A method for intelligent production of stainless steel combining RKEF smelting and twin roll thin strip casting process

Abstract

The present application relates to the technical field of thin strip continuous casting, and particularly relates to a method for intelligently producing stainless steel by combining RKEF smelting and double-roller thin strip casting process, which directly couples the RKEF process with the stainless steel refining and thin strip continuous casting process, calculates the burden of the RKEF furnace based on the target product demand by using an intelligent burdening model, and then cooperatively optimizes and calculates the added burden of the AOD furnace based on the RKEF nickel molten iron composition by using the intelligent burdening model (and a dynamic blowing control model is used to realize real-time optimization of the proportion and flow of O2, Ar and N2 mixed gas), so as to realize cost-optimal dynamic burdening under multiple constraints; then, a neural network algorithm GAN with physical constraint PI is used to predict the composition and temperature, and the predicted values of the composition and temperature and control instructions are output, the metallurgical physical law is embedded in the model training as a hard constraint, the control decision is both consistent with the data and in line with the scientific law, the process control is accurately performed, the overall process is more energy-saving, and the carbon dioxide emission is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of thin strip continuous casting, and particularly relates to a method for intelligently producing stainless steel by combining RKEF smelting and a double-roller thin strip casting process. BACKGROUND

[0002] A traditional stainless steel strip preparation process includes: electric furnace / AOD furnace smelting; continuous casting slab; hot rolling; cold rolling; annealing. The entire process is long, has high energy consumption (comprehensive energy consumption > 500 kgce / ton) and CO2 emission > 2 tons / ton of steel.

[0003] In the prior art, there are also some innovations, such as:

[0004] Patent No. ZL201310619500.9 and patent name A new RKEF nickel-iron alloy production process with high-level short process and energy saving discloses the following steps: (1) drying; (2) batching; (3) roasting-pre-reduction; (4) electric furnace smelting; (5) casting forming, and the application also discloses a new RKEF process nickel-iron alloy production equipment with high-level short process and energy saving, which mainly comprises a rotary kiln, a drying kiln, an electric furnace workshop, a transfer station, a batching station and a raw material warehouse. One end of the rotary kiln is connected with the electric furnace workshop, the other end of the rotary kiln is connected with the transfer station, the drying kiln is located below the rotary kiln, one end of the drying kiln is connected with the electric furnace workshop, one end of the transfer station away from the rotary kiln is connected with the batching station, and the raw material warehouse is connected with one end of the drying kiln close to the electric furnace workshop through a belt conveyor. The patent only changes the process steps and optimizes the production equipment such as the rotary kiln, the drying kiln and the electric furnace, and does not apply the RKEF and double-roller thin strip direct continuous casting method.

[0005] Patent No. ZL201710009624.3 and patent name A method for producing nickel-iron by adopting a rotary kiln direct reduction RKEF combined method include the following steps: drying red clay nickel ore for standby; taking a proper amount of dried red clay nickel ore, crushing, screening and mixing with carbonaceous reducing agent and dolomite to obtain mixed material A, and the mass percentage of the carbonaceous reducing agent in the mixed material A is 17-25%; the mixed material A is made into pellets and sent into a first rotary kiln for roasting, and the roasted sand is screened and treated to obtain coarse nickel-iron particles containing 50-60% of slag; the coarse nickel-iron particles are mixed with dried red clay nickel ore and limestone and sent into a second rotary kiln for roasting, and the roasted sand is directly put into an electric arc furnace for melting, the secondary voltage of the electric arc furnace is 275-315 V, the primary current is 380-420 A, the temperature of the molten iron in the electric arc furnace is 1500-1540℃, and molten iron containing nickel-iron is obtained. The patent adopts a two-stage rotary kiln design, the first kiln is used for high-reducing agent ratio to strengthen direct reduction, the second kiln is used for RKEF, the electric arc furnace is operated at low voltage and high current, and the nickel recovery rate of 93.5% is a highlight, but the complexity of the process is increased.

[0006] The process principle of thin strip casting technology is to directly cast molten steel between a pair of counter-rotating and internally water-cooled crystallization rollers, and the molten pool level exists in a closed space composed of two side sealing plates and roller surfaces, and the molten steel solidifies between the two rollers to form a thin strip. Compared with the traditional continuous casting process, the thin strip casting technology has the advantages of short process, low production cost, energy saving and environmental protection, etc., and can directly cast and roll the molten steel into a strip steel with a thickness of 0.5-6mm without the need for reheating treatment. Therefore, an intelligent stainless steel production method combining RKEF smelting and double-roller thin strip casting process is designed to solve the problems existing in the prior art. SUMMARY

[0007] The purpose of the present application is to provide an intelligent stainless steel production method combining RKEF smelting and double-roller thin strip casting process, and the specific technical solutions are as follows:

[0008] An intelligent stainless steel production method combining RKEF smelting and double-roller thin strip casting process, comprising the following steps:

[0009] Step 1, based on the target product demand, an intelligent batching model is used to calculate the furnace charge of the RKEF furnace; the prepared furnace charge is loaded into the RKEF system, and the nickel iron melt with controllable composition and temperature is obtained by smelting;

[0010] Step 2, the nickel iron melt obtained in step 1 is hot charged to the AOD furnace through the ladle; based on the composition of the RKEF nickel iron melt, an intelligent batching model is used to cooperatively optimize and calculate the addition of the AOD furnace, and a dynamic blowing control model is used to real-time optimize the proportion and flow of O2, Ar and N2 mixed gas, to carry out desiliconization, decarburization and dephosphorization reactions, and obtain primary refined molten steel;

[0011] Step 3, the LF furnace receives the primary refined molten steel from the AOD furnace, and uses a neural network algorithm with physical constraints GAN to intelligently control the composition and temperature; after refining, refined molten steel is obtained, which specifically includes:

[0012] Step 3.1, the LF furnace refines the received molten steel;

[0013] Step 3.2, use the instrument assembly to obtain molten steel temperature data, three-phase electrode data, molten steel composition data and slag composition, and form a data set;

[0014] Step 3.3, use the neural network algorithm with physical constraints GAN to predict the composition and temperature, and output the predicted values of the composition and temperature and the control instructions;

[0015] Step 3.4, judging the composition and temperature predicted in step 3.3, if the composition and temperature meet the standards, entering the fourth step; if the composition and temperature do not meet the standards, adding the required alloy for composition fine-tuning and combining with the control instructions, returning to step 3.1;

[0016] Step 4, pouring the refined molten steel obtained in step 3 into a flow distributor system of a double-roller thin strip continuous casting through a tundish; the refined molten steel is obtained through double-roller casting to obtain an initial thin strip blank;

[0017] Step 5, cooling the initial thin strip blank obtained in step 4 to obtain a stainless steel thin strip.

[0018] Preferably, the furnace charge is obtained by taking laterite nickel ore as the main raw material and adding chromium ore, limestone, dolomite, quartzite, coke and coal; the charging includes at least one of waste stainless steel, high-carbon chromium iron, nickel iron / nickel plate, flux and reducing agent.

[0019] Preferably, the dynamic gas blowing control model is a decision model based on reinforcement learning, which dynamically adjusts the mixing ratio and flow of O2, Ar and N2 by real-time monitoring of furnace gas composition, flue gas temperature, molten pool temperature and sound signals.

[0020] Preferably, the intelligent charging model is a long short-term memory network model optimized by the CPO algorithm, and the charging calculation includes the following steps:

[0021] Step 2.1, collecting data and creating a data set, specifically: the input features of the data set are the target values of the C, Si, Mn, P, S, Cr and Ni elements of the target steel, the composition of the raw materials used, the composition and weight of the RKEF molten iron; the output label is the proportion of each raw material;

[0022] Step 2.2, pre-processing the data including cleaning, normalization and missing value processing to obtain a high-quality data set;

[0023] Step 2.3, initially selecting a long short-term memory network LSTM as an initial model; combining the high-quality data set and using the CPO algorithm to optimize the initial model to search for the optimal super parameter combination to obtain an intelligent charging model;

[0024] Step 2.4, using the intelligent charging model to calculate the charging, outputting the required materials and the corresponding amount.

[0025] Preferably, in step 3.2: using an infrared thermometer and a thermocouple to obtain molten steel temperature data; using a voltage current transformer to obtain the voltage, current and power of the three-phase electrode; using an online spectrometer to obtain molten steel composition data; using an online slag detector to obtain slag composition, including CaO, SiO2, Al2O3, MgO, FeO and MnO.

[0026] Preferably, the GAN in step 3.3 with physical constraints is physically constrained with the metallurgical reaction kinetics equation as a loss function, as follows:

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] ;

[0032] wherein: , , is the adaptive constraint weight; is the total physical loss term; is the energy conservation constraint term; is the mass transfer constraint term; is the slag-steel reaction constraint term; is the specific heat capacity of the molten steel at constant pressure; m is the mass of the molten steel in the ladle; is the change rate of the temperature of the molten pool; is the thermal efficiency of the LF furnace; is the arc voltage; is the arc current, is the power factor; is the total heat loss of the system; is the alloying element C, Si, Mn, P, S, Cr, and Ni; is the change rate of the concentration of the element ; is the mass transfer coefficient of the element ; is the total surface area of the alloying element ; is the density of the molten steel, is the saturation concentration of the element in the molten steel; is the current concentration of the element ; is the sulfur distribution ratio; is the sulfur content in the slag; is the sulfur content in the molten steel; is the equilibrium constant of the desulfurization reaction; is the activity of CaO in the slag; is the activity coefficient of S in the molten steel; is the activity of O in the molten steel; This refers to the phosphorus (P) content in the molten steel. The initial phosphorus (P) content of the molten steel; The rate constant for the phosphorus reversion reaction; For time.

[0033] Preferably, the control commands in step 3.4 include: the feeding speed of the feeders corresponding to low-carbon ferrochrome, ferronickel, and ferromolybdenum; the feeding amount of the feeders corresponding to calcium carbide and aluminum granules; the argon blowing flow rate of the argon control valve; and the arc power of the electrode controller.

[0034] Preferably, it also includes a sixth step, which is a thin strip intelligent defect detection and analysis, including the following steps:

[0035] Step 6.1: Use an online visual inspection system to obtain the defects and locations of the stainless steel strip;

[0036] Step 6.2: Introduce a comprehensive defect quality score to identify defect types, and quantify the severity of defects based on the comprehensive defect quality score obtained in Step 6.2.

[0037] Preferably, the comprehensive defect quality score in step 6.2 is calculated using the following formula:

[0038] ;

[0039] ;

[0040] in: A comprehensive score for defect quality; This is an index for defect categories, where: 1 is edge crack, 2 is subcutaneous pore, 3 is vibration mark, and 4 is dent. This represents the total number of defect categories. These are the weighting coefficients; The output of the model is considered to be the first. The probability value of a class of defects; A user-defined function for calculating defect severity; The physical characteristics that indicate a defect.

[0041] The preferred custom function for calculating defect severity is as follows:

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] in: is a user-defined function for crack severity; crack_length is crack length; L_critical is critical length threshold value; is a user-defined function for subcutaneous porosity severity; Avg_diameter_porosity is average diameter of porosity; D_porosity_critical is critical threshold value of diameter; is a user-defined function for oscillation severity; max_depth_oscillation is maximum depth of oscillation; D_oscillation_critical is critical threshold value of oscillation depth; is a user-defined function for concave severity; Depth_concave is concave depth; D_critical_concave is critical threshold value of concave depth; Area_concave is concave area; A_critical_concave is critical threshold value of concave area.

[0047] The disclosed intelligent stainless steel production method combined with RKEF smelting and double-roller thin strip casting process is as follows: the RKEF process is directly coupled with the stainless steel refining and thin strip continuous casting process, first, the burden of the RKEF furnace is calculated based on the target product demand using an intelligent burdening model, then the addition of the burden of the AOD furnace is cooperatively optimized and calculated based on the composition of the RKEF nickel molten iron using the intelligent burdening model (and the proportion and flow of the O2, Ar and N2 mixed gas are optimized in real time using a dynamic gas blowing control model), realizing cost-optimal dynamic burdening under multiple constraints; the composition and temperature are predicted using a neural network algorithm GAN with physical constraint PI, the predicted values of the composition and temperature and control instructions are output, the metallurgical physical law is embedded in the model training as a hard constraint, the control decision is both consistent with the data and in line with the scientific law, the process control is accurate, the overall process is more energy-saving, and the carbon dioxide (CO2) emission is reduced.

[0048] In addition to the purposes, features and advantages described above, the present application has other purposes, features and advantages. The present application will be further described below with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the drawings shown.

[0050] Figure 1is a schematic diagram of the method for intelligently producing stainless steel in combination with the RKEF smelting and double-roller thin strip casting process of the present application. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort fall within the protection scope of the present application.

[0052] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative positional relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications also change accordingly.

[0053] In addition, the descriptions such as “first”, “second” and the like in the present application are only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features defined as “first”, “second” 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, for example, two, three, etc., unless otherwise explicitly specified and limited.

[0054] In the present application, unless otherwise explicitly specified and limited, the terms “connection”, “fixation” and the like should be understood in a broad sense, for example, “fixation” can be fixed connection, or detachable connection, or integral; can be mechanical connection, or electrical connection; can be direct connection, or indirect connection through an intermediate medium; can be the internal connection of two elements or the interaction relationship between two elements, unless otherwise explicitly limited. For a person of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0055] In addition, the technical solutions of each embodiment of the present application can be combined with each other, but it must be based on the fact that a person of ordinary skill in the art can realize it, and when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, and is not within the protection scope claimed by the present application.

[0056] The present application discloses a method for intelligently producing stainless steel in combination with the RKEF smelting and double-roller thin strip casting process, which is described in detail in Figure 1 , comprising the following steps:

[0057] The first step is to calculate the furnace charge of the RKEF furnace based on the target product demand by using an intelligent batching model; the furnace charge is obtained by using laterite nickel ore as the main raw material and adding chromium ore, limestone, dolomite, quartzite, coke and coal. In this embodiment, the prepared furnace charge is loaded into the RKEF system, dried and pre-reduced in the rotary kiln, and then the hot charge is sent to the submerged arc furnace for high-temperature smelting to produce high-temperature low-grade nickel iron melt. This part of the process can refer to the prior art.

[0058] The second step is to transfer the iron melt obtained in the first step to the AOD furnace through an iron ladle; the intelligent batching model is used for batching and collaborative optimization calculation; at least one of waste stainless steel, high-carbon chromium iron, nickel iron / nickel plate, flux and reducing agent is added for primary refining to obtain primary refined molten steel. This also includes other alloy materials (selected according to actual needs). The dynamic gas blowing control model is used to optimize the proportion and flow of O2, Ar and N2 mixed gas in real time to carry out desiliconization, decarburization, chromium preservation and dephosphorization reactions to obtain primary refined molten steel.

[0059] In the present application, the intelligent batching model is a long short-term memory network model optimized by the CPO algorithm, and the batching calculation includes the following steps:

[0060] Step 2.1, collect data and create a data set, the input features of the data set are the target values of C, Si, Mn, P, S, Cr and Ni elements of the target steel grade, the components of the raw materials (here the raw materials include furnace charge and batching), the components and weight of the RKEF iron melt; the output label is the proportion of each raw material. The preferred operation is to collect historical production data, obtain the standard limit and control range based on the composition requirements (C, Si, Mn, P, S, Cr, Ni) of the target steel grade, the inventory and chemical composition data of each raw material, and the composition and weight of the RKEF iron melt, and construct a data set for batching decision.

[0061] Step 2.2, pre-process the data including cleaning, normalization and missing value processing to obtain a high-quality data set. In the present application, the data set is preferably divided into a training set and a test set in a ratio of 7:3.

[0062] Step 2.3, initially select a long short-term memory network LSTM as an initial model; combine the high-quality data set and use the CPO algorithm to optimize the initial model to search for the optimal combination of hyperparameters to obtain an improved hybrid model (i.e., an intelligent batching model).

[0063] In the application, the initial network structure is 2-layer LSTM layer. The CPO algorithm simulates the defense and attack behavior of the guanhuo porcupine, generates multiple "candidate solutions" in the hyperparameter space (Hidden_size, Dropout_rate, Batch_size, Learning_rate, Lr_factor, Patience), uses these combinations to train the LSTM model, evaluates its performance, and then iteratively updates according to the fitness value. Finally, the optimal hyperparameter combination is searched, and the improved hybrid model is obtained. The evaluation indicators of prediction error include the coefficient of determination (R 2 ), root mean square error (RMSE) and mean absolute percentage error (MAPE). When the prediction error R 2 of the improved hybrid model on the validation set reaches the first threshold, the RMSE and MAPE are less than or equal to the second threshold and the third threshold, respectively, indicating that the prediction accuracy and stability of the model meet the requirements. At this time, the training is terminated, and the obtained improved hybrid model not only performs well on the training data, but also has good generalization ability. It is deployed to the online batching calculation server.

[0064] Step 2.4, using the improved hybrid model for batching calculation, outputting the required substances and corresponding amounts. Specifically, when producing a new batch of molten steel for batching, the system inputs the composition of the current target steel grade, the composition of the used raw materials, the composition and weight of the RKEF molten iron, and the CPO-LSTM model calculates the optimal raw material ratio according to these input features, and outputs the results; record the actual final composition of each batch of molten steel (spectrum detector) and the actual yield of each element, and store this new data in the database.

[0065] In the application, the dynamic blowing control model is a decision model based on reinforcement learning, which dynamically adjusts the mixing ratio and flow of O2, Ar and N2 by real-time monitoring of furnace gas composition (mainly CO, CO2, O2), flue gas temperature, molten pool temperature and sound signal. Its optimization goal is to reduce the carbon content to below the target value in the shortest time, while minimizing the oxidation loss rate of chromium.

[0066] Third step, the LF furnace receives the primary refined molten steel from the AOD furnace, and uses the neural network algorithm GAN with physical constraints (Physical Information, PI for short) to intelligently control the composition and temperature; after refining, refined molten steel is obtained. In this step: through arc heating, the temperature of the molten steel is uniformly and accurately adjusted to the casting temperature range required by the Twin-roll Strip Casting (TRSC) process; through argon blowing stirring and slag refining, the composition is uniform, the inclusions are promoted to float, and the phosphorus reversion is prevented; according to the online rapid composition analysis results, the required alloy is added to accurately fine-tune the composition. When the temperature and composition meet the standards, the molten steel is transferred to the tundish and then subjected to TRSC thin strip casting.

[0067] The preferred specific embodiments of the present application in this step include:

[0068] Step 3.1, the LF furnace refines the received molten steel.

[0069] Step 3.2, use the instrument assembly to obtain molten steel temperature data, three-phase electrode data, molten steel composition data, and slag composition to form a data set. In the present application, it is preferred to use an infrared thermometer and a thermocouple to obtain molten steel temperature data; use a voltage and current transformer to obtain the voltage, current and power of the three-phase electrode; use an online spectrometer to obtain molten steel composition data; use an online slag detector to obtain slag composition, which includes CaO, SiO2, Al2O3, MgO, FeO, MnO and S. After obtaining the corresponding multi-source data, the data is preprocessed (including cleaning, normalization and missing value and outlier processing) to form a high-quality data set.

[0070] Step 3.3, use the neural network algorithm GAN with physical constraints to predict the composition and temperature, and output the predicted values of the composition and temperature and the control instructions. In the present application, the generator G is preferably used to predict the composition and temperature using the neural network algorithm GAN with physical constraints. The internal neural network model of the generator G learns a complex nonlinear mapping relationship from multi-source sensor data to predicted values and control instructions through training, and the training target is to make the predicted values as close to the true values as possible, and to make the control instructions effectively reduce the gap between the predicted values and the target values. The discriminator D receives the output of the generator G and calculates the physical loss using the built-in physical constraint equation. The task of D is to judge whether the instruction of G is "physically reasonable". Through repeated adversarial training of G and D, G generates instructions that meet the target and comply with physical laws. In the neural network algorithm GAN with physical constraints, the metallurgical reaction kinetics equation is used as a loss function for physical constraints, as follows:

[0071] ;

[0072] ;

[0073] ;

[0074] ;

[0075] ;

[0076] wherein: , , is the constraint weight of adaptive adjustment; is the total physical loss term; is the energy conservation constraint term; is the mass transfer constraint term; is the slag-steel reaction constraint term; is the specific constant volume heat capacity of the molten steel; m is the mass of the molten steel in the ladle; is the change rate of the temperature of the molten pool; is the thermal efficiency of the LF furnace; is the arc voltage; is the arc current, is the power factor; is the total heat loss of the system; are alloying elements C, Si, Mn, P, S, Cr and Ni; is the change rate of the concentration of element ; is the mass transfer coefficient of element ; is the total surface area of the alloying material of element ; is the density of the molten steel, is the saturation concentration of element in the molten steel; is the current concentration of element ; is the sulfur distribution ratio; is the sulfur content in the slag; is the sulfur content in the molten steel; is the equilibrium constant of the desulfurization reaction; is the activity of CaO in the slag; is the activity coefficient of S in the molten steel; is the activity of O in the molten steel; is the content of P in the molten steel; is the initial content of P in the molten steel; is the rate constant of the rephosphorization reaction; is time.

[0077] Step 3.4, judging the composition and temperature predicted in step 3.3, if the composition and temperature meet the standards, entering the fourth step; if the composition and temperature do not meet the standards, adding the required alloy to accurately fine-tune the composition and combining the control instructions, returning to step 3.1.

[0078] In the present application, the control instruction set is sent to the PID controller. The control instructions include: by controlling the alloy feeder, adding high-carbon chromium iron, low-carbon chromium iron, nickel-iron, and molybdenum iron alloy for cost-optimal accurate fine-tuning; by controlling the argon regulating valve, optimizing the argon blowing flow to achieve efficient stirring; by controlling the arc power of the electrode controller to control the temperature.

[0079] Step 4, the refined molten steel obtained in step 3 is poured into the TRSC flow distributor system through the tundish; the refined molten steel is obtained by double-roller casting to obtain the initial thin strip blank. That is, the molten steel is rapidly solidified in the molten pool formed by two high-speed rotating internal water-cooled cooling rollers, and is rolled and drawn out at the roll gap of the two rollers to form the initial thin strip blank (for reference to the existing thin strip continuous casting process).

[0080] Step 5, the initial thin strip blank obtained in step 4 is cooled to obtain a stainless steel thin strip. According to the present application, the freshly solidified thin strip blank is cooled by the gas mist cooling system according to the set cooling path; after cooling to the target coiling temperature, intelligent defect detection and analysis of the thin strip are carried out at the outlet position, and finally the thin strip is coiled into a steel coil by the coiler.

[0081] In addition, the method further includes a sixth step, which is intelligent defect detection and analysis of the thin strip, specifically including:

[0082] Step 6.1, using an online visual detection system to obtain the defects and positions of the stainless steel thin strip;

[0083] Step 6.2, introducing a defect quality comprehensive score as a recognition of the defect type, and quantifying the severity of the defect based on the defect quality comprehensive score obtained in step 6.2.

[0084] The preferred embodiment of this invention involves intelligent defect detection and analysis of thin strip steel. An online visual inspection system deployed at the export point acquires images of the strip steel surface. Image data is transmitted in real-time to an edge computing device, where an optimized CNN model performs real-time analysis. This model must complete inference for a single image within 50ms, classifying and locating defects such as surface cracks, edge cracks, dents, and vibration marks. The online visual inspection system is equipped with a high-temperature resistant (≥1400℃), dustproof protective cover and an adaptive light source, with an image acquisition frequency of no less than 100 frames / second. A defect quality comprehensive scoring method is introduced to optimize the CNN model. The optimized CNN model (i.e., the DefectCNN model) is deployed on an edge computing device with GPU acceleration, enabling it to complete the analysis and defect classification of a single image within 50ms. The optimized CNN model is a standard architecture composed of multi-layer convolution, pooling, and fully connected operations, but incorporates a defect quality comprehensive scoring method. This serves as a basis for identifying defect types and quantifying their severity.

[0085] The comprehensive defect quality score in this invention is calculated using the following formula:

[0086] ;

[0087] ;

[0088] in: A comprehensive score for defect quality; This is an index for defect categories, where: 1 is edge crack, 2 is subcutaneous pore, 3 is vibration mark, and 4 is dent. This represents the total number of defect categories. These are the weighting coefficients; The output of the model is considered to be the first. The probability value of a class of defects; A user-defined function for calculating defect severity; The physical characteristics that indicate a defect.

[0089] The custom function for calculating defect severity in this invention is as follows:

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] in: A custom function for crack severity; crack_length is the crack length; L_critical is the critical length threshold. is a user-defined function for the severity of subcutaneous porosity; Avg_diameter_porosity is the average diameter of the porosity; D_porosity_critical is the critical threshold of the diameter; is a user-defined function for the severity of oscillation; max_depth_oscillation is the maximum depth of the oscillation; D_oscillation_critical is the critical threshold of the oscillation depth; is a user-defined function for the severity of concave; Depth_concave is the depth of the concave; D_critical_concave is the critical threshold of the concave depth; Area_concave is the area of the concave; A_critical_concave is the critical threshold of the concave area.

[0095] In the present application, the system hierarchical response mechanism is when the defect quality comprehensive score S total is lower than the confidence L1, the system only records data; when S total is between the confidence L1 and L2, the system automatically calls the PLC execution mechanism to fine-tune the process parameters; when S total exceeds the confidence L2, the system triggers a high-level alarm and submits control suggestions to the operator for final decision confirmation, realizing intelligent production of man-machine cooperation.

[0096] In the present application, according to the output defect type and score, the hierarchical response mechanism based on confidence generates an optimized control instruction set, thereby constructing an intelligent stainless steel production system with “raw material-smelting-rolling-quality” full-process integration, data driving, and self-learning optimization.

[0097] The technical scheme of the present application has the following effects: the core of the present application is to construct an intelligent stainless steel production system with “raw material-smelting-rolling-quality” full-process integration, data driving, and self-learning optimization. The RKEF process is directly coupled with the stainless steel refining and thin strip continuous casting process. The CPO-LSTM model is adopted to realize cost-optimal dynamic proportioning under multiple constraints. The PI-GAN model is adopted to embed the metallurgical physical law as a hard constraint in the model training, so that the control decision not only fits the data but also conforms to the scientific law, and precise process control is realized. The DefectCNN model not only identifies defects, but also innovatively defines a defect quality comprehensive score to quantify the defect severity, providing a basis for intelligent decision-making.

[0098] Application case:

[0099] Take 70 tons of laterite nickel ore as the main raw material, add 5 tons of chromium ore, 8 tons of limestone, 4 tons of dolomite, 2 tons of quartz stone, 7.5 tons of coke, and 3.5 tons of coal. The prepared furnace charge is loaded into the RKEF system. First, dry and pre-reduce in the rotary kiln, then send the hot charge to the submerged arc furnace for high-temperature smelting. Output high-temperature low-grade nickel iron melt, composition includes: Ni 9.5%, Cr 2.5%, C 2.0%, Si 1.5%, Mn 1.2%, P 0.03%, S 0.05%, Fe balance, temperature is 1450℃.

[0100] The high-temperature nickel iron melt produced by RKEF is directly hot charged into the AOD furnace through the ladle. Using waste stainless steel, high-carbon chromium iron, nickel iron / nickel plate, other alloy elements, flux, and reducing agent as raw materials, intelligent raw material batching is carried out to adjust the final composition. In the AOD furnace, O2, Ar, and N2 mixed gas is blown in to carry out desiliconization, decarburization, chromium preservation, and dephosphorization reactions. When the C and P contents meet the requirements, the molten steel is transferred to the ladle.

[0101] Collect historical production data to obtain the standard limits and control ranges based on the target steel composition requirements (C, Si, Mn, P, S, Cr, Ni) (as shown in Table 1), the inventory and chemical composition data of each raw material, and the composition and weight of RKEF molten iron, to build the data basis for batching decision-making.

[0102] Table 1 Target steel composition control range (unit: wt%)

[0103]

[0104] Create a dataset, the input features of the dataset are the target values of the target steel C, Si, Mn, P, S, Cr, and Ni elements, and the compositions of the furnace charge and batching (as shown in Tables 2 and 3), and the composition and weight of RKEF molten iron. The output label is the ratio of each raw material;

[0105] Table 2 Composition of available furnace charge (unit: wt%)

[0106]

[0107] Table 3 Composition of available batching (unit: wt%)

[0108]

[0109] Clean, normalize, and handle missing values of the data to form a high-quality dataset. Divide the training set and test set in a ratio of 7:3.

[0110] The Crested Porcupine Optimizer (CPO) is used to optimize the long short-term memory network (LSTM) hybrid model (i.e., intelligent batching model) for batching calculation. The initial network structure of the CPO-LSTM model is set to two layers of LSTM. The CPO algorithm simulates the defense and attack behavior of the Crested Porcupine in the hyperparameter space (Hidden_size [32, 512], Dropout_rate [0.1, 1.0], Batch_size [16, 256], Learning_rate [0.0001, 0.01], Lr_factor [0.1, 1.0], Patience [1, 10]) to generate multiple "candidate solutions". These combinations are used to train the LSTM model, evaluate its performance, and then iteratively update the fitness value to search for the optimal hyperparameter combination as follows:

[0111] Hidden_size1=184;

[0112] Hidden_size2=78;

[0113] Dropout_rate=0.015;

[0114] Batch_size=32;

[0115] Learning_rate=0.012;

[0116] Lr_factor=0.33;

[0117] Patience=8。

[0118] The evaluation indicators of prediction error include the coefficient of determination (R 2 ), root mean square error (RMSE), and mean absolute percentage error (MAPE). When the prediction error of the CPO-LSTM model on the validation set is R 2 ≥ 0.93 (first threshold), RMSE ≤ 0.015 (second threshold), and MAPE ≤ 2% (third threshold), it indicates that the prediction accuracy and stability of the model meet the requirements. After training, the performance of the model on the validation set is R² = 0.955, RMSE = 0.012, and MAPE = 1.97%. The training is terminated, and the obtained CPO-LSTM model not only performs well on the training data but also has good generalization ability.

[0119] When a new batch of molten steel is produced, the system inputs the composition of the current target steel grade, the composition of the raw materials used, the composition and weight of the RKEF molten iron, and the CPO-LSTM model calculates the optimal raw material ratio that meets the requirements based on these input features:

[0120] Laterite nickel ore: 65.7 tons;

[0121] Chromium ore: 4.7 tons;

[0122] Limestone: 7.5 tons;

[0123] Dolomite: 3.7 tons;

[0124] Quartzite: 1.9 tons;

[0125] Coke: 7 tons;

[0126] Coal: 3 tons;

[0127] RKEF molten iron: 38 tons;

[0128] Waste stainless steel: 73.8 tons;

[0129] High-carbon chromium iron: 7.4 tons;

[0130] Nickel plate: 1.2 tons.

[0131] The dynamic air blowing control model dynamically decides and outputs adjustment instructions for the blowing scheme based on real-time monitoring of the furnace gas composition (CO: 15%, CO2: 5%, O2: 1.5%), flue gas temperature (1650°C), and molten pool temperature (1680°C) through a reinforcement learning algorithm. The blowing process is divided into a decarburization and chromium preservation period and a reduction period: in the decarburization and chromium preservation period, the model controls the O2 flow rate within the range of 1200-1500 Nm³ / h and mixes in a certain proportion of Ar gas (230 Nm³ / h). In the reduction period, when the carbon content decreases to the target value, the model significantly reduces the O2 flow rate and turns off the O2, increases the Ar gas flow rate (350 Nm³ / h) for stirring, and can inject a small amount of N2 (50 Nm³ / h) for composition control or cooling as needed. At the same time, the model controls the addition of reducing agents such as silicon iron to reduce chromium oxide. Through this dynamic control, the carbon content is reduced from the initial 2.0% to below 0.05% within 15 minutes, and the chromium oxidation loss rate is controlled within 1.5%.

[0132] The actual final composition of each batch of molten steel (spectrometer detection: C 0.05%, Si 0.25%, Mn 1.18%, P 0.032%, S 0.018%, Cr 18.35%, Ni 8.15%) and the actual yield of each element (Cr: 98.5%, Ni: 99.2%) are recorded and stored in the database.

[0133] The high-temperature infrared thermometer and continuous temperature thermocouple are used to obtain the temperature data of molten steel. The high-precision voltage and current transformer is used to obtain the voltage (U=250V), current (I=45kA) and power (P=10 MW) data of the three-phase electrode. The online spectrometer is used to obtain the composition data of molten steel (C 0.052%, Si 0.28%, Mn 1.18%, P 0.031%, S 0.018%, Cr 18.35%, Ni 8.15%). The slag online detector is used to obtain the composition of the slag (CaO=55%, SiO2=20%, Al2O3=10%, MgO=8%, FeO=0.8%, S=0.8%).

[0134] After the generator G receives the pre-processed high-quality data set, it outputs a set of preliminary prediction values (element [C]=0.05%, [Si]=0.28%, [Mn]=1.18%, [P]=0.031%, [S]=0.018%, [Cr]=18.48%, [Ni]=8.15%, T pred =1545℃) and preliminary control instructions through its complex neural network structure. The control instructions include: the feeding speed of the low-carbon chromium iron feeder is 50 kg / min, the feeding amount of the calcium carbide feeder is 50 kg, the argon flow of the argon control valve is 100 NL / min, and the arc power of the electrode controller is 9.5 MW.

[0135] The discriminator D receives the output of the generator G and calculates the physical loss using the built-in physical constraint equation. The task of D is to judge whether the instructions of G are “physically reasonable”. In this calculation, the thermal efficiency =0.72; the sulfur distribution ratio =44.4. Through repeated adversarial training of G and D, G generates instructions that meet the target and conform to the physical law.

[0136] The optimized control instruction set is sent to the PID controller. By controlling the alloy feeder, 200 kg of low-carbon chromium iron is added for cost-optimal precise fine-tuning; by controlling the argon regulating valve, the argon flow is optimized to achieve efficient stirring; by controlling the slag feeder, 20 kg of calcium carbide is added to make “white slag”; by controlling the arc power of the electrode controller, the temperature is controlled to 9.5 MW; when the temperature is uniformly reached 1550℃, the composition is fine-tuned to [C]=0.049%, [Si]=0.58%, [Mn]=1.21%, [P]=0.03%, [S]=0.015%, [Cr]=18.5% and [Ni]=8.2%, the molten steel is tapped and transferred to the tundish.

[0137] The online visual inspection system is equipped with a high-temperature-resistant (≥1400 DEG C) dustproof protective cover and a self-adaptive light source, and the image acquisition frequency is not less than 100 frames / s; the optimized CNN model is deployed on an edge computing device with GPU acceleration, and the analysis and defect classification of a single image can be completed within 50 ms. The hierarchical response mechanism of the system is that when the comprehensive quality score of the defect is less than the confidence L1=0.05, the system only records data; when the confidence is between L1=0.05 and L2=0.20, the system automatically calls the PLC execution mechanism to fine-tune the process parameters; when the confidence exceeds L2=0.20, the system triggers a high-level alarm and submits control suggestions to the operator for final decision confirmation, realizing intelligent production of man-machine cooperation. The online visual inspection system is equipped with a high-temperature-resistant (≥1400 DEG C) dustproof protective cover and a self-adaptive light source, and the image acquisition frequency is not less than 100 frames / s; the optimized CNN model is deployed on an edge computing device with GPU acceleration, and the analysis and defect classification of a single image can be completed within 50 ms. The hierarchical response mechanism of the system is that when the comprehensive quality score of the defect is less than the confidence L1=0.05, the system only records data; when the confidence is between L1=0.05 and L2=0.20, the system automatically calls the PLC execution mechanism to fine-tune the process parameters; when the confidence exceeds L2=0.20, the system triggers a high-level alarm and submits control suggestions to the operator for final decision confirmation, realizing intelligent production of man-machine cooperation. The online visual inspection system is equipped with a high-temperature-resistant (≥1400 DEG C) dustproof protective cover and a self-adaptive light source, and the image acquisition frequency is not less than 100 frames / s; the optimized CNN model is deployed on an edge computing device with GPU acceleration, and the analysis and defect classification of a single image can be completed within 50 ms. The hierarchical response mechanism of the system is that when the comprehensive quality score of the defect is less than the confidence L1=0.05, the system only records data; when the confidence is between L1=0.05 and L2=0.20, the system automatically calls the PLC execution mechanism to fine-tune the process parameters; when the confidence exceeds L2=0.20, the system triggers a high-level alarm and submits control suggestions to the operator for final decision confirmation, realizing intelligent production of man-machine cooperation. The online visual inspection system is equipped with a high-temperature-resistant (≥1400 DEG C) dustproof protective cover and a self-adaptive light source, and the image acquisition frequency is not less than 100 frames / s; the optimized CNN model is deployed on an edge computing device with GPU acceleration, and the analysis and defect classification of a single image can be completed within 50 ms. The hierarchical response mechanism of the system is that when the comprehensive quality score of the defect is less than the confidence L1=0.05, the system only records data; when the confidence is between L1=0.05 and L2=0.20, the system automatically calls the PLC execution mechanism to fine-tune the process parameters; when the confidence exceeds L2=0.20, the system triggers a high-level alarm and submits control suggestions to the operator for final decision confirmation, realizing intelligent production of man-machine cooperation.

[0138] The PLC execution mechanism comprises a hydraulic servo system, an aerosol cooling system, a cooling roller control system, a ladle pouring system and a curling system.

[0139] The offline detection results of the final product, such as thickness, plate shape, mechanical properties and metallographic structure (thickness 2.25 mm, yield strength 280 MPa, tensile strength 620 MPa, elongation 55%; metallographic structure: austenite + 5% delta ferrite), are transmitted to the self-learning training module to automatically optimize the control model, and the online control model is continuously improved and optimized through incremental learning combined with real-time production data.

[0140] The above only describes the preferred embodiments of the present application and

Claims

1. A method of intelligent production of stainless steel combining RKEF smelting and twin roll thin strip casting process, characterized in that, The method comprises the following steps: The first step is to calculate the burden of the RKEF furnace based on the requirements of the target product by using an intelligent burdening model, to load the prepared burden into the RKEF system, and to smelt nickel molten iron with controllable composition and temperature; The second step is to hot charge the nickel molten iron obtained in the first step into an AOD furnace through a molten iron ladle; based on the composition of the RKEF nickel molten iron, the burden added to the AOD furnace is calculated and optimized by using the intelligent burdening model, and the proportion and flow rate of the mixed gas of O2, Ar and N2 are optimized in real time by using a dynamic gas blowing control model, so as to carry out desiliconization, decarburization, chromium preservation and dephosphorization reactions, and obtain primary refined molten steel; The third step is to receive the primary refined molten steel from the AOD furnace by using an LF furnace, and to intelligently control the composition, gas blowing and temperature by using a neural network algorithm with physical constraints (GAN); After the refining is completed, refined molten steel is obtained, which specifically comprises: Step 3.1, the LF furnace refines the received molten steel; Step 3.2, an instrument assembly is used to obtain molten steel temperature data, three-phase electrode data, molten steel composition data and slag composition, so as to form a data set; Step 3.3, the neural network algorithm with physical constraints (GAN) is used to predict the composition and temperature, and the predicted values of the composition and temperature and control instructions are outputted; Step 3.4, the composition and temperature predicted in step 3.3 are judged, if the composition and temperature meet the requirements, the fourth step is entered, if the composition and temperature do not meet the requirements, the required alloy is added for accurate fine adjustment of the composition and the control instructions are combined, and step 3.1 is returned to; The fourth step is to inject the refined molten steel obtained in the third step into a flow distributor system of a double-roller thin strip continuous casting through a tundish; the refined molten steel is cast into an initial thin strip blank by using a double-roller casting mill; The fifth step is to obtain stainless steel thin strips by cooling the initial thin strip blank obtained in the fourth step.

2. The method of intelligent production of stainless steel incorporating RKEF melting and twin roll thin strip casting process as claimed in claim 1 wherein, The burden is obtained by taking laterite nickel ore as the main raw material and adding chromium ore, limestone, dolomite, quartzite, coke and coal; the burden includes at least one of waste stainless steel, high-carbon chromium iron, nickel iron / nickel plate, flux and reducing agent.

3. The method of intelligent production of stainless steel incorporating RKEF melting and twin roll thin strip casting process as claimed in claim 1 wherein, The dynamic gas blowing control model is a decision model based on reinforcement learning, which dynamically adjusts the mixing ratio and flow rate of O2, Ar and N2 by real-time monitoring of furnace gas composition, flue gas temperature, molten pool temperature and sound signals.

4. The method of intelligent production of stainless steel incorporating RKEF melting and twin roll thin strip casting process as claimed in claim 1 wherein, The intelligent burdening model is a long short-term memory network model optimized by the CPO optimization algorithm; The calculation of the intelligent burdening model comprises the following steps: Step 2.1, collect data and create a data set, specifically: the input features of the data set are the target values of the C, Si, Mn, P, S, Cr and Ni elements of the target steel, the composition of the raw materials used, the composition and weight of the RKEF molten iron; the output label is the proportion of each raw material; Step 2.2, pre-process the data including cleaning, normalization and missing value processing to obtain a high-quality data set; Step 2.3, initially select a long short-term memory network (LSTM) as an initial model; combine the high-quality data set to search for the optimal super parameter combination of the initial model by using the CPO optimization algorithm to obtain an intelligent burdening model; Step 2.4, calculate the burden by using the intelligent burdening model, and output the required substances and corresponding amounts of the raw materials.

5. The method of intelligent production of stainless steel incorporating RKEF melting and twin roll thin strip casting process as claimed in claim 1 wherein, In step 3.2: the temperature data of the molten steel is obtained by using an infrared temperature measuring instrument and a thermocouple; the voltage, current and power of the three-phase electrode are obtained by using a voltage and current transformer; the composition data of the molten steel is obtained by using an online spectrometer; and the composition of the slag, including CaO, SiO2, Al2O3, MgO, FeO, MnO and S, is obtained by using an online slag detector.

6. The method of intelligent production of stainless steel incorporating RKEF melting and twin roll thin strip casting process as claimed in claim 1 wherein, In the neural network algorithm GAN with physical constraints in step 3.3, the metallurgical reaction kinetics equation is used as a loss function for physical constraints, as follows: ； ； ； ； ； wherein: , , is the constraint weight of adaptive adjustment; is the total physical loss term; is the energy conservation constraint term; is the mass transfer constraint term; is the slag-steel reaction constraint term; is the specific heat capacity of the molten steel at constant pressure; m is the mass of the molten steel in the ladle; is the rate of change of the bath temperature; is the thermal efficiency of the LF furnace; is the arc voltage; is the arc current, is the power factor; is the total heat loss of the system; are the alloying elements C, Si, Mn, P, S, Cr, and Ni; is the concentration of element ; is the mass transfer coefficient of element ; is the total surface area of the alloying elements ; is the density of the molten steel, is the saturation concentration of element in the molten steel; is the current concentration of element ; is the sulfur partition ratio; is the sulfur content in the slag; is the sulfur content in the molten steel; is the equilibrium constant of the desulfurization reaction; is the activity of CaO in the slag; is the activity coefficient of S in the molten steel; is the activity of O in the molten steel; is the content of P in the molten steel; is the initial P content in the molten steel; is the rate constant of the rephosphorization reaction; is time.

7. The method of intelligent production of stainless steel combining RKEF smelting and twin roll thin strip casting process as claimed in claim 6 wherein, The control instructions in step 3.4 include: the feeding speed of the corresponding feeding machine for low-carbon ferrochrome, ferronickel and ferromolybdenum; the feeding amount of the corresponding feeding machine for calcium carbide and aluminum particles; the argon blowing flow of the argon control valve; and the arc power of the electrode controller.

8. The method of intelligent production of stainless steel combining RKEF smelting and twin roll strip casting process as claimed in any one of claims 1 to 7, wherein, The sixth step, which is thin strip intelligent defect detection and analysis, includes the following steps: Step 6.1: obtaining the defects and positions of the stainless steel thin strip by using an online visual detection system; Step 6.2: introducing a defect quality comprehensive score as a recognition of the defect type, and quantifying the severity of the defect based on the obtained defect quality comprehensive score.

9. The method of intelligent production of stainless steel incorporating RKEF melting and twin roll thin strip casting process as claimed in claim 8 wherein, The defect quality comprehensive score in step 6.2 is calculated as follows: ； ； wherein: is the defect quality comprehensive score; is the defect category index, wherein: 1 is edge crack, 2 is subsurface porosity, 3 is shake mark, 4 is indentation; is the total number of defect categories; is the weight coefficient; is the probability value considered by the model to be the category defect; is a self-defined function for calculating defect severity; represents the physical characteristics of the defect.

10. The method of intelligent production of stainless steel incorporating RKEF melting and twin roll thin strip casting process as claimed in claim 9 wherein, The self-defined function for calculating the defect severity is as follows: The self-defined function for calculating the defect severity is as follows: ； ； ； ； wherein: is a custom function for crack severity; crack_length is the crack length; L_critical is the critical length threshold; is a custom function for sub-surface porosity severity; Avg_diameter_porosity is the average diameter of the porosity; D_porosity_critical is the critical threshold for diameter; is a custom function for oscillation severity; max_depth_oscillation is the maximum depth of the oscillation; D_oscillation_critical is the critical threshold for oscillation depth; is a custom function for concave severity; Depth_concave is the depth of the concave; D_critical_concave is the critical threshold for concave depth; Area_concave is the area of the concave; A_critical_concave is the critical threshold for concave area.

Images