Method for optimizing control of feo content of sintered ore and storage medium

Abstract

The application discloses a sinter FeO content optimization control method and a storage medium, belongs to the field of metallurgical industry automation and data science, and through online detection instruments and sensors arranged on a sintering machine site, real-time working condition data of a sintering process is acquired; the working condition data is input into a pre-trained strategy model; a set of optimal process parameter combinations is output in real time according to the input working condition data, and is used for closed loop control of the sintering machine; the T-ccGAN model is improved by internalizing the Grid Search and the Cuckoo Search into model components, the problem of model selection and parameter adjustment of the traditional model on sparse industrial data is solved, the prediction accuracy of the SOTA is realized, and based on the Offline RL framework constructed based on the model, the optimal control strategy capable of guiding production can be trained only by using historical data.

Description

Technical Field

[0001] This invention belongs to the field of metallurgical industrial automation and data science, and specifically relates to a method for optimizing and controlling the FeO content of sintered ore and a storage medium. Background Technology

[0002] Sintering is a crucial step in the steel industry, providing essential raw materials for blast furnace production. The quality of sinter directly impacts blast furnace operations, with the FeO content being a key indicator of quality. Fluctuations in FeO content directly affect energy consumption in the sintering process and the blast furnace's ability to increase iron production while saving coke.

[0003] However, achieving stable control of FeO content is extremely difficult. The sintering process itself has complex characteristics such as large hysteresis, multiple variables, and nonlinearity, making it difficult for humans to detect and control it accurately in real time. Traditional methods are unable to mine the complex sparse relationship between process parameters and FeO content from massive historical data, resulting in inaccurate predictions.

[0004] Furthermore, once a high-precision predictive model is available, how can automatic optimization control be achieved? While traditional reinforcement learning (RL) is an effective means of solving control problems, it relies on "online trial and error," meaning the agent must continuously try in a real environment, even deliberately performing incorrect actions to learn. In the high-cost, high-risk industrial environment of sintering processes, any suboptimal "trial and error" can lead to energy waste, equipment damage, or defective products. Therefore, there is an urgent lack of an effective means to automatically learn and generate optimal process control strategies using only accumulated historical data without interfering with experimental production. Offline reinforcement learning (Offline RL) is a key technology designed to solve this problem. It aims to learn the optimal strategy from only a fixed, pre-collected dataset, providing a completely new technical path for achieving intelligent and safe optimization of industrial processes. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a complete two-stage industrial optimization technology solution. The core of this solution lies in its ability to safely obtain the optimal control strategy in high-risk industrial scenarios where online trial and error is prohibited.

[0006] The present invention adopts the following technical solution: A method for optimizing and controlling the FeO content of sintered ore involves acquiring real-time operating data of the sintering process using online monitoring instruments and sensors deployed at the sintering machine. This data is then input into a pre-trained strategy model, which outputs a set of optimal process parameters in real time based on the input data for closed-loop control of the sintering machine. The policy model is trained through the following steps: S1 sets up a state, action, and reward function for reinforcement learning based on historical production data of the sintering machine. The state consists of 21 parameters that are uncontrollable or slowly changing during production, including: TFe, CaO, MgO, SiO2, Al2O3, P, S, C, K2O, Na2O, Zn, Pb, burn-off rate, shrinkage rate, utilization coefficient, moisture, and five particle sizes. The action consists of 7 process parameters that can be adjusted in real time during production, including: mixture temperature, water addition rate, ignition negative pressure, sintering negative pressure, final temperature, sintering time, and vertical speed.

[0007] The reward function uses a variant of Huber Loss as the loss function, and its formal mathematical expression is as follows: The corresponding reward function is defined as the negative of the loss function: Where y pred This is the predicted FeO content from the T-ccGAN simulator, y target δ is the preset target FeO value, and δ is the preset error penalty threshold. S2 uses the Monte Carlo Dropout method to obtain the prediction uncertainty of the T-ccGAN environment simulator for state-action pairs; S3 Construct the T-Unc-CQL offline reinforcement learning objective function, and embed the prediction uncertainty obtained in step S2 as a penalty term into the loss function of the conservative Q learning algorithm to actively penalize the strategy's exploration in the low confidence region of the simulator, thereby avoiding the accumulation of model error; S4 uses only the static offline dataset and the T-Unc-CQL objective function defined in step S3 to train and solve the policy model.

[0008] Furthermore, the method for constructing the T-ccGAN environment simulator described in step S2 includes the following steps: (a) A TabNet feature engineering module with an embedded Grid Search optimization mechanism is used to adaptively reduce the dimensionality of the acquired multidimensional data and output a low-dimensional feature vector. (b) Construct a ccGAN prediction module with an embedded Cuckoo Search global optimization mechanism, namely the ccGAN prediction module with the CS mechanism. The fitness function of the CS mechanism is designed to be coupled with the dynamic adversarial game process of the ccGAN. The fitness function integrates the gradient statistics of the discriminator loss of the ccGAN and the stability index of the generator output. (c) Use the low-dimensional feature vector output in step (a) as the conditional input of the ccGAN module in step (b) to train the ccGAN module, establish a high-fidelity nonlinear mapping relationship between the low-dimensional feature vector and the FeO content of sinter, and obtain the T-ccGAN environment simulator.

[0009] Furthermore, the TabNet feature engineering module in step (a) internally generates a sparse feature mask through an attention transformer and a Sparsemax activation function to achieve the feature dimensionality reduction.

[0010] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for optimizing and controlling the FeO content of sintered ore based on digital twins and offline strategies.

[0011] The beneficial effects of this invention are: This invention seamlessly integrates high-precision prediction with intelligent decision-making. First, the T-ccGAN model solves the problem of model selection and parameter tuning on sparse industrial data by internalizing GridSearch and Cuckoo Search as model components, achieving state-of-the-art (SOTA) prediction accuracy. Second, the offline RL framework built based on this model can train an optimal control strategy to guide production using only historical data. This invention provides a complete technical solution for achieving intelligent, closed-loop, and optimized control of FeO content without relying on online experiments. Attached Figure Description

[0012] Figure 1 Overall route map; Figure 2 Predicted FeO content of sintered ore; Figure 3 Error comparison chart of FeO content in sintered ore. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with preferred embodiments.

[0014] Example 1

[0015] This invention provides a method for optimizing and controlling the FeO content in sintered ore, the overall roadmap of which is shown below. Figure 1 As shown. The implementation of this method mainly includes the following four stages: 1. Data Preparation and Data Preprocessing This stage involves cleaning and standardizing the historical production data from the sintering machine, laying the foundation for subsequent model training.

[0016] This embodiment selects 3,000 sets of historical data from actual production in steel enterprises, and selects 28 dimensions of data on sintered ore chemical composition, blending particle size distribution, and sintering process parameters. Specifically, these include blending temperature, water addition rate, four particle sizes of blending material (>8mm, 5-8mm, 3-5mm, <3mm), average particle size of blending material, ignition negative pressure, sintering negative pressure, final temperature, sintering time, burn-off rate, shrinkage rate, vertical velocity, utilization coefficient, moisture (residual moisture), and blending chemical composition (TFe, CaO, MgO, SiO2, Al2O3, P, S, C, K2O, Na2O, Zn, Pb).

[0017] The data preprocessing process is as follows: 1) Data cleaning: Inspect and delete invalid data rows containing missing or duplicate values; 2) Outlier removal: Box plots are used to detect data distribution and remove outliers that deviate from the normal range; 3) Standardization: The cleaned data is standardized using Z-score normalization to eliminate the influence of units. Finally, the processed data is divided into (s, a, r, s') tuples to form a static offline dataset.

[0018] 2. Construction of the T-ccGAN High-Precision Environment Simulator To accurately evaluate the effectiveness of a strategy in an offline environment and obtain the uncertainty of predictions, this embodiment first constructs a high-fidelity "digital twin" environment simulator—the T-ccGAN model.

[0019] 2.1 Construction of the Improved TabNet Feature Engineering Module This invention employs an improved TabNet module with an embedded Grid search optimization mechanism to perform adaptive feature dimensionality reduction on the preprocessed 28 columns of input features.

[0020] Sintering industry data is not traditional Gaussian distributed data, but rather exhibits "high-dimensional sparsity" characteristics, including multimodal distribution, strong nonlinearity, and irregular noise. Well-known dimensionality reduction methods in this field, such as principal component analysis (PCA) or autoencoders (AE), are ineffective when processing this type of data because they tend to learn the global distribution structure, thus losing the sparse key features that are crucial for prediction.

[0021] The present invention specifically chose TabNet because its internal "sequence multistep attention mechanism" and "sparse feature mask" are not obviously suitable for capturing sparse nonlinear relationships in such data.

[0022] Furthermore, this invention does not employ the well-known and commonly used TabNet, but instead embeds a Grid Search optimization mechanism within it. This is because TabNet's prediction performance is highly sensitive to the intermediate dimensions after dimensionality reduction. Those skilled in the art will understand that performing Grid Search optimization on TabNet's "intermediate dimensions" (i.e., sparsity) is not a conventional approach in existing technologies (the conventional approach is to directly fix this dimension or only optimize the network width / depth). This invention utilizes Grid Search to enhance TabNet's convergence stability on such sparse data, ensuring that it finds an optimal intermediate dimension (23 dimensions) with the "highest information density," which lays an indispensable foundation for the high-precision predictions of the subsequent ccGAN module.

[0023] The goal of this module is to find an optimal intermediate dimension d. * This dimension is determined by minimizing a 5-fold cross-validation objective function, whose formal mathematical expression is: d * It is the optimal intermediate dimension obtained through optimization, D search It is the candidate dimension search space defined by Grid Search, L CV This embodiment evaluates a multi-metric loss function based on validation trade-offs. It comprehensively evaluates MAE, MSE, and R. 2 By solving this optimization objective, this embodiment determines the optimal intermediate dimension d. * With a value of 23, the model achieved optimal predictive performance in this dimension, with specific evaluation metrics including: MSE decreasing to 0.0968, MAE decreasing to 0.2317, and R... 2 The value was improved to 0.8978. This result fully demonstrates that the embedded Grid Search mechanism can effectively determine the optimal feature dimensionality reduction space of TabNet, thus laying the foundation for high-precision prediction in subsequent modules.

[0024] To demonstrate the superiority of the TabNet structure selected in this invention, this embodiment compares the performance of TabNet with three dimensionality reduction methods (PCA and AE) under the same ccGAN prediction module. With the intermediate dimension uniformly set to 23, the experimental results show that the model using PCA for dimensionality reduction has a larger prediction error (MAE = 0.2015, R0). 2 The accuracy rate was 86.1%, mainly due to its tendency to lose key nonlinear features; the method using an autoencoder (AE) showed a slight improvement (MAE was 0.1866, R...). 2 While the accuracy is 88.3%, it tends to learn global features, thus ignoring the sparsity of the data. In contrast, the TabNet method used in this invention achieves the best prediction performance, with its MAE significantly reduced to 0.1016 and R... 2The accuracy was improved to 97.5%, successfully capturing sparse key features in the data. This comparative experiment fully demonstrates the superiority of the improved TabNet in dimensionality reduction operations for sparse key features.

[0025] 2.2 Construction of the Improved ccGAN Prediction Module This invention employs a conditional generative adversarial network (ccGAN) optimized using Cuckoo Search (CS) as its core prediction module. The core innovation of this module lies in its hyperparameter Θ. * It is obtained through a CS global optimization framework, and its formal mathematical expression is: Where Θ * Θ is the optimal combination of hyperparameters, Θ is the candidate solution of the CS algorithm in the preset search space Ω, and Fitness(Θ) is the fitness function of the CS algorithm, which represents the performance of building and evaluating the cGAN model using the hyperparameters Θ. In this embodiment, Θ includes: learning rate (lr), number of layers (num_layers), number of hidden units (hidden_units), learning rate decay (lr_decay), and mini-batch size (mini_batch).

[0026] This invention specifically designs the fitness function Fitness(Θ) as follows: MAE val V[L] is the mean absolute error of the validation set. D ] is the variance of the discriminator loss within a training batch (epoch) (for assessing stability), V[▽ θD L D ] is the variance of the discriminator gradient (assessing the intensity of the adversarial conflict), w i Preset weights.

[0027] This design shifts the optimization objective of CS from "finding the most accurate prediction" to "finding the most stable and most accurate prediction," thus solving the problem of unstable GAN training in known techniques. Inside the Fitness(Θ) function, the training objective of the evaluated cGAN model remains a standard minimum-maximum game. Where y is the conditional information (i.e., the 23-dimensional feature vector output in 2.1), and z is the noise variable.

[0028] To verify that the CS global optimization mechanism selected in this invention has an "unexpected technical effect" in solving the convergence stability of cGAN, this embodiment compares the CS algorithm with the genetic algorithm (GA) and particle swarm optimization (PSO) algorithms known in the art under the same hyperparameter search space Ω and the same fitness function Fitness(Θ) (i.e., using the MAE of the final cGAN model on the validation set as the fitness).

[0029] The objective of the experiment is to optimize the same combination of cGAN hyperparameters, sharing a common hyperparameter search space (i.e., cGAN model hyperparameters). To ensure that this comparative experiment meets the requirement of "sufficient disclosure" and to demonstrate its reproducibility, this invention discloses the following self-control parameters of the three optimization algorithms during optimization: 1) The cuckoo search (CS) of this invention has the following characteristics: population size (number of nests) of 5, maximum number of iterations of 10, nest abandonment probability of 0.25, and Lévy flight parameter of 1.5. 2) Comparative Genetic Algorithm (GA): Population size 40, number of iterations 20, crossover rate 0.8, mutation rate 0.1; 3) Comparison algorithm: Particle Swarm Optimization (PSO): 40 particles, 20 iterations, inertia weight of 0.729, cognitive factor of 1.494, and social factor of 1.494.

[0030] The evaluation metrics for the experiments included the final prediction accuracy (MAE) of cGAN on the test set and the stability of the training process (assessed by the variance of the training loss function; the smaller the variance, the more stable the model). Comparative results showed that when using the Genetic Algorithm (GA) for hyperparameter optimization, the model was prone to getting trapped in local optima and exhibited poor convergence, with an MAE of 0.1264 and a training stability variance of 0.0271. While the Particle Swarm Optimization (PSO) algorithm converged faster, it was prone to premature convergence, with an MAE of 0.1284 and a training stability variance of 0.0116. In contrast, the Cuckoo Search Algorithm (CS) used in this invention achieved the best performance, with its MAE significantly reduced to 0.1016 and its training stability variance reduced to 0.0058. Its internal Lévy flight mechanism successfully helped the model escape local optima and achieve stable convergence. This comparative experiment fully demonstrates that the Lévy flight mechanism of the CS algorithm has outstanding substantive features and significant progress in the hyperparameter optimization problem of dynamic adversarial games such as cGAN, and solves the convergence and stability problem that the well-known algorithm (GA / PSO) cannot solve.

[0031] 2.3 Solidification of the T-ccGAN Environment Simulator Combining steps 2.1 and 2.2, a complete T-ccGAN model is constructed, and the optimal hyperparameters obtained in step 2.2 are used for final training. During the optimization process, the search ranges for each hyperparameter are set as follows: learning rate (lr) from 0.0001 to 0.01, number of layers (num_layers) from 1 to 3, number of hidden units (hidden_units) from 16 to 64, learning rate decay (lr_decay) from 0 to 0.001, and mini-batch size (mini_batch) from 16 to 128. For the prediction index of FeO content in sintered ore, the optimal fitness obtained after global optimization is 0.0456, and the corresponding optimal hyperparameter combination is: lr = 0.0085, num_layers = 1.255, hidden_units = 35.069, lr_decay = 0.00079, and mini_batch = 32.791. It should be noted that since the optimization algorithm performs optimization in a continuous space, the floating-point results obtained need to be mapped to discrete integer values ​​that meet the requirements of the network structure through rounding or other rounding methods when actually applied to build a neural network model.

[0032] The trained model is the T-ccGAN environment simulator. To verify the prediction performance of the T-ccGAN environment simulator, this embodiment selected four machine learning and deep learning models that performed well in sintering process prediction as comparison models. The prediction results for FeO content in sintered ore are as follows: the XGboost model's prediction evaluation indexes are MAE=0.2131, MSE=0.1708, and R2=0.8569; the RandomForest model's evaluation indexes are MAE=0.2067, MSE=0.1215, and R2=0.9063; the LSTM model's evaluation indexes are MAE=0.2035, MSE=0.1501, and R2=0.8965; and the GA-RNN model's evaluation indexes are MAE=0.1755, MSE=0.0919, and R2=0.9217. Testing showed that the T-ccGAN simulator constructed in this invention exhibited excellent prediction accuracy on the test set, with all evaluation metrics reaching optimal levels: MAE=0.1016, MSE=0.0233, and R²=0.9754. The corresponding simulation results for the predicted FeO content in sinter and the error plot are as follows: Figure 2 and Figure 3 As shown.

[0033] This invention emphasizes that the high prediction accuracy is not the ultimate goal of this invention, but a necessary prerequisite for implementing the third stage. The extremely high data fit exhibited by the model proves its high fidelity as a "digital twin" environment, making it qualified to serve as a reliable and trustworthy environment simulator for subsequent training of reinforcement learning strategies under absolutely safe conditions.

[0034] 3. Training of Offline Policy Models Based on Uncertainty In this phase, we design and train an OfflineRL policy algorithm using the pre-built T-ccGAN environment simulator and offline dataset.

[0035] 3.1 Definition of Reinforcement Learning Elements This embodiment formalizes the optimization control problem of FeO content in sintered ore into a reinforcement learning problem, specifically defined as follows: 1) State Defined as 21 parameters that are uncontrollable or change slowly in production, specifically including: Five particle size distributions (>8mm, 5-8mm, 3-5mm, <3mm, average particle size), 12 mixed mineral chemical components (TFe, CaO, MgO, SiO2, Al2O3, P, S, C, K2O, Na2O, Zn, Pb), uncontrollable sintering process parameters (burn-off rate, shrinkage rate, utilization coefficient), and moisture.

[0036] 2) Action Defined as seven process parameters that can be adjusted in real time during production, specifically including: Mixture temperature, water addition rate, ignition negative pressure, sintering negative pressure, final temperature, sintering time, and vertical speed.

[0037] 3) Reward function To achieve stable control of the target FeO value (8.5% in this embodiment), a variant of Huber Loss is used as the loss function L. FeO This function applies a squared penalty to large errors and an absolute value penalty to small errors. Its formal mathematical expression is: The corresponding reward function is defined as the negative of the loss function: Where y pred It is the prediction of FeO by the T-ccGAN simulator (from step 2.3), where δ is a preset error penalty threshold, preferably δ=1 in this embodiment.

[0038] 3.2 Uncertainty Estimation of T-ccGAN This invention uses the Monte Carlo Dropout (MC Dropout) method to estimate the prediction uncertainty of the T-ccGAN simulator.

[0039] Specifically, during the training of the T-ccGAN model, Dropout is added to the neural network layer of its ccGAN prediction module. When calculating the reward and uncertainty for each pair in the offline dataset, this invention performs T forward propagations (T=20 in this embodiment) on the same (state s, action a) pair, and keeps the Dropout layer active during each forward propagation.

[0040] This will yield T different FeO prediction values ​​{y1, y2, …, y T This invention uses the variance of these T predicted values ​​as the prediction uncertainty u of the simulator for the (s, a) pair. T-ccGAN (s, a): .

[0041] 3.3 Objective Function of CQL with Uncertainty Coupling (T-Unc-CQL) In this embodiment, we implement the aforementioned "uncertainty coupling" strategy. Based on the loss function of the well-known Conservative Q-Learning (CQL) algorithm, this invention introduces the uncertainty as a penalty term to construct the T-Unc-CQL offline reinforcement learning objective function. The standard CQL loss function is: This invention modifies it into a T-Unc-CQL loss function for "uncertainty coupling": Among them, u T-ccGAN (s,a) represents the uncertainty of the data estimated by the T-ccGAN simulator in step 3.2. β is a hyperparameter (β=0.5 in this embodiment) used to adjust the intensity of the penalty.

[0042] Technical effect: By increasing β.E s~D [u T-ccGAN The invention achieves the following non-obvious implementation: when the T-ccGAN simulator has a high prediction uncertainty U for a certain (state s, action a) pair, the optimizer imposes a large penalty, forcing the Q function (Q(s, a)) to reduce its estimate of this "unreliable" (s, a) pair. This fundamentally solves the model bias problem in Offline RL, making the policy "unwilling" to explore regions where the simulator is uncertain, thereby ensuring the safety and stability of the policy.

[0043] 3.4 Offline Policy Model Training and Validation The core of this invention lies in constraining the Offline RL algorithm through the uncertainty of the T-ccGAN simulator to avoid model error accumulation. To demonstrate the non-obvious superiority of this "uncertainty coupling," this embodiment is divided into two groups for comparison: 1) Comparative Example: Training was performed using standard, unconstrained offline RL algorithms (CQL, IQL, BC); 2) Embodiment of the present invention: Training is performed using the aforementioned Offline RL algorithm (T-Unc-CQL) coupled with the uncertainty of T-ccGAN.

[0044] Algorithm training comparison results: In this comparative embodiment, the policy training does not use the uncertainty constraints of the present invention. We directly trained three well-known offline reinforcement learning algorithms (CQL, IQL, BC) on a static offline dataset.

[0045] The "Comparative Examples" and the "Examples of the Invention" (T-Unc-CQL) were trained and compared on a validation set. The comparison of the expected cumulative rewards of different offline reinforcement learning algorithms shows that the average reward for the Behavioral Cloning (BC) algorithm is -1.3611; the average reward for the Implicit Q-Learning (IQL) algorithm is -1.3609; and the average reward for the Conservative Q-Learning (CQL) algorithm is -1.2515. It can be seen that although the known CQL algorithm outperforms IQL and BC, its performance is still limited by the potential model error in the T-ccGAN simulator. In contrast, the T-Unc-CQL algorithm of this invention, by actively utilizing the uncertainty of T-ccGAN to constrain the Q function, achieves the highest average reward (-1.1032), significantly outperforming the standard CQL algorithm. This embodiment fully implements an algorithmic approach from high-precision predictive modeling (T-ccGAN) to prediction-based, uncertainty-aware offline optimization decision-making (T-Unc-CQL) through the above process. This approach does not require online trial and error and can provide optimal parameter recommendations for intelligent control of FeO content in sinter using only historical data.

[0046] 4. Online application of optimized control methods The trained strategy model is then deployed on the sintering production floor to achieve closed-loop optimization control. The specific application process is as follows: 1) Data acquisition: Real-time acquisition of sintering process condition data through online detection instruments and sensor networks deployed on the sintering machine; 2) State mapping: Converting the collected real-time data into a 21-dimensional "state" vector required by the algorithm; 3) Policy reasoning: Input the real-time state vector into the pre-trained T-Unc-CQL policy model; 4) Parameter Output and Control: Based on the current state, the model calculates and outputs a set of optimal "action" vectors (i.e., recommended values ​​for 7 process parameters) in real time. This parameter combination is directly used to guide the DCS system in adjustment, or provided as operational suggestions to personnel, thereby achieving precise and stable control of the FeO content in sintered ore.

[0047] Example 2 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for optimizing and controlling the FeO content of sintered ore based on digital twins and offline strategies.

[0048] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for optimizing and controlling the FeO content of sintered ore, comprising acquiring real-time operating data of the sintering process through online monitoring instruments and sensors deployed at the sintering machine site, inputting the operating data into a pre-trained strategy model, and outputting a set of optimal process parameter combinations in real time based on the input operating data for closed-loop control of the sintering machine; characterized in that, The policy model is trained through the following steps: S1 sets up the state, action, and reward functions for reinforcement learning based on the historical production data of the sintering machine; S2 uses the Monte Carlo Dropout method to obtain the prediction uncertainty of the T-ccGAN environment simulator for state-action pairs; S3 Construct the T-Unc-CQL offline reinforcement learning objective function, and embed the prediction uncertainty obtained in step S2 as a penalty term into the loss function of the conservative Q learning algorithm to actively penalize the strategy's exploration in the low confidence region of the simulator, thereby avoiding the accumulation of model error; S4 uses only the static offline dataset and the T-Unc-CQL objective function defined in step S3 to train and solve the policy model.

2. The method according to claim 1, characterized in that, The state described in step S1 refers to 21 parameters that are uncontrollable or slowly changing during production, including: TFe, CaO, MgO, SiO2, Al2O3, P, S, C, K2O, Na2O, Zn, Pb, burn-off rate, shrinkage rate, utilization coefficient, moisture, and five particle sizes.

3. The method according to claim 1, characterized in that, The actions refer to seven process parameters that can be adjusted in real time during production, including: mixture temperature, water addition rate, ignition negative pressure, sintering negative pressure, final temperature, sintering time, and vertical speed.

4. The method according to claim 1, characterized in that, The reward function described in step S1 uses a variant of Huber Loss as the loss function, and its formal mathematical expression is as follows: The corresponding reward function is defined as the negative of the loss function: Where y pred This is the predicted FeO content from the T-ccGAN simulator, y target δ is the preset target FeO value, and δ is the preset error penalty threshold.

5. The method according to claim 1, characterized in that, The method for constructing the T-ccGAN environment simulator described in step S2 includes the following steps: (a) A TabNet feature engineering module with an embedded Grid Search optimization mechanism is used to adaptively reduce the dimensionality of the acquired multidimensional data and output a low-dimensional feature vector. (b) Construct a ccGAN prediction module with an embedded Cuckoo Search global optimization mechanism, namely the CS mechanism ccGAN prediction module. The fitness function of the CS mechanism is designed to be coupled with the dynamic adversarial game process of the ccGAN. The fitness function integrates the gradient statistics of the discriminator loss of the ccGAN and the stability index of the generator output. (c) Use the low-dimensional feature vector output in step (a) as the conditional input of the ccGAN module in step (b) to train the ccGAN module, establish a high-fidelity nonlinear mapping relationship between the low-dimensional feature vector and the FeO content of sinter, and obtain the T-ccGAN environment simulator.

6. The method according to claim 5, characterized in that, The TabNet feature engineering module in step (a) internally generates a sparse feature mask through an attention transformer and a Sparsemax activation function to achieve feature dimensionality reduction.

7. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of any one of claims 1 to 6.

Images