Steel structure performance control method and system based on large language model

Abstract

The invention relates to a steel structure performance control method and system based on a large language model, and belongs to the field of steel manufacturing. In order to solve the problems that traditional steel rolling water cooling parameter control depends on an empirical formula, prediction precision is low, adjusting efficiency is poor and working condition adaptability is insufficient, initial parameters (chemical components, temperature, technological parameters and the like) and mechanical property data of steel are collected, and a historical knowledge base is established; preprocessing the to-be-predicted data, calculating the similarity between the to-be-predicted data and historical cases, screening reference cases, and constructing Prompt input containing the cases and the current working condition; a large language model (GPT / GLM) is utilized to generate high-precision water tank parameters, and the model is optimized through a continuous learning mechanism. The method has the technical effects that the prediction error is smaller than or equal to 5%, the parameter adjustment efficiency is improved by 70%, the rejection rate is reduced by 3%-5%, complex working conditions such as equipment aging and component fluctuation are dynamically adapted, and an intelligent and high-robustness cooling control scheme is provided for steel production.

Description

Technical Field

[0001] The present invention belongs to the field of steel manufacturing and relates to a steel structure and performance control method and system based on a large language model. Background Art

[0002] As the steel industry develops towards intelligence and high precision, process control of steel cooling process has become the core link that determines the mechanical properties and microstructure uniformity of the product. Traditional methods for controlling water cooling parameters in steel rolling mainly rely on the following two types of technologies:

[0003] (1) Empirical formula driven control: Based on the mathematical model constructed based on metallurgical principles (such as heat conduction equation, phase change kinetic model), the cooling rate, water volume, pressure and other parameters are set by combining manual experience. This method has significant defects:

[0004] Poor adaptability: Dynamic factors such as fluctuations in steel composition and changes in equipment status are difficult to cover with fixed formulas, resulting in cumulative prediction errors;

[0005] Ignoring parameter coupling: The multi-physics coupling effects of temperature field, water flow field, and microstructure evolution during the cooling process are not fully modeled, resulting in "parameter mismatch" (such as local overcooling or reheating) in actual production;

[0006] Dependence on expert intervention: Frequent manual adjustment of parameters is required, the degree of automation is low, and the response speed cannot meet the needs of high-speed rolling production lines (for example, the delay in handling abnormal working conditions exceeds 10 minutes).

[0007] (2) Traditional machine learning-assisted control: In recent years, some studies have attempted to use algorithms such as random forests and support vector machines to train prediction models using historical data. However, such methods have inherent limitations:

[0008] Insufficient data representation capabilities: The complex nonlinear relationship between process parameters (such as "crystallizer water inlet temperature" and "refining bottom blowing time") and performance indicators (such as "super-yield ratio") is difficult to fully explore with shallow models;

[0009] Cold start problem: When new steel grades and new processes are launched, the model fails due to lack of annotated data;

[0010] Lack of interpretability: The parameter adjustment suggestions output by the black box model lack metallurgical logic support and are not well accepted by engineers.

[0011] In addition, existing technologies generally have the problem of low utilization of historical knowledge: although steel companies have accumulated a large amount of production data (for example, a steel plant stores an average of 200,000 rolling records annually), traditional methods only use it for offline statistical analysis and fail to achieve real-time case matching and knowledge transfer, resulting in frequent "repeated debugging of similar working conditions".

[0012] Existing methods cannot meet the production needs of high-value-added steel grades (such as automotive plates and pipeline steel) in terms of prediction accuracy (average error > 15%), real-time performance (adjustment cycle > 5 minutes), and process interpretability, becoming a key bottleneck restricting the steel industry from improving quality and efficiency.

[0013] To address the above problems, the present invention innovatively introduces large language model technology. Through its powerful semantic understanding, context association and generation capabilities, it transforms the historical case library into a reasonable "process knowledge graph", realizes intelligent recommendation and dynamic optimization of cooling parameters, and ultimately breaks through the performance ceiling of traditional methods. Summary of the Invention

[0014] In view of this, an object of the present invention is to provide a method and system for controlling steel structure and properties based on a large language model.

[0015] In order to achieve the above object, the present invention provides the following technical solutions:

[0016] A method for controlling steel microstructure and properties based on a large language model comprises the following steps:

[0017] Collect the initial parameters of steel, including basic information parameters, chemical composition parameters, equipment parameters, temperature parameters, heating furnace parameters and process parameters;

[0018] Establishing a historical knowledge base, wherein the historical knowledge base includes initial parameters of historical data and corresponding cooling process water tank parameters;

[0019] Preprocess the data to be predicted and calculate the similarity between it and historical data;

[0020] Filter reference cases based on similarity, and construct an instruction guide prompt input containing the reference cases and the data to be predicted;

[0021] Use a large language model to train and predict the Prompt to generate target water tank parameters;

[0022] Deploy the predicted water tank parameters to the production environment and optimize the model based on actual production feedback.

[0023] Furthermore, the similarity calculation adopts at least one algorithm of Euclidean distance and cosine similarity.

[0024] Furthermore, the large language model is a pre-trained language model based on the Transformer architecture, including a generative pre-trained transformer (GPT) or a general language model (GLM).

[0025] Furthermore, the construction of the Prompt includes:

[0026] Perform structured text splicing on the initial parameters of the reference case, the water tank parameters and the initial parameters of the data to be predicted;

[0027] Add natural language instructions to guide the model to generate target tank parameters.

[0028] Furthermore, the model optimization includes:

[0029] Regularly update the historical knowledge base to incorporate new production data;

[0030] Tune model parameters via online learning or incremental learning.

[0031] The enhanced large language model-based steel structure and property control method based on the method further includes:

[0032] Use data augmentation technology to expand historical data and generate synthetic data;

[0033] Use autoencoders or variational autoencoders to reduce the dimensionality of the data to optimize similarity calculation;

[0034] Evaluate nonlinear similarity relationships through fuzzy logic or neural networks;

[0035] Build a dynamic prompt to dynamically adjust the input content based on the weight of the reference case.

[0036] Furthermore, the data enhancement technology includes synthetic minority over-sampling technology (Synthetic Minority Over-sampling Technique, SMOTE) or generative adversarial networks (Generative Adversarial Networks, GANs).

[0037] A steel structure and performance control system based on a large language model, comprising:

[0038] A data collection module is configured to collect the initial parameters of the steel and the mechanical properties parameters after cooling from the steel rolling production line in real time, wherein the initial parameters include chemical composition, temperature parameters and process parameters, and transmit them to the preprocessing module via a data bus;

[0039] A historical knowledge base module is connected to the preprocessing module and stores a historical data set after cleaning and normalization, wherein the data set includes a mapping relationship between initial parameters and corresponding water tank parameters;

[0040] The preprocessing module receives the real-time data from the data collection module and the offline data from the historical knowledge base, performs normalization and missing value filling operations, and outputs the processed data to be predicted to the similarity calculation module;

[0041] The similarity calculation module calls the historical data in the historical knowledge base, calculates the similarity matrix between the data to be predicted and the historical data based on the Euclidean distance or cosine similarity algorithm, selects the reference case set with a similarity higher than the threshold θ, and transmits the case set to the Prompt generation module;

[0042] The prompt generation module receives the reference case set output by the similarity calculation module and the preprocessed data to be predicted, and splices the initial parameters, water tank parameters and current working condition parameters of the case into structured text according to a preset template to generate a prompt instruction stream that meets the input specifications of the large language model;

[0043] The large language model inference module loads the pre-trained Transformer architecture language model, receives the instruction stream from the Prompt generation module, and outputs a set of predicted water tank parameters, including flow rate, pressure value and its confidence score;

[0044] The deployment and feedback module communicates with the rolling line PLC control system, sends the predicted parameters to the cooling water tank actuator, collects actual production results in real time, calculates the error between the predicted value and the measured value, and feeds the error data back to the historical knowledge base module and the large language model inference module to trigger online fine-tuning of the model parameters.

[0045] Furthermore, the system further includes a data enhancement module connected between the historical knowledge base module and the preprocessing module, configured as follows:

[0046] Receive raw data from a historical knowledge base and generate synthetic data using synthetic minority oversampling technology (SMOTE) or generative adversarial networks (GANs). The synthetic data meets preset metallurgical constraints, including the composition conservation law Ceq ≤ 0.50% and the temperature gradient ≤ 200°C / m.

[0047] The synthetic data is mixed with the original data and then fed into the historical knowledge base module to expand the training data set, where the synthetic data accounts for no more than 30% of the total data volume;

[0048] Trigger condition: When the amount of data for a certain steel grade in the historical knowledge base is less than 500, the data enhancement process is automatically started.

[0049] Furthermore, the similarity calculation module further includes:

[0050] The fuzzy logic evaluation submodule is configured as follows:

[0051] Define a fuzzy rule base, including a "composition difference-weight mapping table," such as C content difference ΔC∈[0,0.1%] → weight coefficient 0.8 to 1.0; perform fuzzy correction on the Euclidean distance calculation result, and output the corrected similarity score S'=S×α, where S is the original similarity score and α is the fuzzy weight coefficient;

[0052] Neural network evaluation submodule, configured as:

[0053] Train a similarity network based on the attention mechanism, with the input being the difference vector between the data to be predicted and the historical data, and the output being a nonlinear similarity score; starting condition: when the data missing rate is greater than 15%, the neural network submodule is preferentially called for scoring.

[0054] The beneficial effects of the present invention are:

[0055] (1) By utilizing the strong contextual reasoning capabilities of large language models (such as GPT and GLM), we can effectively model the complex nonlinear relationships between chemical composition (such as Ceq carbon equivalent), temperature field (such as "opening rolling bar temperature"), process parameters (such as "refining bottom blowing time") and mechanical properties (such as "super yield ratio"), and reduce the prediction error of water tank parameters (water volume, pressure) from an average of more than 15% of traditional methods to less than 5%.

[0056] By matching historical case similarities (such as filtering reference cases using Euclidean distance), cold start problems can be avoided and the parameter tuning cycle when a new steel grade is launched can be shortened by 70% (from several hours to within 30 minutes).

[0057] (2) Based on data enhancement technology (such as SMOTE, GANs) and continuous learning mechanism, the model can automatically adapt to production line fluctuations (such as equipment aging, fine-tuning of steel grade composition), and the parameter prediction accuracy fluctuation range is controlled within ±2% when the steel specifications are switched.

[0058] The similarity assessment module that combines fuzzy logic and neural networks solves the sensitivity of traditional algorithms to noisy data and missing values, and can still maintain a prediction reliability of more than 85% when 20% of the data is missing.

[0059] (3) The Prompt project makes the association between the reference case and the current working conditions explicit (e.g., “When the steel grade is X, the flow parameter should refer to case Y”), so that the model output has metallurgical logic support, and engineers’ acceptance of parameter adjustment suggestions has increased by more than 40%.

[0060] The continuous expansion of the historical knowledge base (an average of 100,000 new production data items per year) forms a "process decision-making knowledge graph", realizing the digital precipitation and reuse of the company's experience assets.

[0061] (4) Real-time parameter recommendations (response time < 1 second) support full-speed operation of the rolling production line (for example, the rolling speed is increased from 12 m / s to 15 m / s), and the annual production capacity of a single production line increases by 8%-12%.

[0062] By precisely controlling cooling uniformity, the need for re-melting and re-rolling due to substandard structural performance is reduced (scrap rate drops by 3%-5%), and energy consumption per ton of steel is reduced by approximately 50kWh.

[0063] (5) The modular design of the system (such as independent data enhancement modules and feedback mechanism modules) supports seamless integration with existing MES and PLC systems, shortening the transformation implementation cycle to within 2 weeks.

[0064] The model supports cross-steel grade transfer learning (such as migration from automotive plates to pipeline steel control), reducing the data required for new scenario model training by 60%, significantly lowering deployment costs.

[0065] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0067] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0068] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0069] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.

[0070] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0071] See also Figure 1 It is a method for predicting water cooling parameters of steel rolling based on large language model technology. By constructing a language model with strong generalization and reasoning capabilities, it can achieve accurate prediction of water tank parameters during the steel cooling process.

[0072] Before the steel is cooled, the initial parameters of the steel are collected, including basic information parameters of the steel, chemical composition parameters, equipment parameters, temperature parameters, heating furnace parameters and process parameters.

[0073] Basic steel information parameters: "Bill number", "Furnace number", "Shift", "Shift type", "Steel grade", "Weighing (kg)", "Real weight (kg)", "Final length", "Finished product length", "Cold and hot charging", "Process", "Actual billet length (mm)", "Bill size", "Specifications", "Source".

[0074] Chemical composition parameters: "C","Si","Mn","P","S","Cr","Ni","Cu","N","V","Ceq".

[0075] Equipment parameters: current, torque, speed.

[0076] Temperature parameters: "Converter exit temperature (°C)", "Refining entry temperature (°C)", "Refining exit temperature (°C)", "Continuous temperature measurement of tundish (°C)", "Manual temperature measurement of tundish (°C)", "Temperature of low-pressure turbid water supply pipe of water tank", "Temperature of turbid water return pipe of water tank", "Temperature of high-pressure turbid water supply pipe", "Start rolling temperature", "1# shear temperature", "Start rolling rod temperature", "Intermediate rolling outlet temperature", "2# shear temperature", "Pre-finishing rolling outlet temperature 1", "Pre-finishing rolling outlet temperature 2", "3# shear temperature 1", "3# shear temperature 2", "3# shear temperature 3", "Temperature before double-length shear 1", "Temperature before double-length shear 2", "Temperature before double-length shear 3", "Upper cooling bed temperature", "Temperature of rod clearing of 1# shear", "Temperature of rod clearing of 2# shear".

[0077] Heating furnace parameters: "Second heating temperature (°C)", "Second heating duration", "First heating temperature (°C)", "First heating duration", "Soaking section temperature (°C)", "Soaking section time".

[0078] Process parameters: "Zone 1 flow rate", "Zone 2 flow rate", "Zone 3 flow rate", "Zone 4 flow rate", "Zone 1 pressure", "Zone 2 pressure", "Zone 3 pressure", "Zone 4 pressure", "Mold flow rate", "Mold inlet water temperature", "Mold outlet water temperature", "1st baffle waiting time (s)", "1st baffle arrival time", "1st baffle arrival temperature", "Furnace entry temperature (°C)", "Furnace entry time", "Furnace exit time", "Duration in furnace (min)", "3rd heating temperature (°C)", "3rd heating duration", "Rolling interval", "2# stand rolling interval"), "Cutting start time", "Cutting end time", "Stream number", "Pulling speed", "Tundish weight", "Refining bottom blowing duration (min)", "Refining power-on duration (min)".

[0079] After the steel is cooled, the mechanical properties of the steel are collected. The mechanical properties include: tensile strength (MPa), yield strength (MPa), yield strength ratio (MPa), and yield strength ratio (MPa).

[0080] We refer to the initial parameters and mechanical properties of steel as the initial conditions of steel, and use vector x i To represent, vector x i Each dimension represents a physical quantity, vector x i The dimension size is d x , we use X to represent the set of initial conditions of historical data.

[0081] During the cooling process of steel, people need to control the amount of water and pressure in the water tank during the cooling process. We use the vector y i Used to represent the water volume and pressure of the water tank, vector y i Each dimension represents a physical quantity, vector y i The dimension size is d y , we use Y to represent the set of water tank parameters of historical data.

[0082] In historical data, Under the initial parameters of the steel, the water tank parameters are adjusted according to y i Setting, we can reach vector x i The desired temperature in .

[0083] The steel cooling parameter prediction problem can be expressed as: given X, Y, for the new data of the steel plant We need to j Predict water tank parameters y j, so that the water tank parameters are in accordance with y j Setting, we can reach vector x j Expected mechanical properties.

[0084] The present invention proposes two steel structure and performance control models based on large language models. The specific implementation steps are as follows:

[0085] Model 1: Large language steel microstructure and performance control model based on reference cases

[0086] Step 1: Build a historical knowledge base

[0087] Collect historical data, including the initial conditions of the steel (basic information parameters of the steel, chemical composition parameters, equipment parameters, temperature parameters, heating furnace parameters, process parameters and mechanical properties parameters) and the corresponding water tank parameters (water volume, pressure).

[0088] Clean and preprocess the data, including removing outliers, filling missing values, feature engineering, etc.

[0089] Step 2: Similarity calculation

[0090] Data preprocessing: Normalize the predicted data and historical data to ensure the consistency of similarity calculation.

[0091] Similarity evaluation: Use methods such as Euclidean distance and cosine similarity to calculate the similarity between the data to be predicted and the historical data.

[0092] Step 3: Reference Case Selection

[0093] Threshold setting: Set a similarity threshold θ. Only when the similarity of historical data exceeds this threshold will it be considered as a reference case.

[0094] Case screening: Based on the similarity calculation results, the historical data that is most similar to the data to be predicted is screened out as a reference case.

[0095] Step 4: Prompt construction and model training

[0096] Reference case integration: Writes the initial conditions and tank parameters of the selected reference case to the prompt, along with the initial conditions of the data to be predicted.

[0097] Select a suitable large language model, such as GPT, GLM, etc., and perform customized training.

[0098] Model training: Use historical datasets to train large language models, including Prompt construction and parameter prediction.

[0099] Apply cross-validation and hyperparameter tuning techniques to optimize model performance.

[0100] Step 5: Model Validation and Testing

[0101] An independent validation set is used to validate the trained model and evaluate its prediction accuracy and generalization ability.

[0102] Conduct model testing to ensure the stability and reliability of the model in practical applications.

[0103] Step 6: Model deployment and application

[0104] Deploy the trained model to a production environment, which can be a cloud server or edge computing device.

[0105] Develop user interfaces and APIs so operators can easily input forecast data and receive forecast results.

[0106] Implement real-time monitoring and logging systems to facilitate model maintenance and troubleshooting.

[0107] Step 7: Model maintenance and continuous learning

[0108] Models are updated regularly to include the latest production data and trends.

[0109] Implement a continuous learning strategy to maintain the predictive power of the model through online learning or periodic retraining.

[0110] Based on user feedback and production results, adjust model parameters and similarity thresholds to optimize prediction results.

[0111] Model 2: Enhanced large-scale steel microstructure and performance control model based on reference cases

[0112] Step 1: Build an enhanced historical knowledge base

[0113] Collect and integrate more comprehensive data on steel initial conditions, and use data augmentation techniques such as SMOTE (Synthetic Minority Over-sampling Technique) to generate synthetic data to address the imbalance of samples in certain categories in the dataset. Generative adversarial networks (GANs) are also used to synthesize new steel cooling process data to increase data diversity for model training.

[0114] Step 2: Advanced Similarity Calculation

[0115] Use deep learning techniques such as autoencoders or variational autoencoders (VAEs) to learn low-dimensional representations of the data for similarity calculations.

[0116] Apply fuzzy logic or neural networks for similarity assessment to handle uncertainty and nonlinear relationships.

[0117] Step 3: Reference Case Selection

[0118] Threshold setting: Set a similarity threshold θ. Only when the similarity of historical data exceeds this threshold will it be considered as a reference case.

[0119] Case screening: Based on the similarity calculation results, the historical data that is most similar to the data to be predicted is screened out as a reference case.

[0120] Step 4: Dynamic Prompt Construction and Deep Learning Prediction

[0121] Reference case integration: Writes the initial conditions and tank parameters of the selected reference case to the prompt, along with the initial conditions of the data to be predicted.

[0122] Select a pre-trained large language model that performs well in related fields as the base model.

[0123] Model training: Use historical datasets to train large language models, including Prompt construction and parameter prediction.

[0124] Apply cross-validation and hyperparameter tuning techniques to optimize model performance.

[0125] Step 5: Prediction result verification and feedback mechanism

[0126] Develop a prediction result verification module to verify the effectiveness of prediction parameters through actual production data.

[0127] Establish a feedback mechanism to feed actual production results back to the model for continuous learning and optimization of the model.

[0128] Example 1: Cooling parameter control of a large language model based on a reference case

[0129] Workflow:

[0130] (1) Data collection and knowledge base construction:

[0131] A steel plant collected rolling data for three months, including 5000 sets of initial parameters (steel grade Q235B, C content 0.18%, rolling temperature 1150℃, rolling speed 12m / s) and corresponding water tank parameters (flow rate in zone 1 180m 3 / h, pressure 0.8MPa).

[0132] Data cleaning: Remove outliers (such as invalid records with temperatures greater than 1500°C) and use the KNN algorithm to fill in missing values ​​(such as the missing "refining power-on time").

[0133] Establish a historical knowledge base and store it by steel grade (Q235B, X70, etc.).

[0134] (2) Similarity matching and case screening:

[0135] New billet parameter input: steel grade Q235B (change C content to 0.20%), rolling temperature 1130℃, pulling speed 13m / s.

[0136] Normalization: Scale each parameter to the [0,1] interval (e.g., divide the temperature value by 1500).

[0137] Calculate the cosine similarity and select the top 5 historical cases (similarity > 0.85).

[0138] (3) Prompt construction and model reasoning:

[0139] Prompt template:

[0140] The water tank parameters for the following cases are known:

[0141] Case 1: C = 0.18%, rolling temperature 1150℃ → flow rate in zone 1 180m 3 / h, pressure 0.8MPa;

[0142] ...(Case 2-5)

[0143] Current working conditions: C = 0.20%, rolling temperature 1130℃ → Please generate recommended parameters.

[0144] Call the fine-tuned GPT-3.5 model to generate parameters: traffic flow in zone 1 is 185m 3 / h (+2.7%), pressure 0.78MPa.

[0145] (4) Deployment and verification:

[0146] The parameters were sent to the PLC control system, and the mechanical properties were tested after actual cooling: yield strength 345 MPa (expected value 340-350 MPa).

[0147] The feedback results are stored in the knowledge base, triggering incremental learning of the model (learning rate 0.001).

[0148] Results: The prediction error for the new operating conditions was reduced to 2.7%, the adjustment took 45 seconds (the traditional method took 15 minutes), and the scrap rate dropped from 4.5% to 3.8%.

[0149] Example 2: Application of enhanced models in data-scarce scenarios

[0150] Workflow:

[0151] (1) Data enhancement and knowledge base expansion:

[0152] Objective: A workshop added pipeline steel X80 production, with only 200 sets of initial data.

[0153] Use GANs to generate synthetic data: input the C and Mn composition distribution of X80 (C: 0.06-0.10%, Mn: 1.60-1.80%), and generate 1,000 sets of virtual parameters (such as "refining bottom blowing time" expanded from 15 minutes to 12-18 minutes).

[0154] SMOTE oversampling: For the minority class (samples with "superbuckling ratio > 1.25" account for 5%), synthesize to 20%.

[0155] (2) Depth similarity calculation:

[0156] Training variational autoencoder (VAE): input dimension 50 (initial parameters), hidden layer compressed to 10 dimensions.

[0157] New working condition input: X80 steel, C=0.08%, Mn=1.75%, starting rolling temperature 1100℃.

[0158] The Euclidean distance is calculated in the latent space to filter similar cases (threshold < 0.3).

[0159] (3) Fuzzy logic evaluation and dynamic prompt:

[0160] Define fuzzy rules:

[0161] If the difference in "C content" is greater than 0.03%, the similarity weight is reduced by 30%.

[0162] If the “pulling speed” difference is greater than 2m / s, the “flow compensation coefficient” adjustment is triggered.

[0163] Dynamic Prompt: Assign case priorities based on weight, and parameters of high-weight cases are displayed in a magnified manner.

[0164] (4) Verification and feedback:

[0165] Model recommendation: Zone 3 flow rate 220m 3 / h, pressure 1.2MPa.

[0166] After actual production, it was detected that the "pre-finishing rolling outlet temperature" deviated by 8°C, triggering the feedback mechanism:

[0167] Back-tune the VAE encoder (adding a temperature error term to the loss function);

[0168] Updated fuzzy rules (speed difference tolerance increased to 3m / s).

[0169] Results: The prediction error in data-scarce scenarios is 7.2% (traditional methods are >25%), with synthetic data contributing 40% of the total, and the model iteration cycle is shortened from 2 weeks to 3 days.

[0170] Example 3: Deployment of a multi-production line collaborative control system

[0171] Workflow:

[0172] (1) System module configuration:

[0173] Data collection module: collects real-time data (10 records per second) from 3 rolling lines through OPC protocol.

[0174] Data enhancement module: Deploy a GANs server to generate 500 pieces of synthetic data every day.

[0175] Fuzzy logic module: configure the rule base (such as "IF steel type is automobile plate THEN flow threshold = 200m 3 / h”).

[0176] (2) Knowledge transfer across production lines:

[0177] Production line A (mature production line): 100,000 historical data items, Model A accuracy rate of 94%.

[0178] Production line B (new production line): 3000 data points, initialize and load model A, and start transfer learning:

[0179] Freeze the Prompt encoding layer of model A and only fine-tune the similarity calculation layer;

[0180] After three days of training using data from production line B, the accuracy increased from 62% to 85%.

[0181] (3) Real-time control and exception handling:

[0182] Abnormal operating condition: The "water tank turbid water return temperature" of production line C suddenly rises by 10°C.

[0183] System response:

[0184] The similarity calculation module screens historical similar failure cases (recorded in May 2023);

[0185] The prompt generation module inserts an emergency instruction: "The current temperature is abnormal, please refer to the pressure reduction strategy of case #205 first";

[0186] Large language model output: Reduce the pressure in the four zones from 1.5MPa to 1.2MPa to avoid pipe bursting.

[0187] (4) Centralized feedback optimization:

[0188] Production feedback data from all production lines is fed into the central server, and batch training (BERT+Transformer structure) is performed once a day.

[0189] Model version management: Version V2.1 adds the impact factor of "mid-package weight" on the flow formula.

[0190] Results: The cold start training cycle of the new production line model is shortened by 80%, the response time to abnormal working conditions is less than 500ms, and the parameter migration error across production lines is less than 3%.

[0191] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for controlling steel microstructure and properties based on a large language model, characterized by: The following steps are involved: Collect the initial parameters of steel, including basic information parameters, chemical composition parameters, equipment parameters, temperature parameters, heating furnace parameters and process parameters; Establishing a historical knowledge base, wherein the historical knowledge base includes initial parameters of historical data and corresponding cooling process water tank parameters; Preprocess the data to be predicted and calculate the similarity between it and historical data; Filter reference cases based on similarity, and construct an instruction guide prompt input containing the reference cases and the data to be predicted; Use a large language model to train and predict the Prompt to generate target water tank parameters; Deploy the predicted water tank parameters to the production environment and optimize the model based on actual production feedback.

2. The steel structure and property control method based on a large language model according to claim 1, characterized in that: The similarity calculation adopts at least one algorithm of Euclidean distance and cosine similarity.

3. The steel structure and property control method based on a large language model according to claim 1, characterized in that: The large language model is a pre-trained language model based on the Transformer architecture, including a generative pre-trained transformer (GPT) or a general language model (GLM).

4. The steel structure and property control method based on a large language model according to claim 1, characterized in that: The Prompt build includes: Perform structured text splicing on the initial parameters of the reference case, the water tank parameters and the initial parameters of the data to be predicted; Add natural language instructions to guide the model to generate target tank parameters.

5. The steel structure and property control method based on a large language model according to claim 1 is characterized in that: The model optimization includes: Regularly update the historical knowledge base to incorporate new production data; Tune model parameters via online learning or incremental learning.

6. An enhanced large language model-based steel structure and property control method based on the method of claim 1, characterized in that: Further including: Use data augmentation technology to expand historical data and generate synthetic data; Use autoencoders or variational autoencoders to reduce the dimensionality of the data to optimize similarity calculation; Evaluate nonlinear similarity relationships through fuzzy logic or neural networks; Build a dynamic prompt to dynamically adjust the input content based on the weight of the reference case.

7. The enhanced steel structure and property control method based on a large language model according to claim 6 is characterized in that: The data enhancement technology includes synthetic minority over-sampling technology (Synthetic Minority Over-sampling Technique, SMOTE) or generative adversarial networks (Generative Adversarial Networks, GANs).

8. A steel structure and property control system based on a large language model, characterized by: include: A data collection module is configured to collect the initial parameters of the steel and the mechanical properties parameters after cooling from the steel rolling production line in real time, wherein the initial parameters include chemical composition, temperature parameters and process parameters, and transmit them to the preprocessing module via a data bus; A historical knowledge base module is connected to the preprocessing module and stores a historical data set after cleaning and normalization, wherein the data set includes a mapping relationship between initial parameters and corresponding water tank parameters; The preprocessing module receives the real-time data from the data collection module and the offline data from the historical knowledge base, performs normalization and missing value filling operations, and outputs the processed data to be predicted to the similarity calculation module; The similarity calculation module calls the historical data in the historical knowledge base, calculates the similarity matrix between the data to be predicted and the historical data based on the Euclidean distance or cosine similarity algorithm, selects the reference case set with a similarity higher than the threshold θ, and transmits the case set to the Prompt generation module; The prompt generation module receives the reference case set output by the similarity calculation module and the preprocessed data to be predicted, and splices the initial parameters, water tank parameters and current working condition parameters of the case into structured text according to a preset template to generate a prompt instruction stream that meets the input specifications of the large language model; The large language model inference module loads the pre-trained Transformer architecture language model, receives the instruction stream from the Prompt generation module, and outputs a set of predicted water tank parameters, including flow rate, pressure value and its confidence score; The deployment and feedback module communicates with the rolling line PLC control system, sends the predicted parameters to the cooling water tank actuator, collects actual production results in real time, calculates the error between the predicted value and the measured value, and feeds the error data back to the historical knowledge base module and the large language model inference module to trigger online fine-tuning of the model parameters.

9. The steel structure and property control system based on a large language model according to claim 8, characterized in that: The system further includes a data enhancement module connected between the historical knowledge base module and the preprocessing module, configured as follows: Receive raw data from a historical knowledge base and generate synthetic data using synthetic minority oversampling technology (SMOTE) or generative adversarial networks (GANs). The synthetic data meets preset metallurgical constraints, including the composition conservation law Ceq ≤ 0.50% and the temperature gradient ≤ 200°C / m. The synthetic data is mixed with the original data and then fed into the historical knowledge base module to expand the training data set, where the synthetic data accounts for no more than 30% of the total data volume; Trigger condition: When the amount of data for a certain steel grade in the historical knowledge base is less than 500, the data enhancement process is automatically started.

10. The steel structure and property control system based on a large language model according to claim 8, characterized in that: The similarity calculation module also includes: The fuzzy logic evaluation submodule is configured as follows: Define a fuzzy rule base, including a "component difference-weight mapping table"; perform fuzzy correction on the Euclidean distance calculation result and output the corrected similarity score S'=S×α, where S is the original similarity score and α is the fuzzy weight coefficient; Neural network evaluation submodule, configured as: Train a similarity network based on the attention mechanism, with the input being the difference vector between the data to be predicted and the historical data, and the output being a nonlinear similarity score; starting condition: when the data missing rate is greater than 15%, the neural network submodule is preferentially called for scoring.

Images