combines theoretical analysis and experimental verification to ensure the reliability and practicality of the model. (For specific principles, refer to the attached Figure 9 )

[0148] Therefore, the present embodiment has important technical significance in improving the prediction accuracy of artificial neural network models, optimizing parameter selection, enhancing the robustness and adaptability of models, and promoting technological progress and application expansion.

[0149] Further, as shown in Figure 10 , the process of predicting the corresponding target parameters in step S31 specifically includes the following steps:

[0150] Step S311: Collecting training artificial neural network data set, the data including input features and corresponding target parameters; after the model training is completed, using SHAP method to calculate the contribution value of each input feature to the target parameter; setting the contribution value standard, comparing, selecting the features with contribution value higher than the standard;

[0151] Step S312: According to the contribution value obtained by comparison, the importance is sorted; the higher the contribution value, the greater the influence of the feature on the target parameter, and the higher the relative importance; according to the sorting result, the key feature ranked first is selected;

[0152] Step S313: The selected key features are combined and input into the trained artificial neural model to predict the target parameter; by comparing the actual parameter with the predicted parameter, if it is higher than a certain threshold, retraining and prediction are performed; repeat the above steps until the maximum iteration number or the preset standard is reached.

[0153] Preferably, step S311 of the embodiment provides basic data for model training by collecting training data set including input features and corresponding target parameters. Using SHAP method to calculate the contribution value of each input feature to the target parameter can quantify the influence degree of each feature on the model prediction result. By setting the contribution value standard and comparing, selecting the features with contribution value higher than the standard helps to filter out the key features that have greater influence on the target parameter, thereby improving the prediction accuracy and efficiency of the model. Step S312 sorts the features according to the importance of the contribution value, the higher the contribution value, the greater the influence of the feature on the target parameter, and the higher the relative importance. Selecting the key feature ranked first can focus on the most important feature for analysis and prediction. Through importance sorting, the feature with the greatest influence on the target parameter can be identified, and the decision-making process of the model can be understood to provide a basis for feature engineering and model optimization. Step S313 combines the selected key features and inputs them into the trained artificial neural model to predict the target parameter, and compares the actual parameter with the predicted parameter. If it is higher than a certain threshold, retraining and prediction are performed. Repeat the above steps until the maximum iteration number or the preset standard is reached. Through iterative training and prediction, the model can be continuously optimized to improve the accuracy and stability of prediction; at the same time, by setting the threshold and the maximum iteration number, the complexity and time cost of model training can be controlled.

[0154] Therefore, the embodiment provides basic data for model training, and quantifies the importance of features through SHAP method to filter out key features; importance sorting identifies the feature with the greatest influence on the target parameter to provide a basis for subsequent analysis and prediction; and through iterative training and prediction, the model is continuously optimized to improve the accuracy and stability of prediction, while controlling the complexity and time cost of model training.

[0155] Further, as shown in Figure 11 The process of randomly generating the initial structure parameter population in step S32 by using the genetic algorithm specifically includes the following steps:

[0156] Step S321: Determine the key structure parameters that have the greatest impact on the target parameters through analysis of feature importance, and set parameter combinations. Input the parameters into the trained artificial neural network model to predict the corresponding target parameters.

[0157] Step S322: Adjust the structure parameters to minimize the gap between the target parameters predicted by the artificial neural network and the actual target parameters.

[0158] Randomly generate an initial structure parameter population, and predict the SN, Re, and TRe of each parameter combination through an artificial intelligence network model. Calculate the fitness function of each combination, which is the gap between the predicted values of SN, Re, and TRe and the target SN, Re, and TRe. Select individuals with better performance for crossover and mutation according to the fitness function to generate a new parameter population.

[0159] Step S323: Repeat the process of selection, crossover, and mutation, and continuously iterate and optimize until the maximum number of iterations is reached or the fitness function converges. Finally, the parameter combination with the optimal fitness is verified by CFD-DEM simulation. Set a verification threshold, if higher than the threshold, continue to select individuals for crossover and mutation according to the fitness function to generate the parameter combination with the optimal fitness.

[0160] Preferably, in step S321 of this embodiment, the key structural parameters that have the greatest impact on the target parameters are determined through feature importance analysis, and parameter combinations are set. These parameters are then input into the trained artificial neural network model to predict the corresponding target parameters. The residual learning framework in deep learning is used to optimize the network by learning the residual function, making the network easier to optimize and improving accuracy. Furthermore, this demonstrates the powerful ability of neural networks to handle complex problems, such as in image recognition tasks. Step S322 uses the artificial neural model to predict the difference between the output target parameters and the actual target parameters, and adjusts the structural parameters to minimize the difference. The Adam optimization algorithm is utilized, which performs well when handling large-scale data and parameters, and has good convergence and computational efficiency. In addition, the application of genetic algorithms in optimization problems, such as real-valued genetic algorithms in PID parameter optimization, uses real-number encoding and genetic operations to find the optimal solution. Step S323 involves randomly generating an initial population of structural parameters and predicting the SN, Re, and TRe for each parameter combination using an artificial intelligence network model. The fitness function for each combination is calculated, which is the difference between the predicted SN, Re, and TRe values ​​and the target SN, Re, and TRe. Based on the fitness function, individuals with better performance are selected for crossover and mutation to generate a new parameter population. By combining differential evolution and genetic algorithms, parameter combinations are optimized through adaptive parameter adjustment and selection of the best individuals, thereby improving prediction accuracy.

[0161] Therefore, the significance of this embodiment lies in the fact that by combining deep learning, optimization algorithms and genetic algorithms, it achieves efficient optimization and prediction of parameters of complex systems, demonstrating the wide application and technological advantages of artificial intelligence in engineering and scientific computing.

[0162] Furthermore, such as Figure 12 As shown, the process of verifying the optimal parameters using CFD-DEM simulation in step S33 specifically includes the following steps:

[0163] Step S331: Set the initial flow field and particle distribution. Based on the simulation results of the artificial neural network model, run the CFD-DEM simulation and record key variables. Key variables include flow velocity, pressure, particle concentration, collision frequency, etc.

[0164] Step S332: Calculate the target parameters based on the simulation results. The target parameters include SN, Re, TRe, etc. Set an optimization target value and compare the simulation results with the optimization target value. If the simulation results exceed the optimization target value, re-mesh the grid. Repeat the above steps until the simulation results are less than the optimization target value.

[0165] Step S333: Evaluate the performance of the optimized parameter combination according to the simulation results, set performance indicators, and if not up to standard, continue to adjust the simulation; the performance of the optimized parameter combination includes but is not limited to the service life of the spray gun, manufacturing cost and other constraint conditions.

[0166] Preferably, step S331 of the present embodiment provides basic data for subsequent CFD-DEM simulation by setting initial flow field and particle distribution. According to the simulation results of the artificial neural network model, the accuracy and reliability of the simulation results are ensured. By recording key variables such as flow rate, pressure, particle concentration, collision frequency, etc., important data support can be provided for subsequent analysis and optimization. Step S332 calculates the target parameters and optimizes them by analyzing the simulation results. Compare the simulation results with the optimization target value, if the simulation results exceed the optimization target value, then re-divide the grid and repeat the above steps until the simulation results are less than the optimization target value. The accuracy and optimization effect of the simulation results are ensured, and a reliable basis is provided for subsequent performance evaluation. Step S333 evaluates the performance of the optimized parameter combination according to the simulation results, and sets performance indicators. If not up to standard, continue to adjust the simulation. The performance of the optimized parameter combination includes but is not limited to the service life of the spray gun, manufacturing cost and other constraint conditions; ensure that the performance of the optimized parameter combination meets the expectations, and provide technical support for practical application.

[0167] Therefore, the significance of the present embodiment lies in setting initial conditions, analyzing simulation results, and evaluating optimized parameter combinations to ensure the accuracy of simulation results and optimization effects, and to provide reliable technical support for practical applications.

[0168] Further, as shown in Figure 13 the process of running CFD-DEM simulation in step S331 specifically includes the following steps:

[0169] Step S3311: Input the optimal parameter combination in the CFD software, simulate the flow field according to the optimal parameter combination, and optimize the flow field characteristics;

[0170] Step S3312: Numerically simulate the motion of the cyclone chaotic dry powder spray gun during the flow field running process by coupling CFD with DEM, obtain the characteristics and performance of the cyclone chaotic dry powder spray gun, and record the key variables;

[0171] Among them, the key variables include flow rate, pressure, particle concentration, collision frequency, etc.

[0172] Step S3313: Calculate the target parameters according to the characteristics, performance and key variables of the cyclone chaotic dry powder spray gun, set the optimization target value, compare the simulation results with the optimization target value, and if the simulation results exceed the optimization target value, re-divide the grid.

[0173] Preferably, step S3311 of the embodiment can ensure that the initial conditions of the flow field simulation are optimal by inputting the optimal parameter combination, thereby improving the accuracy and reliability of the simulation results. Optimizing the flow field characteristics makes the flow of fluid in a certain area more stable and efficient. Step S3312 can consider the motion of particles and fluid simultaneously through the DEM-CFD coupling method, accurately simulate the complex behavior of the cyclone chaotic dry powder spray gun in the flow field. By recording key variables, the performance and characteristics of the spray gun can be comprehensively understood, providing data support for optimization. Step S3313 can evaluate the accuracy of the simulation results by calculating the target parameters and comparing them with the optimization target values, and adjust the grid division as needed to improve the simulation accuracy.

[0174] Therefore, the experimental cost and time can be reduced, and the design efficiency can be improved, especially in cases where multiple parameter adjustments are needed to achieve optimal performance; the embodiment comprehensively analyzes the performance and characteristics of the cyclone chaotic dry powder spray gun through the DEM-CFD coupling method; the simulation-based optimization method can significantly reduce the number of experiments, save resources, and improve the accuracy and efficiency of the design.

[0175] As shown in Figure 4 The embodiment also provides an embodiment of a molten pool smelting cyclone chaotic dry powder spray gun structure parameter optimization system, in which the molten pool smelting cyclone chaotic dry powder spray gun structure parameter optimization system is applied to the machine learning-based smelting cyclone dry powder spray gun design method in the above embodiment, and the molten pool smelting cyclone chaotic dry powder spray gun structure parameter optimization system comprises a flow field simulation and data acquisition module 1, a data cleaning and model training module 2, and a model training and verification module 3 which are electrically connected in sequence.

[0176] The flow field simulation and data acquisition module 1 is used to simulate the cyclone chaotic dry powder spray gun under different architecture parameter configurations using CFD-DEM coupling software, and the grid at the boundary of the cyclone blade and at the outlet of the spray gun is encrypted before calculation; the characteristic parameters of the flow field are extracted in the CFD-DEM coupling software; the ratio of the tangential momentum flux to the axial momentum flux of the fluid and the turbulent intensity of the jet in the spray gun are calculated for the density, coordinates, tangential velocity and axial velocity of each data point, and the calculated data is transmitted to the data cleaning and model training module 2; the data cleaning and model training module 2 is used to match the flow heat evidence parameters with the calculated data to build a data set; check the missing values in the data set, describe the distribution of the data and identify abnormal values using the box plot method of data distribution; and standardize the extracted structure parameters and flow field characteristics; output the processed training set to the artificial neural network model, adjust the model using the grid search method; adjust the network parameters through the back propagation algorithm to reduce the loss function value; evaluate the trained artificial neural network model using the validation set; the model training and verification module 3 is used to determine the key structure parameters that have the greatest impact on the target parameters through feature importance analysis, and set the parameter combinations; input the set parameter combinations into the trained artificial neural network model to predict the corresponding target parameters; compare the target parameters predicted by the artificial neural network model with the actual target parameters to minimize the difference between the structure parameters; use a genetic algorithm to randomly generate an initial structure parameter population, and predict the SN, Re and TRe of each parameter combination through the artificial neural network model; calculate the fitness function of each combination, that is, the difference between the predicted values of SN, Re and TRe and the target SN, Re and TRe; set a fitness function threshold, select the individual with the smallest difference from the threshold for crossover and mutation to form a new parameter population; repeat the selection, crossover and mutation process for continuous iteration and optimization until the maximum iteration number or the fitness function converges; perform CFD-DEM simulation verification on the most optimal fitness parameter combination; set a verification threshold, if higher than the threshold, continue to select individuals for crossover and mutation according to the fitness function to generate the most optimal fitness parameter combination;

[0177] Preferably, the use of the CFD-DEM coupling software of the present embodiment is critical, which can efficiently simulate the flow field and analyze the performance, and combined with the neural network technology, the prediction accuracy can be significantly improved and the calculation time can be saved. In addition, the application of neural networks in the core processes of CFD such as grid generation and numerical solution also shows its advantages of high efficiency and intelligence. The present embodiment uses deep learning to model the flow field, which not only can simulate and predict the flow field by learning the internal variation law of the flow field system, but also can reduce the prediction time and computing resources. In practical applications, the CFD-DEM coupling software can provide guidance for experiments to save a lot of manpower, material resources and time, and let researchers free from tedious programming and testing.

[0178] Therefore, the significance of the embodiment is that the CFD-DEM coupling software is combined with the neural network technology, which not only improves the efficiency and accuracy of the flow field simulation, but also provides a new idea and method for the optimal design of a complex fluid system.

[0179] As shown in Figure 15 The electronic device 4 includes a processor 41 and a memory 42 coupled to the processor 41.

[0180] The memory 42 stores program instructions for implementing the machine learning-based smelting cyclone dry powder spray gun design method of any of the above embodiments.

[0181] The processor 41 is configured to execute the program instructions stored in the memory 42 to optimize the structure parameters of the molten pool smelting cyclone chaotic dry powder spray gun.

[0182] The processor 41 can also be referred to as a CPU (Central Processing Unit). The processor 41 can be an integrated circuit chip with signal processing capability. The processor 41 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0183] Further, Figure 16 The storage medium 5 of the embodiment of the present application stores program instructions 51 capable of implementing all the above methods. The program instructions 51 can be stored in the above storage medium in the form of a software product, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media capable of storing program codes, or a computer, a server, a mobile phone, a tablet, and other terminal devices.

[0184] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the division of the units is only a logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0185] In addition, each function unit in the various embodiments of the present application can be integrated into a processing unit, or each unit can exist physically independently, or two or more units can be integrated into one unit. The integrated unit can be implemented in the form of hardware or in the form of a software functional unit. The above is only an embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

[0186] The specific embodiments of the application are described in detail above, but they are only examples. The present application is not limited to the specific embodiments described above. Any equivalent modification or substitution made by those skilled in the art to the present application is also within the scope of the present application, and therefore, any equivalent transformation, modification, improvement, etc. made without departing from the spirit and principle range of the present application should be included in the scope of the present application.

Claims

1. A machine learning based design method for a smelting cyclone dry powder lance, characterized by, The machine learning-based smelting cyclone dry powder spray gun design method comprises: Simulate the cyclone chaotic dry powder spray gun under different structural parameter configurations using CFD-DEM coupling software; extract the key feature parameters of the flow field in the CFD-DEM coupling software, calculate the ratio of the tangential momentum to the axial momentum of the fluid through the key feature parameters, and form the target parameter; Match the flow field feature parameters with the target parameters, construct a data set, and clean the data set; divide the data set into a training set and a validation set, train an artificial neural network model according to the training set, and evaluate the trained artificial neural network model using the validation set; Analyze the influence of each hyperparameter in the artificial neural network model on the model performance; determine the structural parameter that has the greatest influence on the target parameter, and set the parameter combination; calculate the target parameter of the parameter combination, and select individuals for crossover and mutation according to the calculation result to form an optimal parameter combination with the best fitness; Calculate the maximum number of iterations or the convergence of the fitness function, and if the new model combination reaches the simulation verification, if not, continue to perform crossover and mutation; The process of forming the target parameter comprises: Encrypt the cyclone chaotic dry powder spray gun locally, and simulate the cyclone chaotic dry powder spray gun under different structural parameter configurations using CFD-DEM coupling software at the key positions of the boundary of the cyclone blade and the outlet of the spray gun; Extract the key feature parameters of the flow field in the CFD-DEM software; Wherein, the key characteristic parameters include the density, coordinate, tangential velocity U tan , axial velocity U axi , inner diameter of the annular tube of the lance R, the range from R0 to R1 of the annular zone to the mixing zone, the characteristic length L of the lance, the velocity u of the jet and the integral volume V; Calculate the ratio SN of the tangential momentum flux to the axial momentum flux of the fluid and the turbulent intensity of the jet in the spray gun.

2. The machine learning based smelter cyclone dry powder lance design method of claim 1, wherein, The process of calculating the ratio SN of the tangential momentum flux to the axial momentum flux of the fluid and the turbulent intensity of the jet in the spray gun comprises: Extract the key feature parameters of the flow field in the CFD-DEM software; Calculate the ratio SN of the tangential momentum flux to the axial momentum flux of the fluid; the equation of SN is: where SN local represents the ratio of tangential momentum flux to axial momentum flux, represents the momentum distribution characteristics of the local region, G angular,local represents the tangential momentum flux of the local region, represents the momentum transfer of the fluid in the tangential direction, G linear,local represents the axial momentum flux of the local region, represents the momentum transfer of the fluid in the axial direction, R represents the radius, refers to the characteristic radius of the flow field or the radius of a certain specific region, and p represents the density of the fluid, represents the mass of the fluid per unit volume, r represents the radial distance, represents the distance from the center of the flow field to a certain point, U axi represents the axial velocity, represents the velocity component of the fluid in the axial direction, U tan represents the tangential velocity, represents the velocity component of the fluid in the tangential direction, R0 represents the lower limit of integration, represents the inner boundary radius of the flow field, R1 represents the upper limit of integration, represents the outer boundary radius of the flow field; The turbulent Reynolds number Re and the turbulent Reynolds TRe of the gas-phase introduced jet in the lance are used to evaluate the turbulent intensity of the jet in the lance; wherein Re is defined as the ratio of inertial force and friction force, and TRe is defined as the ratio of the turbulent kinetic energy k of the fluid and the friction work W g The expression is listed as: where Re represents the Reynolds number, which is the ratio of inertial forces to viscous forces, used to evaluate the intensity of turbulence in a fluid, p g represents the density of the gas phase fluid, represents the mass per unit volume of the gas phase fluid, u g represents the characteristic velocity of the gas phase fluid, represents the velocity of the gas phase fluid over a certain characteristic length, L represents the characteristic length, represents the characteristic dimension of the flow field, which is the width or height of the flow field, μ g represents the dynamic viscosity of the gas phase fluid, represents the viscous properties of the gas phase fluid, TRe represents the turbulent Reynolds number, which is the ratio of turbulent kinetic energy to the work done by friction, used to evaluate the intensity of turbulence, k g represents the turbulent kinetic energy of the gas phase fluid, represents the energy of turbulent motion in the gas phase fluid, V represents the volume of the fluid, represents the total volume of fluid in the flow field, represents the mean square value of the fluid velocity, represents the square of the mean velocity, represents the square value of the mean velocity of the fluid, W represents the work done by friction, represents the work done by friction in the fluid, μ represents the dynamic viscosity of the fluid, represents the viscous properties of the fluid, represents the velocity gradient, represents the rate of change of the fluid velocity in a certain direction, n v represents the normal vector, represents the unit vector in a certain direction in the flow field, <u i ′u j ′> represents the Reynolds stress tensor, represents the correlation of the velocity fluctuations in the fluid.

3. The machine learning based smelter cyclone dry powder lance design method of claim 2, wherein, The equation for evaluating the trained artificial neural network model using the validation set comprises: Match the flow field feature parameters with the target parameters, construct a data set, and clean the data set; divide the data set into an 80% training set and a 20% validation set; Adjust the artificial neural network model using the grid search method, select the artificial neural network model with the smallest loss function for training, train the artificial neural network model according to the training set, adjust the network parameters through the back propagation algorithm, and adopt the early stopping strategy to avoid overfitting; Evaluate the trained artificial neural network using the validation set, calculate the error index of the artificial neural network model on the validation set, set a threshold, if the threshold is exceeded, readjust the artificial neural network model, if the threshold is below, perform feature extraction and analysis.

4. The machine learning based smelter cyclone dry powder lance design method of claim 3, wherein, The process of constructing the data set comprises: Check the missing values in the data set; if there are missing values, handle them by filling in the mean values as follows: wherein x i′ the value at the position of the missing value, n' represents the total number of non-missing values in the dataset, x i′j′ represents the j'th feature value of the i'th sample or data point; Draw a box plot of each mechanism parameter and flow feature value; calculate the upper quartile, lower quartile and interquartile range of each parameter to determine the upper and lower threshold values; mark the values below the lower threshold or above the upper threshold as abnormal values; analyze the abnormal values by replacement or deletion method; Standardize the extracted structural parameters and flow field features; the standardization formula is as follows: where x norm is the standardized data, x is the original data, μ is the mean of the data, and σ is the standard deviation of the data.

5. The machine learning based smelter cyclone dry powder lance design method of claim 3, wherein, The process of training an artificial neural network model includes: designing an artificial neural network model network structure, selecting ReLU as an activation function, and selecting Huner loss as a loss function; inputting the training set into the artificial neural network model, and using the Adam optimizer to adjust the weights and biases of the artificial neural network model network; wherein the input layer of the artificial neural network model network structure includes 7 input features: blade length, blade rotation angle, blade installation height, blade tip clearance, blade thickness, CH4 pipe diameter, and mixing zone length; a plurality of hidden layers and an output layer include 3 output features: SN, Re, and TRe; determining the value range of the learning rate, the number of hidden layers, the number of nodes, the batch number, and the batch size parameters; all hyperparameters form a grid; for each combination of hyperparameters in the grid, the training set is used to train the artificial neural network model; using the validation set to verify the mean square error and accuracy of the artificial neural network model after each training, setting a threshold, and selecting the artificial neural network model corresponding to the hyperparameter combination with the smallest error compared to the threshold as the final artificial neural network model.

6. The machine learning based smelter cyclone dry powder lance design method of claim 1, wherein, The process of forming the optimal fitness parameter combination includes: using the SHAP importance evaluation method to analyze the influence of each structural parameter in the artificial neural network model on the target parameter; through the analysis of the importance of the target feature, determining the key structural parameter that has the greatest influence on the target parameter, setting the parameter combination, and inputting the combination into the trained artificial neural network model to predict the corresponding target parameter; adjusting the parameter combination according to the gap between the target parameter predicted by the artificial neural network model and the actual parameter, randomly generating an initial structural parameter population using a genetic algorithm; continuously iterating and optimizing until the maximum iteration number is reached or the function converges, and finally obtaining the optimal fitness parameter combination; selecting the optimal fitness parameter combination for CFD-DEM simulation verification, and testing and correcting the performance after verification.

7. The machine learning based smelter cyclone dry powder lance design method of claim 6, wherein, The process of predicting the corresponding target parameter includes: collecting the artificial neural network dataset for training, including input features and corresponding target parameters; after the model training is completed, using the SHAP method to calculate the contribution value of each input feature to the target parameter; setting a contribution value standard, comparing, and selecting features with a contribution value higher than the standard; sorting the contribution values obtained by comparison according to their importance; the higher the contribution value, the greater the influence of the feature on the target parameter, and the higher the relative importance; according to the sorting result, the key feature ranked first is selected; inputting the selected key feature combination into the trained artificial neural model to predict the target parameter; by comparing the actual parameter with the predicted parameter, if it is higher than a certain threshold, retraining and prediction are performed; repeat the process of predicting the corresponding target parameter until the maximum iteration number is reached or a preset standard is reached.

8. The machine learning based smelter cyclone dry powder lance design method of claim 6, wherein, The process of randomly generating an initial structural parameter population using a genetic algorithm includes: By analyzing the importance of features, determine the key structural parameters that have the greatest impact on the target parameters, and set the parameter combination, input the parameters into the trained artificial neural network model to predict the corresponding target parameters; By artificial neural network prediction output target parameters and the gap between the actual target parameters, adjust the structural parameters to minimize the gap; Randomly generate an initial population of structural parameters, and predict the SN, Re and TRe of each parameter combination through the artificial intelligence network model; Calculate the fitness function of each combination, that is, the gap between the predicted values of SN, Re and TRe and the target SN, Re and TRe; According to the fitness function, select individuals with better performance for crossover and mutation to generate a new parameter population; Repeat the process of selection, crossover and mutation, and continuously iterate and optimize until the maximum iteration number or the adaptive function converges; Finally, the parameter combination with the optimal fitness is verified by CFD-DEM simulation; Set the verification threshold, if higher than the threshold, continue to select individuals according to the fitness function for crossover and mutation to generate the optimal parameter combination.

9. A machine learning based design system for smelting cyclone dry powder injection lances, applied to the machine learning based design method for smelting cyclone dry powder injection lances according to any one of claims 1 to 8, characterized in that, It includes: Flow field simulation and data acquisition module, used to simulate the cyclone chaotic dry powder spray gun under different architecture parameter configurations using CFD-DEM coupling software; Before calculation, the grid at the boundary of the cyclone blade and the outlet of the spray gun is encrypted; Extract the characteristic parameters of the flow field in the CFD-DEM coupling software; Calculate the ratio of tangential momentum flux to axial momentum flux and the turbulent intensity of the jet in the spray gun for each data point density, coordinates, tangential velocity and axial velocity, and transfer the calculated data to the data cleaning and model training module; Data cleaning and model training module, used to match the flow field characteristic parameters with the calculated data to build a data set; Check for missing values in the data set, use the box plot method to describe the distribution of the data and identify outliers; And standardize the extracted structural parameters and flow field characteristics; Output the processed training set to the artificial neural network model, adjust the model using the grid search method; Adjust the network parameters through the back propagation algorithm to reduce the loss function value; Evaluate the trained artificial neural network model using the validation set; The model training and verification module is configured to determine key structural parameters with the greatest impact on the target parameter through feature importance analysis and set a parameter combination; input the set parameter combination into the trained artificial neural network model to predict the corresponding target parameter; compare the target parameter predicted by the artificial neural network model with the actual target parameter to minimize the difference between the structural parameters; generate an initial structural parameter population randomly by using a genetic algorithm and predict the SN, Re and TRe of each parameter combination by the artificial neural network model; calculate the fitness function of each combination, i.e., the difference between the predicted values of SN, Re and TRe and the target SN, Re and TRe; set a fitness function threshold, select the individual with the smallest difference from the threshold to perform crossover and mutation to form a new parameter population; repeat the process of selection, crossover and mutation, and continuously iterate and optimize until the maximum number of iterations or the fitness function converges; and perform CFD-DM simulation verification on the most optimal parameter combination with the best fitness. A verification threshold is set, and if the threshold is higher, the process of selecting individuals for crossover and mutation according to the fitness function will continue to generate the most optimal parameter combination with the best fitness.

Images