Model construction method and system based on titanium sponge production mechanism

Abstract

The invention provides a titanium sponge production mechanism-based model construction method and system, the method comprises the following steps: S1, data acquisition, S2, data preprocessing, S3, mechanism model establishment, S4, fusion modeling, S5, model optimization, and S6, model verification, and when the mechanism model is established in the step S3, an adopted mechanism formula comprises a shrinkage core model of a reaction of magnesium and titanium tetrachloride. The titanium sponge production model with physical interpretability and dynamic adaptability is constructed through a closed loop of'full data acquisition-domain knowledge feature-multi-scale mechanism-intelligent fusion optimization ', so that the problems of high construction cost, long period and poor industrial production adaptability of a mechanism model in the prior art are solved; and double improvement of production efficiency and product quality is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of titanium sponge production, and in particular to a method and system for constructing a model based on the titanium sponge production mechanism. Background Art

[0002] In modern industrial production, mechanism models are crucial tools for describing the internal mechanisms of production processes. Traditional mechanism model construction methods often rely on the programming skills of algorithm engineers, have long construction cycles, and struggle to quickly respond to production demands. Furthermore, existing methods lack adaptability when handling complex dynamic conditions, making them difficult to meet the diverse demands of actual production.

[0003] In titanium sponge production, magnesium reduction of titanium tetrachloride (TiCl₄) is a core process, but several key issues remain unresolved: product purity fluctuates due to unstable residual MgCl₂ levels; energy consumption remains high due to lags in electric furnace power regulation. Traditional process parameter adjustments rely heavily on manual experience, while existing mechanistic models are based on simplified chemical reaction equations, ignoring the influence of dynamic factors such as furnace temperature gradients and material flow rates. This significantly increases model prediction errors, making it impossible to provide real-time, accurate guidance for actual production.

[0004] Chinese patent CN119960956A discloses a cloud computing system resource load prediction method and system based on full lifecycle data, including: obtaining cloud computing full lifecycle data, pre-processing and extracting features for calculation, and generating prediction data; constructing corresponding prediction models for different system characteristics, inputting the prediction data into the prediction model, generating load prediction results, and formulating resource allocation and scheduling strategies based on the load prediction results and resource status. It also has a feedback link to adjust the resource allocation and scheduling strategies based on the difference between the actual value and the predicted value. This invention discloses data processing and modeling ideas, but does not disclose a mechanism model applicable to sponge titanium production, and cannot be used to guide specific sponge titanium production practices.

[0005] It can be seen that a model construction method and system based on the sponge titanium production mechanism is needed to meet the actual production needs. Summary of the Invention

[0006] The purpose of the present invention is to provide a model construction method and system based on the sponge titanium production mechanism, so as to solve the problems of high cost, long cycle and poor adaptability to industrial production in the existing technology of mechanism model construction, and achieve a dual improvement in production efficiency and product quality.

[0007] To achieve the above objectives, the present invention provides a method and system for constructing a model based on the production mechanism of titanium sponge. The technical solution of the present invention is implemented as follows:

[0008] A method for constructing a model based on the production mechanism of titanium sponge, comprising the following steps: S1 data acquisition, S2 data preprocessing, S3 mechanism model establishment, S4 fusion modeling, S5 model optimization, and S6 model verification. When establishing the mechanism model in step S3, the mechanism formula used includes a shrinking core model of the reaction between magnesium and titanium tetrachloride, which is:

[0009] in, is the conversion rate of Mg, is the chemical reaction rate constant, is the concentration of TiCl4 in the gas phase. It is used to describe the interface migration of the Mg particle surface reaction.

[0010] Furthermore, in step S1, the data types include continuous data and discrete data. Continuous data includes temperature, pressure, and flow, while discrete data includes equipment status and alarm information. Data collection is performed at least once per minute. This enables the full collection of temperature field data and chemical reaction parameters within the reduction furnace, enabling the capture of transient operating conditions, improving data integrity, and providing accurate input for the mechanism model. Further steps are to preprocess the data using domain knowledge driven feature engineering methods, including the following steps

[0011] S21, data cleaning, including removing outliers and filling missing values;

[0012] S22, data screening, including correlation analysis and feature selection;

[0013] S23, feature construction, including feature extraction and feature construction;

[0014] S24, establish the original data set and integrate the cleaned, filtered and constructed data into a structured data set.

[0015] The signal-to-noise ratio of the preprocessed data is improved, the domain knowledge features make the model physically interpretable, and the model's extrapolation capability is improved, avoiding the "black box" problem of conventional data-driven models.

[0016] Furthermore, the data screening in step S22 is as follows: selecting alternative features that are highly correlated with the MgCl2 residue through correlation analysis, eliminating redundant features in combination with a feature selection algorithm, and finally retaining the core features.

[0017] Furthermore, the feature extraction in step S23 includes extracting the thermal efficiency gradient and analyzing the frequency domain characteristics of the furnace wall temperature fluctuations through Fourier transform to identify abnormal heat conduction conditions.

[0018] Furthermore, the feature construction in step S23 includes constructing a reduction reaction balance index based on the principle of conservation of materials, that is, the ratio of actual Mg consumption to theoretical consumption, which reflects the degree of reaction.

[0019] Furthermore, in step S4, the output of the mechanism model and the original data features are input into the neural network for training, and the physical constraint loss function is added during training:

[0020] (2)

[0021] in, is the mean square error loss function, λ is the weight coefficient, for The conversion rate of Mg, The conversion rate predicted by the mechanism Ensure that data-driven results do not violate the basic laws of chemical reactions.

[0022] Furthermore, in step S4, the preprocessed dataset is divided into a training set, a validation set, and a test set, with the training set, validation set, and test set ratios being 60%, 20%, and 20%, respectively. Increasing the validation set ratio improves the reliability of hyperparameter optimization.

[0023] Furthermore, in step S5, a Bayesian optimization algorithm is used to optimize hyperparameters. The optimization parameters are: 1 to 3 LSTM layers, 64 to 256 hidden layer nodes, a step size of 32, a learning rate of 1e-4 to 1e-3 on a logarithmic scale, and a physical constraint weight λ of 0.1 to 1.0. Bayesian optimization can find near-optimal solutions within a small number of iterations and is applicable to continuous, discrete, and even hybrid hyperparameter types.

[0024] A titanium sponge production system is constructed using the above-mentioned model construction method.

[0025] Compared with the existing technology, the model construction method based on the titanium sponge production mechanism described in the present invention has the following advantages:

[0026] (1) The construction cycle of the mechanism model is significantly shortened through automated modeling tools.

[0027] (2) By integrating data and mechanisms, the model’s adaptability to complex dynamic working conditions is enhanced.

[0028] (3) By constructing a fusion model through this method, the purity prediction error can be controlled within 5% and energy consumption can be reduced by more than 10%.

[0029] (4) It provides visual construction tools, which lowers the threshold for modeling and can be completed without the need for professional algorithm engineers. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of the model building method described in Example 1 of the present invention;

[0031] Figure 2 This is a data preprocessing flow chart described in Example 1 of the present invention. DETAILED DESCRIPTION

[0032] To make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described are part of the embodiments of the present invention, rather than all of the embodiments. The specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention. It should be noted that, unless there is a conflict, the features in the embodiments and embodiments of the present invention may be combined with each other.

[0033] Example 1

[0034] A method for constructing a model based on the production mechanism of titanium sponge, the method comprising the following steps:

[0035] S1, data acquisition: obtain the set data and actual detection data of the production process through industrial data sets.

[0036] Acquire set data and actual test data from the automated control system of an industrial production process. Automated control systems include DCS systems or PLC systems. Set data includes temperature setpoints, pressure setpoints, and flow setpoints. Actual test data includes actual temperature, pressure, flow, and product quality indicators.

[0037] Data types include continuous data and discrete data. Continuous data includes temperature, pressure and flow, while discrete data includes device status and alarm information.

[0038] According to the characteristics and needs of the production process, set a reasonable data collection frequency, and the data collection frequency should be at least once per minute.

[0039] Preferably, setting data and detection data are collected in real time from the reduction furnace DCS system, including axial temperature distribution sensor data in the furnace, weight change of Mg ingots and Cl2 concentration in the tail gas, with a collection frequency of once per second.

[0040] The collected data is stored in an industrial database for subsequent processing and analysis. Industrial databases include relational databases and time series databases.

[0041] Preferably, the storage tool is the time series database InfluxDB, which is synchronously backed up to the relational database MySQL.

[0042] Step S1 realizes the full collection of temperature field data and chemical reaction parameters in the reduction furnace, which can capture transient operating conditions, improve data integrity, and provide accurate input for the mechanism model.

[0043] S2, data preprocessing: Use domain knowledge-driven feature engineering methods to clean, filter and construct features of the data to establish the original data set.

[0044] Domain knowledge-driven feature engineering leverages domain expertise to guide feature selection, construction, and optimization. This approach helps models better capture underlying patterns and relationships in the data, thereby improving predictive capabilities. Domain knowledge refers to specialized background information related to the problem, such as theories, laws, and experience in the field. Feature engineering transforms raw data into a form suitable for machine learning models through steps such as data preprocessing, feature extraction, and feature selection.

[0045] Domain knowledge-driven feature engineering can reduce data complexity. Data from many real-world problems possess complex structures and hidden relationships, making it difficult to effectively extract key features using algorithms alone. Domain knowledge can help identify which features significantly influence the target variable, thereby reducing noise and enhancing the model's generalization capabilities. Features constructed based on domain knowledge are more easily understood by humans and help explain the model's decision-making process.

[0046] S21, data cleaning, including removing outliers and filling missing values.

[0047] Outliers are identified and removed using statistical analysis methods, including Z-score or IQR algorithms. Preferably, the IQR algorithm is used on the temperature data to identify and remove impulse noise in the early stages of the reduction reaction.

[0048] Fill missing values ​​using interpolation or model-based methods. Interpolation methods include linear interpolation or spline interpolation. Model-based methods include K-nearest neighbor interpolation.

[0049] S22, data screening, including correlation analysis and feature selection.

[0050] Leveraging domain knowledge, correlation analysis is used to identify candidate features that are highly correlated with the target variable. Feature selection algorithms are then used to eliminate redundant features, ultimately retaining the core features. Correlation analysis methods include Pearson correlation coefficient analysis or Spearman rank correlation coefficient analysis. Feature selection algorithms include recursive feature elimination or LASSO regression.

[0051] Preferably, the Spearman rank correlation coefficient analysis method is used to screen out alternative features that are highly correlated with the MgCl2 residual amount, and the LASSO regression algorithm is used to eliminate redundant features, and finally the core features are retained. The alternative features include axial temperature distribution, Mg ingot weight, tail gas Cl2 concentration, magnesium ingot purity, and TiCl4 feeding rate fluctuation. The core features include at least one of the axial temperature difference, Cl2 concentration change rate, magnesium ingot Al / Mn impurity content, and TiCl4 feeding rate fluctuation.

[0052] S23, feature construction, includes feature extraction and feature construction.

[0053] Feature extraction includes time series feature extraction and domain knowledge feature extraction.

[0054] Time series feature extraction involves extracting statistical and frequency domain features from time series data. Statistical features include mean, variance, and peak. Methods for obtaining frequency domain features include Fourier transform or wavelet transform.

[0055] Domain knowledge feature extraction refers to extracting features with physical meaning by combining production process knowledge, such as energy or material conservation principles.

[0056] Preferably, the thermal efficiency gradient is extracted, and the frequency domain characteristics of the furnace wall temperature fluctuation are analyzed by Fourier transform to identify abnormal heat conduction conditions.

[0057] Feature construction is to create new features based on domain knowledge, which may be more representative than the original data.

[0058] Preferably, the reduction reaction balance index is constructed based on the principle of conservation of materials, that is, the ratio of actual Mg consumption to theoretical consumption, which reflects the degree of reaction.

[0059] S24, establish the original data set and integrate the cleaned, filtered and constructed data into a structured data set.

[0060] The cleaned, filtered and constructed data are integrated into a structured data set as the basis for subsequent modeling.

[0061] The signal-to-noise ratio of the preprocessed data is improved by about 60%. The domain knowledge features make the model physically interpretable, and the model's extrapolation capability is improved, avoiding the "black box" problem of conventional data-driven models.

[0062] S3, mechanism model establishment: The mechanism model prediction data is obtained based on the mechanism formula of the production process.

[0063] S31, Mechanism formula selection

[0064] Using the actual production process as input, and based on the physical and chemical principles of titanium sponge production, an appropriate mechanism formula is selected. The mechanism formula includes a rolling force calculation formula, a rolling temperature calculation formula, or a shrinking core model formula for the reaction between magnesium and titanium tetrachloride.

[0065] The rolling force calculation formula is based on the theory of metal plastic deformation, and calculates the relationship between rolling force and rolling parameters (such as roll diameter, rolling speed, material hardness, etc.).

[0066] The rolling temperature calculation formula is based on the principles of heat conduction and heat convection, and calculates the temperature changes during the rolling process.

[0067] Based on the gas-solid reaction kinetics model of magnesium reduction of TiCl₄, a mathematical model describing the gradual consumption of solid particles during chemical reactions or physical processes—the Shrinking Core Model—is introduced to explain the reaction mechanism between solid particles and the surrounding fluid. The Shrinking Core Model quantitatively describes the behavior of solid particles during the reaction. By analyzing the reaction rate and influencing factors, it provides theoretical support for research and applications.

[0068] Preferably, when the mechanism model is established, the mechanism formula used includes a shrinking core model of the reaction between magnesium and titanium tetrachloride, and the formula is:

[0069] (1)

[0070] in is the conversion rate of Mg, is the chemical reaction rate constant, is the concentration of TiCl₄ in the gas phase. Equation (1) describes the interfacial migration of the Mg particle surface reaction. Specifically, it refers to the process in which macromolecules form first-order reactions through different chemical reactions. In the case of a single reactant, the reactant is directly converted into a product. This reaction rate is linearly related to the concentration of the reactant: the higher the concentration, the faster the reaction rate.

[0071] Based on the mechanism formula of the production process, the mechanism model prediction data is calculated.

[0072] S32, parameter estimation

[0073] The unknown parameters in the mechanism model are estimated using experimental data or historical data using methods such as least squares and maximum likelihood estimation.

[0074] Preferably, the least squares method is used to fit The activation energy E was estimated by combining the Arrhenius relationship with temperature and historical batch data (n=500). a =85kJ / mol.

[0075] S33, Model Validation

[0076] By comparing the predicted data of the mechanism model with the actual test data, the accuracy and reliability of the model can be verified. If the error is large, the model needs to be revised or the mechanism formula needs to be reselected.

[0077] Comparing the Mg conversion rate predicted by the mechanism model with the actual weighing data, the initial error was about 12%. By correcting the assumption of the gas flow field distribution in the furnace, the error was reduced to about 8%.

[0078] By constructing a multi-scale mechanism model that takes into account the particle reaction kinetics and mass and heat transfer in the furnace, physical constraints are provided for the data-driven model, avoiding the "extrapolation failure" problem of pure data models.

[0079] S4, fusion modeling: Fuse the preprocessed data with the mechanism model to build a data and mechanism fusion model.

[0080] S41, Data and Mechanism Model Fusion

[0081] Use machine learning algorithms to build data-driven models to capture implicit patterns in the data. Long-short-term memory networks are used to process time series data such as temperature and flow rate to capture the dynamic hysteresis characteristics of the reduction reaction. Machine learning algorithms include linear regression, neural networks, or support vector machines. Preferably, a static regression model is used; if this is not possible, a time series correlation model can be used.

[0082] The fusion modeling strategy is to use the prediction results of the mechanism model as one of the input features of the data-driven model to build a fusion model. Preferably, the output of the mechanism model is input into the neural network together with the original data features for training.

[0083] The theoretical Mg conversion rate and theoretical energy consumption values ​​predicted by the mechanism model are used as input features of LSTM (Long Short-Term Memory) to construct a "mechanism constraint-data fitting" fusion model.

[0084] Add physical constraint loss function during training:

[0085] (2)

[0086] in, is the mean square error loss function, λ is the weight coefficient, for The conversion rate of Mg, The conversion rate predicted by the mechanism ensures that the data-driven results do not violate the basic laws of chemical reactions.

[0087] By integrating data and mechanism models, the adaptability to complex working conditions is greatly improved, avoiding unreasonable predictions.

[0088] S42, model training

[0089] The preprocessed dataset is divided into training, validation, and test sets. The training set is used to train the model. The validation set is used to tune hyperparameters and monitor performance. The test set is used to ultimately evaluate the model's generalization ability. The optimal ratio of the training set, validation set, and test set is 60%, 20%, and 20%. Compared to the conventional 7:3 training / test set split, increasing the validation set ratio improves the reliability of hyperparameter optimization.

[0090] The fusion model is trained using the training set data, and model performance is optimized by adjusting model parameters. Model parameters include the learning rate and regularization parameter. To prevent model overfitting, cross-validation, early stopping, or dropout techniques are preferably used.

[0091] Preferably, the TensorFlow framework is adopted, the Adam optimizer is used, and EarlyStopping and Dropout are set to prevent overfitting.

[0092] Through model training, the prediction error of MgCl2 residue of the fusion model on the test set was reduced to about 4.2%, which is significantly improved compared with the single mechanism model and the single LSTM model, and the prediction deviation of abnormal operating conditions such as furnace wall slagging was reduced by about 50%.

[0093] S5, model optimization: Optimize the model through the Bayesian optimization algorithm and select the optimal hyperparameter combination.

[0094] S51, Hyperparameter Optimization Using Bayesian Optimization Algorithm

[0095] In Bayesian optimization, the optimization problem we need to solve can be expressed as:

[0096] max x∈Xf(x) (3)

[0097] Where f(x) is the objective function, x is the parameter vector, and X is the parameter space.

[0098] The range of hyperparameters that need to be optimized include the number of hidden layer nodes of the neural network, the learning rate, and the regularization parameter.

[0099] Compared with grid search and random search, Bayesian optimization can find near-optimal solutions in fewer iterations and is applicable to continuous, discrete, and even mixed hyperparameters.

[0100] The optimization parameters are: the number of LSTM layers is 1 to 3, the number of hidden layer nodes is 64 to 256, the step size is 32, the learning rate is 1e-4 to 1e-3 on a logarithmic scale, and the physical constraint weight λ is 0.1 to 1.0.

[0101] The optimization process involves searching the hyperparameter space using a Bayesian optimization algorithm to find the hyperparameter combination that optimizes model performance. Bayesian optimization algorithms include Gaussian Process Regression (GPR), and model performance metrics include mean squared error or accuracy.

[0102] We used GPR as the surrogate model and found the optimal parameter combination after 20 iterations: a 2-layer LSTM, 128 nodes, a learning rate of 5e-4, λ = 0.6, and a validation set MSE of 0.018. This setup took only about one-third the time compared to a conventional grid search of over 50 iterations.

[0103] S52, Performance Evaluation

[0104] The model performance is evaluated using validation set data to select the optimal hyperparameter combination.

[0105] S6, Model Validation: Evaluate model performance through test set data to ensure the accuracy and stability of the model.

[0106] S61, test set evaluation

[0107] The optimized model is finally evaluated using the test set data to calculate performance indicators, including mean square error and coefficient of determination R².

[0108] The method of the present invention calculates the MgCl2 residual with an R²=0.92 and a root mean square error (RMSE)=0.35%, which is better than the industry standard R²>0.85.

[0109] S62, model stability test

[0110] The stability and robustness of the model can be tested by introducing noisy data or simulating abnormal operating conditions.

[0111] When 5% Gaussian noise is injected into the temperature data, the model prediction error only increases by 0.8%, compared with the conventional model error increase of more than 3%, and the robustness is significantly improved.

[0112] S63, practical application verification

[0113] The model was deployed in a real-world production environment, and its predictive effectiveness and guidance were verified through real-time data. It was also deployed to a PLC system, which provided real-time recommendations for optimal furnace power adjustment. During a three-month trial run, the titanium sponge purity compliance rate increased from 82% to 95%, unit power consumption decreased from 18kWh / kg to 16.2kWh / kg (a 10% decrease), and the frequency of manual intervention decreased by 70%.

[0114] In addition, a visual modeling tool has been developed, which allows process engineers to configure data preprocessing processes and fusion strategies by dragging and dropping components without writing code. Non-professionals can model independently, shortening the modeling cycle from 45 days to 10 days.

[0115] Through the above steps, an efficient, accurate and adaptable model based on industrial production mechanisms can be constructed to provide strong support for the optimization and control of industrial production processes.

[0116] A titanium sponge production system is constructed using the above-mentioned model construction method.

[0117] Through a closed loop of "full data collection - domain knowledge features - multi-scale mechanisms - intelligent fusion optimization," a titanium sponge production model with physical interpretability and dynamic adaptability was constructed. Compared to conventional methods, this invention breaks the traditional "separation of mechanism and data" model, lowering the technical threshold through Bayesian optimization and visualization tools, ultimately achieving a dual improvement in production efficiency and product quality, providing a replicable implementation paradigm for intelligent process industries.

[0118] Although the present invention is disclosed as above, the present invention is not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the scope defined by the claims.

Claims

1. A method for constructing a model based on the production mechanism of titanium sponge, the method comprising the following steps: S1 data acquisition, S2 data preprocessing, S3 mechanism model establishment, S4 fusion modeling, S5 model optimization, S6 model verification, characterized in that when the mechanism model is established in step S3, the mechanism formula used includes a shrinking core model of the reaction between magnesium and titanium tetrachloride, and the formula is: （1） in, is the conversion rate of Mg, is the chemical reaction rate constant, is the concentration of TiCl4 in the gas phase.

2. The model building method according to claim 1, characterized in that In step S1, the types of data include continuous data and discrete data. The continuous data includes temperature, pressure and flow, and the discrete data includes device status and alarm information. The data collection frequency is at least once per minute.

3. The model building method according to claim 1, characterized in that Step S2 uses a domain knowledge-driven feature engineering method to preprocess the data, including the following steps: S21, data cleaning, including removing outliers and filling missing values; S22, data screening, including correlation analysis and feature selection; S23, feature construction, including feature extraction and feature construction; S24, establish the original data set and integrate the cleaned, filtered and constructed data into a structured data set.

4. The model building method according to claim 3, characterized in that: The data screening in step S22 is as follows: selecting alternative features that are highly correlated with the MgCl2 residue through correlation analysis, eliminating redundant features in combination with a feature selection algorithm, and finally retaining the core features.

5. The model building method according to claim 3, characterized in that: The feature extraction in step S23 includes extracting the thermal efficiency gradient.

6. The model building method according to claim 3, characterized in that: The feature construction in step S23 includes constructing a reduction reaction balance index based on the principle of conservation of materials, that is, the ratio of actual Mg consumption to theoretical consumption.

7. The model building method according to claim 1, characterized in that: In step S4, the output of the mechanism model and the original data features are input into the neural network for training. The physical constraint loss function is added during training: （2） in, is the mean square error loss function, λ is the weight coefficient, for The conversion rate of Mg, The conversion rate predicted by the mechanism.

8. The model building method according to claim 1, characterized in that: In step S4, the preprocessed data set is divided into a training set, a validation set, and a test set, with the ratios of the training set, the validation set, and the test set being 60%, 20%, and 20%.

9. The model building method according to claim 1, characterized in that: In step S5, the Bayesian optimization algorithm is used to optimize hyperparameters. The optimization parameters are: the number of LSTM layers is 1 to 3, the number of hidden layer nodes is 64 to 256, the step size is 32, the learning rate is 1e-4 to 1e-3 on a logarithmic scale, and the physical constraint weight λ is 0.1 to 1.

0.

10. A titanium sponge production system, characterized in that: The system is constructed using the model construction method of any one of claims 1 to 9.

Images