An electric remelting endpoint element dynamic prediction and process decision optimization method and system

Abstract

The application provides an electric remelting smelting endpoint element dynamic prediction and process decision optimization method and system, the method comprises the following steps: obtaining historical data sets and real-time data streams in the smelting process, and preprocessing the data; introducing an online stochastic gradient descent and incremental learning algorithm, constructing an element dynamic prediction model, receiving sensor data in the smelting process in real time, and dynamically optimizing the element concentration prediction model; based on the deep deterministic policy gradient algorithm, an optimization decision model is constructed, the element composition, smelting conditions and other environmental parameters predicted by the element dynamic prediction model are used to adjust the smelting operation strategy in real time; a real-time feedback mechanism is set, so that the element dynamic prediction model and the optimization decision model work together to realize the real-time synchronous adjustment of the element concentration prediction and the control strategy. The application can adjust the smelting parameters in real time, improve the element prediction accuracy, reduce the energy consumption, and reduce the manual intervention, and significantly improve the smelting efficiency and quality stability.

Description

Technical Field

[0001] This invention relates to the field of metallurgical technology, and more particularly to a method and system for dynamic prediction of endpoint elements and optimization of process decisions in electroslag remelting smelting. Background Technology

[0002] Electroslag remelting (ESR) is an important metallurgical process widely used in the production of high-end steel and alloy materials. This process uses electric current to melt electrodes, causing metal to drip and solidify in a water-cooled crystallizer to form an ingot, thus achieving metal refining and shaping. In ESR, precise control of element concentration is crucial to ensuring the compositional uniformity and quality stability of the final product.

[0003] Currently, the prediction of element concentrations during electroslag remelting mainly relies on static mathematical models or machine learning models trained offline. Traditional thermodynamic prediction methods are primarily based on the molecular-ion coexistence theory (MICT) and the principle of material conservation, while kinetic-based modeling methods describe element transport and interfacial reaction processes through complex partial differential equations and numerical calculations. Furthermore, existing intelligent control systems typically employ fixed-rule control (such as PID control) or manual experience-based control, along with some models based on simple machine learning or deep learning, to control the smelting process.

[0004] However, these existing technologies have significant limitations. Static mathematical models and offline-trained machine learning models struggle to adapt to the complex multiphysics coupling effects and dynamic changes during smelting, resulting in low prediction accuracy. Traditional thermodynamic and kinetic models are deficient in terms of timeliness, accuracy, and computational complexity, making it difficult to meet the precise composition control requirements of electroslag remelting processes. Fixed-rule control and manual experience-based control lack adaptability to nonlinear dynamic changes and are difficult to standardize and quantify, limiting control accuracy and consistency in the production process. Furthermore, existing intelligent control systems lack self-learning and optimization capabilities, failing to adapt to dynamic changes in the smelting process in real time, requiring significant resource investment for model updates and maintenance. Reinforcement learning and online learning methods also face challenges in practical applications, including low sample efficiency, long training times, high costs, and dependence on environmental disturbances. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for dynamic prediction of elemental endpoints and optimization of process decisions in electroslag remelting smelting. This invention enables real-time adjustment of smelting parameters, improves elemental prediction accuracy, reduces energy consumption, minimizes manual intervention, and significantly enhances smelting efficiency and quality stability.

[0006] The technical means employed in this invention are as follows:

[0007] A method for dynamic prediction of endpoint elements and optimization of process decisions in electroslag remelting smelting includes:

[0008] S1. Acquire historical datasets and real-time data streams from the smelting process, and preprocess the acquired data;

[0009] S2. Introduce Online Stochastic Gradient Descent (Online SGD) and incremental learning algorithms to construct a dynamic element prediction model, receive sensor data during the smelting process in real time, and dynamically optimize the element concentration prediction model.

[0010] S3. Based on the Deep Deterministic Policy Gradient (DDPG) algorithm, an optimization decision model is constructed, and the smelting operation strategy is adjusted in real time according to the element composition, smelting conditions and other environmental parameters predicted online by the element dynamic forecast model.

[0011] S4. Set up a real-time feedback mechanism to enable the element dynamic forecasting model and the optimization decision model to work together, realize the real-time synchronous adjustment of element concentration prediction and control strategy, and ensure the accuracy of element concentration prediction and stable control of the smelting process.

[0012] Further, step S1 specifically includes:

[0013] S11. Clean the data to remove invalid, abnormal, or erroneous data and ensure the quality of the input data.

[0014] S12. Normalize the data by using normalization methods to standardize or scale the data to maintain a uniform scale and ensure the stability of model training and prediction accuracy.

[0015] S13. Time series processing: Using the sliding window method, a fixed-length historical data sequence is input each time, enabling the model to use past information for real-time prediction and optimization.

[0016] S14. Feature extraction: Extract key features from the raw smelting data, including equipment operating status, raw material ratio, process smelting parameters, and environmental data.

[0017] S15. Store the processed data in a relational database and a time-series database to achieve efficient data management.

[0018] Furthermore, in step S12, a data standardization process is adopted, and the specific formula is as follows:

[0019]

[0020] Where x is the original value, μ is the characteristic mean, and σ is the characteristic standard deviation.

[0021] Further, step S2 specifically includes:

[0022] S21. When new data arrives, the online stochastic gradient descent algorithm is used for local updates to gradually adjust the parameters of the element concentration prediction model in order to minimize the prediction error.

[0023] S22. The parameters of the element concentration prediction model are periodically updated using an incremental learning algorithm, so that the element concentration prediction model adapts to the accumulation of data streams and ensures a high degree of consistency with the actual data of the current smelting environment.

[0024] S23. By comparing the prediction errors of the online stochastic gradient descent algorithm and the incremental learning algorithm through the error feedback mechanism, the element concentration prediction model with the smallest error is automatically selected.

[0025] Furthermore, the underlying layer of the element dynamic prediction model is a multi-layer feedforward neural network, consisting of an input layer, multiple hidden layers, and an output layer. The key parameters of the model adjusted by the coupling strategy of the online stochastic gradient descent (SGD) algorithm and the incremental learning algorithm include the learning rate, the number of neurons in the hidden layer, and the selection of the activation function.

[0026] Furthermore, the element concentration prediction model supports online AdaGrad or online Adam optimization algorithms to adaptively adjust the learning rate and accelerate model convergence. At the same time, it combines a sliding window or forgetting mechanism to avoid interference from old data on the element concentration prediction model, allowing the element concentration prediction model to focus on the latest data stream, thereby improving prediction accuracy and real-time response capability.

[0027] Further, step S3 specifically includes:

[0028] S31. Based on the Deep Deterministic Policy Gradient (DDPG) algorithm, construct an optimization decision model, including a policy network (Actor) and a value network (Critic). The policy network (Actor) generates continuous control actions (such as melting rate and current) according to the current smelting state. The value network (Critic) evaluates the impact of state-action on model performance and calculates the Q value to optimize smelting operation parameters.

[0029] S32. Define the state space (smelting state parameters) and the action space (continuous operation parameters such as melting rate, current, and voltage);

[0030] S33. Design a reward function that comprehensively considers element content error, energy consumption, and efficiency, as follows:

[0031]

[0032] Among them, R tα1 is the value of the reward function, representing the reward obtained by the agent at time t; the reward value is used to guide the agent to optimize its policy; α1, α2, and α3 are all weighting coefficients, used to balance the influence of different parts of the reward function, and determine the relative importance of target element content error, smelting energy consumption, and production efficiency in the reward; C pred,i C represents the predicted content of the i-th element, i.e., the element content predicted by the agent based on the current smelting state; target,i This represents the target content of the i-th element, i.e., the element content target that the smelting process hopes to achieve.

[0033] S34. Utilize experience replay and target network mechanisms to train the model, enabling dynamic adjustment of smelting parameters. This improves the stability and convergence of the training process.

[0034] Furthermore, in step S33, the designed reward function specifically includes:

[0035] When the content of the endpoint element is close to the target value, a positive reward is given;

[0036] When smelting energy consumption is too high or efficiency is reduced, a negative reward is given. This guides the algorithm to gradually adjust smelting operations and optimize the smelting process.

[0037] This invention also provides a system for dynamic prediction of endpoint elements and optimization of process decisions in electroslag remelting based on the aforementioned method, comprising: a data preprocessing module, an online learning module, a reinforcement learning module, and a synergistic effect module, wherein:

[0038] The data preprocessing module is used to acquire historical datasets and real-time data streams from the smelting process, and to preprocess the acquired data.

[0039] The online learning module introduces online stochastic gradient descent and incremental learning algorithms to construct a dynamic element prediction model, receive sensor data during the smelting process in real time, and dynamically optimize the element concentration prediction model.

[0040] The reinforcement learning module, based on the deep deterministic policy gradient algorithm, constructs an optimization decision model and adjusts the smelting operation strategy in real time according to the element composition, smelting conditions and other environmental parameters predicted online by the element dynamic prediction model.

[0041] The synergistic module is equipped with a real-time feedback mechanism, enabling the dynamic element forecasting model and the optimization decision-making model to work together to achieve real-time synchronous adjustment of element concentration prediction and control strategies, thereby ensuring the accuracy of element concentration prediction and stable control of the smelting process.

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] 1. This invention utilizes online stochastic gradient descent (SGD) and incremental learning algorithms to receive sensor data during the smelting process in real time and dynamically optimize the element concentration prediction model. This significantly improves the accuracy and timeliness of element concentration prediction while reducing computational costs. Through error feedback mechanisms and optimization algorithms (such as online AdaGrad and Adam), the model can continuously and adaptively adjust, ensuring the stability and accuracy of the smelting process. Furthermore, by combining a sliding window or forgetting mechanism, interference from old data is avoided, allowing the model to focus more on the latest data stream, thereby improving prediction accuracy and real-time responsiveness. Through these technical means, this invention can achieve high-precision dynamic prediction of element concentration during the smelting process, providing accurate data support for subsequent intelligent control.

[0044] 2. This invention utilizes the Deep Deterministic Policy Gradient (DDPG) algorithm to adjust smelting operation strategies, such as melting rate, current, and electrode position, in real time based on the elemental composition, smelting conditions, and other environmental parameters predicted online by the elemental dynamic prediction model. The DDPG algorithm can efficiently handle continuous action spaces, ensuring precise control of the smelting process under complex conditions. The combination of a policy network (Actor) and a value network (Critic), along with experience replay and a target network mechanism, improves the stability and convergence of the training process. The reward function comprehensively considers the target element content error, smelting energy consumption, and efficiency improvement, guiding the algorithm to optimize the operation strategy. Through these techniques, this invention achieves precise control of the smelting process, optimizes operating parameters, improves production efficiency, reduces energy consumption, minimizes resource waste, and promotes the development of smelting technology towards energy conservation and environmental protection.

[0045] 3. This invention, through a real-time feedback mechanism, enables the online learning algorithm and DDPG policy network to continuously self-optimize, ensuring efficient operation and precise control of the smelting process under various dynamic environments. This collaborative optimization closed-loop mechanism makes the smelting process not only more efficient but also possesses strong adaptability and stability, greatly enhancing the intelligence and self-regulation capabilities of the entire process.

[0046] Based on the above reasons, the method of this invention is not only applicable to electroslag remelting processes, but can also be extended to other metallurgical smelting processes, such as blast furnaces, electric furnaces, vacuum induction furnaces, and vacuum arc remelting furnaces. Only appropriate adjustments to the model and optimization of input and output parameters are needed to adapt to the characteristics of different smelting processes. Through precise parameter adjustment and intelligent control strategies, efficient control of the smelting process can be achieved, ensuring maximum utilization of resources during production while reducing environmental burden. Precise control strategies can significantly reduce resource waste during smelting, promoting the development of smelting processes towards energy conservation and environmental protection. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of the method of the present invention.

[0049] Figure 2 This is a flowchart of the system operation of the present invention. Detailed Implementation

[0050] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0051] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0052] like Figure 1 As shown, this invention provides a method for dynamic prediction of endpoint elements and optimization of process decisions in electroslag remelting smelting, including:

[0053] S1. Acquire historical datasets and real-time data streams from the smelting process, and preprocess the acquired data;

[0054] S2. Introduce Online Stochastic Gradient Descent (Online SGD) and incremental learning algorithms to construct a dynamic element prediction model, receive sensor data during the smelting process in real time, and dynamically optimize the element concentration prediction model.

[0055] S3. Based on the Deep Deterministic Policy Gradient (DDPG) algorithm, an optimization decision model is constructed, and the smelting operation strategy is adjusted in real time according to the element composition, smelting conditions and other environmental parameters predicted online by the element dynamic forecast model.

[0056] S4. Set up a real-time feedback mechanism to enable the element dynamic forecasting model and the optimization decision model to work together, realize the real-time synchronous adjustment of element concentration prediction and control strategy, and ensure the accuracy of element concentration prediction and stable control of the smelting process.

[0057] In a specific implementation, as a preferred embodiment of the present invention, step S1 specifically includes:

[0058] S11. Clean the data to remove invalid, abnormal, or erroneous data and ensure the quality of the input data.

[0059] In this embodiment, the specific steps include: for historical data, fields with few missing values ​​are filled using the mean, median, or interpolation method; if too much data is missing and cannot be reasonably filled, that part of the data is removed; for abnormal data, the 3σ principle and Isolation Forest statistical method are used to detect and process abnormal data; for real-time data streams, since the API may repeatedly collect data with the same timestamp, deduplication is required to ensure that the record at each time point is unique. Furthermore, an Isolation Forest algorithm trained based on historical data is used for abnormal data alerts and processing.

[0060] S12. Normalize the data by using normalization methods to standardize or scale the data to maintain a uniform scale and ensure the stability of model training and prediction accuracy.

[0061] In this embodiment, a data standardization process is adopted, and the specific formula is as follows:

[0062]

[0063] Where x is the original value, μ is the feature mean, and σ is the feature standard deviation. Normalized data can effectively avoid the impact of numerical scale imbalance on the model learning process, improving training efficiency and prediction accuracy.

[0064] S13. Time series processing: Using the sliding window method, a fixed-length historical data sequence is input each time, enabling the model to use past information for real-time prediction and optimization.

[0065] In this embodiment, since the smelting data is time-series data, and data at different points in time exhibit time dependencies, it is necessary to maintain the temporal consistency of the data during data preprocessing. Sliding window processing is a crucial part of this method. Its window size and step size need to be pre-set based on the sensor characteristics and sampling measurement conditions of the target device to ensure that the data can effectively capture the dynamic changes in the smelting process. Through the sliding window method, a fixed-length historical data sequence is input each time, enabling the model to utilize past information for real-time prediction and optimization. Furthermore, time-series processing also includes timestamp alignment and feature extraction (such as mean, trend, and periodicity analysis) steps to enhance the model's predictive ability and adapt to the continuous changes in the smelting process.

[0066] S14. Feature extraction: Extract key features from the raw smelting data, including equipment operating status, raw material ratio, process smelting parameters, and environmental data.

[0067] In this embodiment, to improve the model's predictive accuracy and generalization ability, key features need to be extracted from the original smelting data. These features mainly include equipment operating status, raw material and auxiliary material ratios, process smelting parameters, and environmental data. Equipment features encompass core parameters such as current, voltage, and melting rate; raw material and auxiliary material features involve the feeding ratio and chemical composition of key materials; process features include furnace temperature changes and slag layer thickness; and environmental features consider external influencing factors such as humidity and airflow. Parameters that cannot be directly or indirectly measured, such as furnace temperature changes, can be calculated using kinetic or other mechanistic models. Furthermore, to reduce redundant data and optimize computational efficiency, filtering methods (Spearman correlation coefficient analysis), embedding methods (principal component analysis, Lasso regression, etc.), and packaging methods (recursive feature elimination, etc.) can be used to select the most representative features. Simultaneously, algorithms such as PCA and KPCA can be used for feature value dimensionality reduction. This ensures that the model can more accurately learn the key patterns of the smelting process, achieving intelligent optimization and dynamic adjustment of the smelting state.

[0068] S15. Store the processed data in a relational database and a time-series database to achieve efficient data management.

[0069] In this embodiment, to efficiently manage these data, a relational database is used to store historical data, a time-series database is used to process real-time data streams, and distributed storage and caching mechanisms are combined to ensure efficient data access and real-time processing. Furthermore, a data stream management platform enables real-time monitoring, utilizing stream processing technology to promptly capture changes during the smelting process. Finally, database deployment must prioritize data security and access control, employing encryption, access control, and security auditing measures to ensure data security and privacy protection, thereby providing a solid data foundation for the dynamic optimization of the smelting process.

[0070] In a specific implementation, as a preferred embodiment of the present invention, step S2 specifically includes:

[0071] S21. When new data arrives, the online stochastic gradient descent (Online SGD) algorithm is used for local updates to gradually adjust the parameters of the element concentration prediction model in order to minimize the prediction error.

[0072] In this embodiment, the Online Stochastic Gradient Descent (SGD) algorithm minimizes prediction error by progressively adjusting model parameters and performing local updates each time new data arrives, avoiding the high computational cost of training with full data sets. This enables the model to keep pace with changes in the data stream in real time and reflect the actual situation during the smelting process promptly.

[0073] S22. The parameters of the element concentration prediction model are periodically updated using an incremental learning algorithm, so that the element concentration prediction model adapts to the accumulation of data streams and ensures a high degree of consistency with the actual data of the current smelting environment.

[0074] In this embodiment, the incremental learning algorithm updates the model parameters periodically, enabling the model to automatically adjust during data stream accumulation. This ensures that the model remains highly consistent with the actual data of the smelting environment at all times. This approach avoids frequent retraining during model updates, improving the system's real-time performance and adaptability.

[0075] S23. By comparing the prediction errors of the online stochastic gradient descent algorithm and the incremental learning algorithm through the error feedback mechanism, the element concentration prediction model with the smallest error is automatically selected.

[0076] In this embodiment, to further improve the model's accuracy and stability, an error feedback mechanism is established. This mechanism automatically selects the model with the smallest error by comparing the prediction results of online SGD and incremental learning strategies with the actual values ​​in real time. Specifically, the error feedback mechanism compares the prediction errors of the two algorithms and adjusts the model parameters according to the optimal algorithm. If the adjustment result of online SGD is better than incremental learning, the SGD optimization strategy is applied first, and vice versa, ensuring the model's accuracy in dynamic environments. This process continuously corrects the model weights through error feedback, thereby improving the model's real-time performance and accuracy.

[0077] In a specific implementation, as a preferred embodiment of the present invention, the underlying layer of the element dynamic prediction model is a multi-layer feedforward neural network, consisting of an input layer, multiple hidden layers, and an output layer. The key parameters of the model adjusted by the coupling strategy of the Online Stochastic Gradient Descent (OnlineSGD) algorithm and the incremental learning algorithm include the learning rate, the number of neurons in the hidden layer, and the selection of the activation function.

[0078] In a specific implementation, as a preferred embodiment of the present invention, the element concentration prediction model supports online AdaGrad or online Adam optimization algorithms to adaptively adjust the learning rate and accelerate model convergence. At the same time, it combines a sliding window or forgetting mechanism to avoid interference from old data on the element concentration prediction model, so that the element concentration prediction model focuses on the latest data stream, thereby improving prediction accuracy and real-time response capability.

[0079] In a specific implementation, as a preferred embodiment of the present invention, step S3 specifically includes:

[0080] S31. Based on the Deep Deterministic Policy Gradient (DDPG) algorithm, construct an optimization decision model, including a policy network (Actor) and a value network (Critic). The policy network (Actor) generates continuous control actions (such as melting rate and current) according to the current smelting state. The value network (Critic) evaluates the impact of state-action on model performance and calculates the Q value to optimize smelting operation parameters.

[0081] In this embodiment, DDPG, as a continuous control reinforcement learning algorithm based on the Actor-Critic structure, is more suitable for high-dimensional, continuous control problems, such as the optimization and adjustment of key smelting parameters like melting rate and current. DDPG can output a smooth and differentiable action space adjustment strategy, making the optimization of smelting operations more stable.

[0082] S32. Define the state space (smelting state parameters) and the action space (continuous operation parameters such as melting rate, current, and voltage);

[0083] In this embodiment, the state space defines the smelting state parameters at a certain moment in the smelting process. These parameters reflect the specific situation of the smelting process, including physical quantities collected by sensors, historical smelting states, environmental parameters, and auxiliary material parameters. These parameters typically do not change easily during the smelting process, therefore, as a dimension of the state space, they can relatively stably reflect the current smelting status. The action space consists of the operational parameters that the intelligent control system needs to adjust according to the current smelting state. It belongs to the continuous action space and includes parameters such as melting rate, voltage, current, cooling system temperature control, and electrode parameters.

[0084] S33. Design a reward function that comprehensively considers element content error, energy consumption, and efficiency, as follows:

[0085]

[0086] Among them, R tα1 is the value of the reward function, representing the reward obtained by the agent at time t; the reward value is used to guide the agent to optimize its policy; α1, α2, and α3 are all weighting coefficients, used to balance the influence of different parts of the reward function, and determine the relative importance of target element content error, smelting energy consumption, and production efficiency in the reward; C pred,i C represents the predicted content of the i-th element, i.e., the element content predicted by the agent based on the current smelting state; target,i This represents the target content of the i-th element, i.e., the element content target that the smelting process hopes to achieve.

[0087] In this embodiment, the reward function guides the agent to choose appropriate actions during the learning process to achieve the optimization goal. In this model, the reward function needs to ensure that the content of the endpoint element remains within the target range, while also minimizing smelting energy consumption and improving smelting efficiency. Therefore, a reward function that comprehensively considers the target element content error, smelting energy consumption, and efficiency improvement is designed.

[0088] S34. Utilize experience replay and target network mechanisms to train the model, enabling dynamic adjustment of smelting parameters. This improves the stability and convergence of the training process.

[0089] In summary, during the training process of the DDPG model, the agent continuously optimizes the policy by maximizing long-term rewards, gradually learning how to adjust smelting operation parameters to ensure that the content of the endpoint element remains within a reasonable range. To improve training efficiency and stability, an experience replay pool is used to store state-action-reward triplets, and a priority experience replay mechanism is combined to select experiences that contribute more to policy optimization for learning. Finally, optimization suggestions are made to the action space to achieve process optimization. In this way, DDPG can adjust operation parameters in real time during the smelting process, ensuring precise control of the endpoint element content, while optimizing the energy efficiency and production efficiency of the smelting process, achieving intelligent optimization and precise control of the smelting process.

[0090] In a specific implementation, as a preferred embodiment of the present invention, the reward function designed in step S33 specifically includes:

[0091] When the content of the endpoint element is close to the target value, a positive reward is given;

[0092] When smelting energy consumption is too high or efficiency is reduced, a negative reward is given. This guides the algorithm to gradually adjust smelting operations and optimize the smelting process.

[0093] Corresponding to the method for dynamic prediction of endpoint elements and process decision optimization in electroslag remelting smelting in this application, this application also provides a system for dynamic prediction of endpoint elements and process decision optimization in electroslag remelting smelting, such as... Figure 2As shown, it includes: a data preprocessing module, an online learning module, a reinforcement learning module, and a synergy module, wherein:

[0094] The data preprocessing module is used to acquire historical datasets and real-time data streams from the smelting process, and to preprocess the acquired data.

[0095] The online learning module introduces online stochastic gradient descent (Online SGD) and incremental learning algorithms to construct a dynamic element prediction model, receive sensor data during the smelting process in real time, and dynamically optimize the element concentration prediction model.

[0096] The reinforcement learning module, based on the Deep Deterministic Policy Gradient (DDPG) algorithm, constructs an optimization decision model and adjusts the smelting operation strategy in real time according to the element composition, smelting conditions and other environmental parameters predicted online by the element dynamic prediction model.

[0097] The synergistic module is equipped with a real-time feedback mechanism, enabling the dynamic element forecasting model and the optimization decision-making model to work together to achieve real-time synchronous adjustment of element concentration prediction and control strategies, thereby ensuring the accuracy of element concentration prediction and stable control of the smelting process.

[0098] The embodiments of the present invention are described simply because they correspond to those in the embodiments above. For any similarities, please refer to the descriptions in the embodiments above, which will not be elaborated here.

[0099] Example

[0100] In the verification phase of this embodiment, nearly one year's worth of data from a certain electroslag remelting furnace at a special steel plant in China was used to construct a dynamic prediction and optimization decision-making method for the endpoint titanium content, as detailed below:

[0101] I. Data Acquisition and Preprocessing

[0102] First, data is acquired based on the production database provided by the factory. This data includes equipment parameters (crystallizer size, etc.), slag parameters (slag quantity, slag composition, etc.), types and amounts of alloy additives, environmental parameters (atmospheric humidity, atmospheric temperature, etc.), cooling system parameters (inlet water temperature, outlet water temperature), smelting parameters during electroslag remelting (current, voltage, melting rate, protective gas state parameters, etc.), electrode state parameters (electrode size, shape, and state, etc.), ingot parameters (composition content, size, shape, final ingot temperature, etc.), and element content at the ingot's final casting point.

[0103] Data cleaning then proceeded, comprising two parts: historical data cleaning and real-time data cleaning. In historical data cleaning, fields with few missing values ​​were imputed using the mean, while fields with many missing values ​​that could not be adequately imputed were removed. The 3σ principle and the Isolation Forest algorithm were also used to handle outliers. In real-time data cleaning, timestamp deduplication ensured the uniqueness of each record, and an Isolation Forest model trained on historical data was used for anomaly detection and alerts, thus guaranteeing data accuracy and consistency.

[0104] Then, all data are normalized, and real-time data is processed to be time-series data. To further improve model performance and reduce costs, mutual information is used for model selection. The specific principle and formula are as follows:

[0105]

[0106] The final selected input variables are 21, including ingot weight, slag quantity, initial titanium, tungsten, copper, aluminum, ferrosilicon powder, silicon, carbon, aluminum powder, sulfur, carbon powder, manganese, oxygen, phosphorus, inlet water temperature, outlet water temperature, ambient humidity, ambient temperature, real-time current and voltage, and melting rate. The output variable is the final titanium content.

[0107] II. Building an Online Learning Module

[0108] In the online learning module, an 8-layer deep neural network (DNN) was constructed to predict the dynamic changes in element concentration during the smelting process in real time. This model combines online stochastic gradient descent (SGD) and incremental learning methods, and introduces an error feedback mechanism. Based on the model's prediction error, it dynamically selects between online learning and incremental learning for updates, ensuring the model can adapt to real-time data changes during the smelting process. Furthermore, the Adam adaptive optimization algorithm is introduced to dynamically adjust the learning rate to further optimize hyperparameters and improve the model's convergence speed and generalization ability.

[0109] The online learning module employs an 8-layer deep neural network (DNN) whose input features encompass smelting parameters such as current, voltage, melting rate, furnace temperature, slag thickness, and key element concentrations. Six hidden layers utilize the ReLU activation function, with layer numbers sequentially set to 128-256-512-512-256-128, ultimately outputting predicted element concentration values. During training, the mean squared error (MSE) loss function is used, and a sliding window mechanism (10-minute window length, 1-minute step) is employed for data sampling to ensure the model can capture dynamic changes in smelting conditions.

[0110] To optimize the model, an error feedback mechanism was designed. Using data nodes from every 30 heats as a baseline, the prediction error (MSE) of both online SGD and incremental learning strategies was calculated, and the model with the smallest error was automatically selected for application. Specifically, the system periodically uses the error feedback mechanism to compare the prediction errors of the two algorithms and adjusts the model parameters based on the optimal algorithm. If the adjustment result of online SGD is better than incremental learning (lower MSE), the SGD optimization strategy is applied first for rapid updates; if the prediction error of incremental learning is lower, the incremental learning strategy is used for stable updates. This error feedback process continuously adjusts the model weights to ensure the model's accuracy in dynamic environments, improving real-time performance and adaptability. Furthermore, the Adam optimization algorithm was introduced, combined with momentum optimization and RMSprop, to adaptively adjust the learning rate, improving the model's convergence speed and stability. Adam's dynamic adjustment mechanism effectively avoids the problem of excessively large or small learning rates when the smelting data stream changes significantly, ensuring efficient model updates.

[0111] To verify the effectiveness of the model, it was compared with traditional machine learning methods (random forest, support vector machine, XGBoost), methods using neural networks alone (without online / incremental learning), and methods using incremental learning and online learning respectively on top of neural networks. Experimental results show that:

[0112] Traditional machine learning methods perform well on static data, but their generalization ability is weak in dynamic smelting environments. The MSE is too high, reaching a maximum of 53.21 and a minimum of 0.21. As data streams are input, a single machine learning model cannot update autonomously, and it cannot make predictions to guide production smelting.

[0113] When using only neural networks (without online / incremental learning), the model struggles to adapt to real-time data changes, resulting in large prediction errors. As data flows in, the MSE gradually increases from small to large, and the fluctuations are highly uncertain, with the maximum MSE reaching 12.231.

[0114] When incremental learning and online learning are used alone, both strategies reduce the model's error, but the error range is between 0.02 and 10.2. By combining online learning and incremental learning and adopting an error feedback mechanism, the model can dynamically select the optimal update strategy, keeping the overall MSE value within 0.02 and significantly improving prediction accuracy.

[0115] Experimental results show that this method can effectively reduce prediction error (MSE from 0.02 to below 0.008) and improve the model's adaptability and stability through error feedback mechanism and dynamic model update strategy. During the smelting process, the model achieves more accurate element concentration prediction and control, meeting process requirements.

[0116] III. Building a Reinforcement Learning Module

[0117] Based on the online learning module, a Deep Deterministic Policy Gradient (DDPG) reinforcement learning algorithm was built for real-time optimization of smelting operation strategies. DDPG adopts an Actor-Critic structure, where the Actor network adjusts operation parameters (such as electrode position and charging rate) in real time based on online predicted elemental composition, smelting conditions, and environmental parameters; while the Critic network evaluates the merits of the operation strategy through a reward function and continuously optimizes the Actor network through a feedback mechanism.

[0118] The state space is defined with seven variables: inlet water temperature, outlet water temperature, ambient humidity, ambient temperature, real-time current, voltage, and melting rate, thus constructing a seven-dimensional state space. Based on this state space, a three-dimensional action space is defined to optimize operating parameters according to state changes. Specifically, the action space includes melting rate, current, and voltage, and is a continuous three-dimensional space.

[0119] The Actor network has an input dimension of 7 (corresponding to 7 state variables) and an output dimension of 3 (corresponding to 3 actions: melting rate, current, and voltage). It uses a 9-layer fully connected neural network (FCN) with 256, 128, and 64 neurons in each layer (repeated layers), uses ReLU as the activation function, and adds an L2 regularization layer to reduce overfitting.

[0120] The Critic network takes as input a concatenation of state and action variables, with an input dimension of 10, and outputs a Q-value (scalar). It uses a 5-layer fully connected neural network (FCN), with each layer containing 256, 128, 128, 64, and 64 neurons respectively. The activation function is ReLU, and an L2 regularization layer is also included.

[0121] Hyperparameter settings:

[0122] Learning rate: Using the Adam optimizer, the learning rate for both the Actor and Critic networks is set to 0.001, β1 = 0.9, and β2 = 0.999. Experience replay pool size: 10000. Batch size: 64 samples are randomly sampled for each update. Target network update frequency: The target network is updated every 50 steps. Discount factor (gamma): 0.99. Exploration rate: Initial exploration rate of 0.1, gradually decreasing during training.

[0123] Reward function design:

[0124] The reward function incorporates multiple indicators from the smelting process, aiming to minimize deviations in titanium content and improve electrical efficiency. The specific design is as follows:

[0125] Composition error: The reward is calculated based on the error between the final titanium content and the target value during the smelting process. The smaller the error, the higher the reward.

[0126] The target concentration is set within a tolerance range of ±1%, and negative rewards are given when the concentration is exceeded.

[0127] A positive reward is given when the error in the titanium content at the endpoint is lower than a set threshold (e.g., 1%); otherwise, a negative reward is given.

[0128] Temperature error: The target temperature range is set as follows:

[0129] Inlet water temperature: ±2℃, outlet water temperature: ±3℃.

[0130] If the temperature deviation exceeds the set threshold, the reward will be negative; the smaller the error, the higher the reward.

[0131] Electrical Efficiency: When electrical efficiency increases by 1%, the reward increases by +10; if electrical efficiency decreases, a negative reward is given, and the smaller the error, the higher the reward.

[0132] Abnormal Operation Alarm: An alarm mechanism will be triggered and penalties will be imposed when the following abnormal operations occur during the smelting process:

[0133] When the final titanium content exceeds the specified range, the inlet and outlet water temperatures exceed ±5℃, or the composition deviation exceeds ±2%, the system will trigger an alarm mechanism and make automatic adjustments.

[0134] Optimization results:

[0135] This method can significantly improve the operational efficiency in the smelting process, specifically in the following ways:

[0136] The deviation of titanium content at the endpoint was reduced by 3%, the number of abnormal operation alarms was reduced by 47%, and the electrical efficiency was improved by 11.5%.

[0137] By optimizing control, temperature and composition deviations during the smelting process can be kept within the target range, significantly improving the stability and energy efficiency of the smelting process while reducing losses caused by abnormal operations.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dynamic prediction of endpoint elements and optimization of process decisions in electroslag remelting smelting, characterized in that, include: S1. Acquire historical datasets and real-time data streams from the smelting process, and preprocess the acquired data; S2. Introducing online stochastic gradient descent and incremental learning algorithms to construct a dynamic element prediction model, receiving sensor data from the smelting process in real time, and dynamically optimizing the element concentration prediction model, including: S21. When new data arrives, the online stochastic gradient descent algorithm is used for local updates to gradually adjust the parameters of the element concentration prediction model in order to minimize the prediction error. S22. The parameters of the element concentration prediction model are periodically updated using an incremental learning algorithm, so that the element concentration prediction model adapts to the accumulation of data streams and ensures a high degree of consistency with the actual data of the current smelting environment. S23. By comparing the prediction errors of the online stochastic gradient descent algorithm and the incremental learning algorithm through the error feedback mechanism, the element concentration prediction model with the smallest error is automatically selected. S3. Based on the deep deterministic strategy gradient algorithm, an optimization decision model is constructed. Based on the elemental composition, smelting conditions, and other environmental parameters predicted online by the elemental dynamic forecasting model, the smelting operation strategy is adjusted in real time, including: S31. Based on the deep deterministic policy gradient algorithm, construct an optimization decision model, including a policy network and a value network. The policy network generates continuous control actions based on the current smelting state; the value network evaluates the impact of state-action on model performance and calculates the Q value to optimize smelting operation parameters. S32. Define the state space and action space; S33. Design a reward function that comprehensively considers element content error, energy consumption, and efficiency, as follows: in, It is the value of the reward function, representing the value of the reward function. t The rewards obtained by the agent at any given moment; , , All represent weighting coefficients; Indicates the first i Predicted content of each element; Indicates the first i The target content of each element; S34. Train the model using experience playback and target network mechanisms to achieve dynamic adjustment of smelting parameters; S4. Set up a real-time feedback mechanism to enable the element dynamic forecasting model and the optimization decision model to work together, realize the real-time synchronous adjustment of element concentration prediction and control strategy, and ensure the accuracy of element concentration prediction and stable control of the smelting process.

2. The method for dynamic prediction of endpoint elements and optimization of process decision in electroslag remelting smelting according to claim 1, characterized in that, Step S1 specifically includes: S11. Clean the data to remove invalid, abnormal, or erroneous data and ensure the quality of the input data. S12. Normalize the data by using normalization methods to standardize or scale the data to maintain a uniform scale and ensure the stability of model training and prediction accuracy. S13. Time series processing: Using the sliding window method, a fixed-length historical data sequence is input each time, enabling the model to use past information for real-time prediction and optimization. S14. Feature extraction: Extract key features from the raw smelting data, including equipment operating status, raw material ratio, process smelting parameters, and environmental data. S15. Store the processed data in a relational database and a time-series database to achieve efficient data management.

3. The method for dynamic prediction of endpoint elements and optimization of process decision in electroslag remelting smelting according to claim 1, characterized in that, In step S12, a data standardization process is adopted, and the specific formula is as follows: in, The original value, The characteristic mean, The characteristic standard deviation is denoted as .

4. The method for dynamic prediction of endpoint elements and optimization of process decision in electroslag remelting smelting according to claim 1, characterized in that, The underlying layer of the element dynamic prediction model is a multi-layer feedforward neural network, consisting of an input layer, multiple hidden layers, and an output layer. The key parameters of the model adjusted by the coupling strategy of the online stochastic gradient descent algorithm and the incremental learning algorithm include the learning rate, the number of neurons in the hidden layer, and the selection of the activation function.

5. The method for dynamic prediction of endpoint elements and optimization of process decision in electroslag remelting smelting according to claim 1, characterized in that, The element concentration prediction model supports online AdaGrad or online Adam optimization algorithms to adaptively adjust the learning rate and accelerate model convergence. At the same time, it combines a sliding window or forgetting mechanism to avoid interference from old data on the element concentration prediction model, so that the element concentration prediction model focuses on the latest data stream.

6. The method for dynamic prediction of endpoint elements and optimization of process decision in electroslag remelting smelting according to claim 1, characterized in that, In step S33, the designed reward function specifically includes: When the content of the endpoint element is close to the target value, a positive reward is given; When smelting energy consumption is too high or efficiency is reduced, a negative reward is given.

7. A system for dynamic prediction of endpoint elements and optimization of process decisions in electroslag remelting based on the method for dynamic prediction of endpoint elements and optimization of process decisions in electroslag remelting according to any one of claims 1-6, characterized in that, include: The module includes a data preprocessing module, an online learning module, a reinforcement learning module, and a synergy module, among which: The data preprocessing module is used to acquire historical datasets and real-time data streams from the smelting process, and to preprocess the acquired data. The online learning module introduces online stochastic gradient descent and incremental learning algorithms to construct a dynamic element prediction model, receive sensor data during the smelting process in real time, and dynamically optimize the element concentration prediction model. The reinforcement learning module, based on the deep deterministic policy gradient algorithm, constructs an optimization decision model and adjusts the smelting operation strategy in real time according to the element composition, smelting conditions and other environmental parameters predicted online by the element dynamic prediction model. The synergistic module is equipped with a real-time feedback mechanism, enabling the dynamic element forecasting model and the optimization decision-making model to work together to achieve real-time synchronous adjustment of element concentration prediction and control strategies, thereby ensuring the accuracy of element concentration prediction and stable control of the smelting process.

Images