Novel magnesium alloy refining control system

Abstract

The invention provides a novel magnesium alloy refining control system which comprises the steps that according to melt velocity field and concentration field data, a real-time data set is combined, melt state feature vectors are extracted, and an initial state feature set is generated and used for follow-up prediction and optimization processing; aiming at different sensor sampling frequencies and data formats, applying a timestamp aligned multi-source data synchronization algorithm, fusing the real-time data sets, and generating a fused data set with a unified format; according to the fusion data set and the initial state feature set, applying a long short-term memory network time sequence prediction algorithm to generate future change trend data of melt component distribution and flow state; and according to the change trend data and the initial state feature set, a refining parameter optimization model based on gradient descent is constructed, an optimized refining process parameter set is generated, and the melt state is adjusted.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a novel magnesium alloy refining control system. Background Art

[0002] In novel magnesium alloy refining control systems, visualization of the refining process is crucial for achieving precise control. However, in practice, traditional visualization technologies struggle to meet these requirements due to the high temperature, high activity, and uneven composition distribution of magnesium alloy melts. Specifically, the compositional distribution of magnesium alloy melts complicates during the refining process due to temperature gradients and changes in flow state. Existing sensors are susceptible to interference in high-temperature environments, making it difficult to ensure accurate and real-time data acquisition. Furthermore, predicting the melt's flow state relies on a large amount of historical data. However, due to the unique nature of the magnesium alloy refining process, insufficient data accumulation limits the effectiveness of machine learning model training. Furthermore, constructing a dynamic visualization interface requires fusing data from multiple sources, but differences in sampling frequencies and data formats among different sensors complicate data synchronization and processing. These issues make it difficult for operators to accurately grasp the real-time state of the melt during the refining process, which in turn affects the effectiveness of refining parameter adjustments and ultimately leads to unstable quality of magnesium alloy products. Therefore, achieving high-precision data acquisition in a high-interference environment and building efficient data fusion and prediction models are pressing technical challenges for novel magnesium alloy refining control systems. Summary of the Invention

[0003] The present invention provides a novel magnesium alloy refining control system, which mainly includes: 1. Deploy high-temperature resistant and anti-interference sensors in high-temperature and high-activity environments to collect melt temperature and composition distribution data in real time, generate high-precision real-time data sets, and reduce the impact of environmental interference on data accuracy.

[0004] 2. Based on the melt temperature and composition distribution in the real-time data set and combined with the melt fluid characteristics, a melt flow dynamics model based on computational fluid dynamics is constructed to generate melt velocity field and concentration field data to reflect the flow change trend.

[0005] 3. Based on the melt velocity field and concentration field data, combined with the real-time data set, the melt state feature vector is extracted to generate the initial state feature set for subsequent prediction and optimization processing.

[0006] 4. According to different sensor sampling frequencies and data formats, a multi-source data synchronization algorithm with timestamp alignment is applied to fuse the real-time data sets and generate a fused data set in a unified format.

[0007] 5. Based on the fused data set, a dynamic visualization interface is constructed to render the melt temperature distribution heat map, component concentration 3D map and flow velocity vector map in real time to present the melt state change trend.

[0008] 6 Based on the fused data set and the initial state feature set, the long short-term memory network time series prediction algorithm is applied to generate future change trend data of melt composition distribution and flow state.

[0009] 7. According to the change trend data and the initial state feature set, a refining parameter optimization model based on gradient descent is constructed to generate the optimized refining process parameter set and adjust the melt state.

[0010] 8 If the optimized refining process parameter set does not make the melt state meet the preset composition uniformity threshold, the real-time data set is re-collected, the long short-term memory network prediction model is updated, and a new refining process parameter set is generated.

[0011] 9 Based on the new refining process parameter set, the long short-term memory network prediction model and the refining process parameter set are iteratively updated until the melt state meets the preset composition uniformity threshold and the quality of the magnesium alloy product is stabilized.

[0012] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: The present invention discloses a novel magnesium alloy refining control system. To address the uneven distribution of magnesium alloy melt composition in high-temperature, highly reactive environments, the system uses high-temperature, anti-interference sensors to collect melt data in real time, constructs a computational fluid dynamics model to generate melt flow data, extracts state feature vectors, and fuses data sets using a multi-source data synchronization algorithm. This system implements a dynamic visualization interface to render the melt state in real time. Furthermore, it uses a long-short-term memory network to predict melt change trends and constructs a refining parameter optimization model to adjust the melt state. Through high-precision data acquisition, multi-source data fusion, and intelligent prediction and optimization, the present invention achieves precise visualization and intelligent control of the magnesium alloy refining process, effectively improving the quality stability of magnesium alloy products and providing technical support for the optimization of novel magnesium alloy refining processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of a new type of magnesium alloy refining control system of the present invention.

[0014] Figure 2 This is a schematic diagram of a novel magnesium alloy refining control system of the present invention.

[0015] Figure 3 This is another schematic diagram of a novel magnesium alloy refining control system of the present invention. DETAILED DESCRIPTION

[0016] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] like Figure 1-3 In this embodiment, a novel magnesium alloy refining control system may specifically include: Step S101, 1 deploys high-temperature resistant and anti-interference sensors in a high-temperature and high-activity environment to collect melt temperature and composition distribution data in real time, generate high-precision real-time data sets, and reduce the impact of environmental interference on data accuracy.

[0018] The raw data of melt temperature and composition distribution are collected by high-temperature resistant and anti-interference sensors to obtain a real-time raw data set. The real-time raw data set is denoised using a preset signal filtering algorithm to obtain a denoised data set. If the signal-to-noise ratio of the denoised data set is lower than a preset threshold, the denoised data set is secondary optimized using a Kalman filter algorithm to obtain an optimized data set. Based on the optimized data set, the characteristic parameters of the melt temperature and composition distribution are extracted to obtain a characteristic data set. The characteristic data set is classified using a support vector machine algorithm to determine the dynamic change trend of the melt temperature and composition distribution and obtain a change trend result. The change trend result is predicted using a preset time series analysis model to obtain a predicted data set. Based on the predicted data set, a high-precision real-time data set is generated and stored in a database.

[0019] For example, high-temperature, high-activity environments can be equipped with high-temperature, anti-interference sensors to collect melt temperature and composition distribution data in real time and generate high-precision real-time data sets. First, a zirconia-based sensor capable of stable operation at temperatures up to 950 degrees Celsius can be selected. A shielding layer design can be used to reduce electromagnetic interference and ensure sensor signal stability in high-interference environments. The signal acquisition frequency is set to 10 times per second to capture instantaneous changes. A built-in temperature compensation algorithm is used to correct the collected temperature data, for example, by comparing the collected values ​​with standard thermocouple data, with an error control within ±0.5 degrees Celsius. Next, spectral analysis technology is used to monitor the melt composition distribution in real time. A near-infrared spectrometer is used to scan the melt surface with a wavelength range set to 800-2500 nanometers. Principal component analysis (PCA) is used to extract characteristic spectral data. This data is then matched against a pre-set component spectral library to calculate the percentage of major components, such as iron oxide, with an accuracy of ±0.1%. This data is then spatially and temporally aligned with the temperature distribution to form a comprehensive data set. To reduce the impact of environmental interference on data accuracy, adaptive filtering algorithms such as Kalman filtering are introduced to denoise the collected signals. Assuming a noise variance of 0.02, the signal-to-noise ratio is increased to over 15:1 through iterative calculations. Machine learning models are also used to train historical data, predict interference patterns, and dynamically adjust filtering parameters to ensure that data errors do not exceed ±0.3%. Finally, the processed temperature and composition data are stored in a cloud database, and time series analysis methods are used to predict data trends. For example, the ARIMA model is used to predict temperature trends within the next 10 minutes, with a prediction error within ±1 degree Celsius. The results are fed back to the control system in real time to optimize melt process parameters, forming a complete closed-loop logic from data acquisition to analysis and feedback, ensuring data accuracy and process stability in high-temperature environments.

[0020] Step S102, 2: Based on the melt temperature and composition distribution in the real-time data set and the melt fluid characteristics, a melt flow dynamics model based on computational fluid dynamics is constructed to generate melt velocity field and concentration field data to reflect the flow change trend.

[0021] Raw data on melt temperature and composition distribution are acquired from a real-time data set and preprocessed using standardized methods to produce a unified basic data set. Based on this basic data set and combined with fluid characteristic parameters, a computational fluid dynamics framework is constructed. A pre-established kinetic model is used to simulate melt flow behavior and determine preliminary velocity and concentration field data. Finite element analysis (FEA) is then used to perform meshing and numerical calculations on this preliminary velocity and concentration field data to produce refined flow field data. Based on this refined flow field data, the correlation between melt flow and its changing trend is analyzed. If the change in the flow field data exceeds a preset threshold, the flow field data is subjected to secondary correction to determine whether the corrected flow field data conforms to the expected distribution. Using this corrected flow field data, combined with real-time changes in the melt temperature and composition distribution, the parameters of the kinetic model are dynamically updated to obtain an updated model output, and the final velocity and concentration field data are determined. Based on this final velocity and concentration field data, the melt flow trend distribution is mapped and visualized using multi-dimensional visualization tools to produce flow analysis results. Based on the analysis results of the flow changes and the continuous updating of the real-time data set, the input parameters of the dynamic model are cyclically adjusted to obtain the latest change trend distribution and determine whether the latest change trend distribution is consistent with the actual flow behavior.

[0022] For example, for a real-time data set of melt temperature and composition distribution, melt temperature data is first collected through sensors. Assume that the temperature distribution range measured in an industrial furnace is 750°C to 950°C, and the concentration of iron in the composition distribution ranges from 0.3 to 0.5. The data is stored in a database at a frequency of once per minute. A data cleaning algorithm is used to remove outliers. For example, data with a temperature exceeding 1000°C is considered noise and smoothed using a median filter to ensure data reliability. Next, a computational fluid dynamics model is constructed based on the Navier-Stokes equations, combining the melt flow characteristics. The melt density is set to 7000kg / m³, the viscosity is 0.01Pa·s, the boundary conditions are no slip on the furnace wall, and the inlet velocity is 0.2m / s. The equations are discretized using the finite volume method, and the grid is divided into a three-dimensional structure of 100×100×50. The calculation time step is 0.001 seconds to ensure numerical stability. After the model is solved, a melt velocity field is generated, assuming a peak velocity of 0.35 m / s in the center and a trend toward 0.1 m / s at the edges. Concentration field data is also generated, showing that the iron concentration accumulates to 0.48 at the furnace bottom and drops to 0.32 at the top, reflecting the gravitational settling effect. To analyze flow trends, a time series analysis algorithm is used to extract the rates of change of the velocity and concentration fields over time. The velocity fluctuation period is found to be approximately 30 seconds, consistent with the stirring frequency within the furnace. The concentration gradient changes at a rate of approximately 0.002 per minute, indicating slow diffusion of the components. Further integrating business-related insights, the results are linked to optimizing heat distribution within the furnace. By calculating the heat flux (assuming 5000 W / m²), stirring parameters are adjusted to reduce concentration unevenness, ultimately forming a closed-loop control logic to improve melt uniformity. All of the above processes are implemented through an automated computing platform. Data processing and model solving rely on a high-performance computing cluster, and analysis results are fed back to the control system in real time to ensure process continuity and accuracy.

[0023] Step S103, 3: Based on the melt velocity field and concentration field data, combined with the real-time data set, the melt state feature vector is extracted to generate an initial state feature set for subsequent prediction and optimization processing.

[0024] Real-time data is acquired from the melt velocity and concentration distribution. Data fusion technology is used to integrate this multi-source information to generate a preliminary fused dataset. Based on this preliminary fused dataset, features are extracted from the melt state. A support vector machine algorithm is used to classify and quantify these state features, and their distribution pattern is determined. If the distribution pattern deviates from a preset threshold, the preliminary fused dataset undergoes secondary correction processing to generate a corrected feature dataset. Feature vectors are constructed from this corrected feature dataset, and cluster analysis is used to group these feature vectors to generate a classified feature vector set. Based on this classified feature vector set, an initial state feature set is generated. Data format conversion is performed based on the prediction analysis requirements to determine whether the initial state feature set meets the input conditions of the prediction model. If the initial state feature set meets the input conditions, it is transferred to the prediction analysis module to generate prediction result data. If not, the process returns to the feature extraction stage to obtain an updated feature vector set. Using this prediction result data, combined with dynamic monitoring information, optimization parameters are adjusted to determine the final optimization solution data.

[0025] For example, the process of extracting melt state feature vectors and generating an initial state feature set based on melt velocity and concentration field data, combined with real-time data sets, can be achieved using the following techniques. First, three-dimensional velocity component data is acquired from the melt velocity field, assuming a velocity field resolution of 100×100×50, a sampling frequency of 10 Hz, and a velocity component range of 0.1 to 1.5 m / s. A fast Fourier transform (FFT) algorithm is used to perform frequency domain analysis on the velocity field to extract the dominant frequency features. The calculation formula is F(k)=∑[u(x,t)e^(-2πikx / N)], where u(x,t) is the velocity component, k is the wavenumber, and N is the number of sampling points. Results show that the dominant frequency is concentrated between 0.5 and 2 Hz, reflecting the periodic fluctuations of the melt flow. Next, the concentration field data is processed, assuming a concentration range of 0.01 to 0.1 mol / L, with a grid resolution consistent with the velocity field. Principal component analysis (PCA) was used for dimensionality reduction, retaining 95% of the variance. Five principal components were obtained, calculated as X' = X·W, where X is the normalized concentration matrix and W is the eigenvector matrix. This reduced the feature dimension from 10,000 to 5, preserving key concentration distribution information. Subsequently, dynamic features were extracted through time series analysis, combined with real-time datasets (e.g., temperature fields ranging from 1600 to 1800 K). The mean and variance of velocity and concentration were calculated using a sliding window (window size of 1 second, step size of 0.1 second), resulting in the eigenvectors [v_mean, v_var, c_mean, c_var, T_mean] = [0.8, 0.02, 0.05, 0.001, 1700]. To generate the initial state feature set, the velocity dominant frequency, concentration principal component, and dynamic features were concatenated to form a 20-dimensional feature vector. K-means clustering (K=3) was used to classify the feature set, with cluster centers [c1, c2, c3], reflecting the three modes of the melt state: stable, transitional, and unstable. Finally, the feature set was stored in HDF5 format for subsequent prediction models (such as LSTM). The entire process was implemented in Python, relying on the NumPy, SciPy, and Scikit-learn libraries, with a computation time of approximately 0.5 seconds per frame, meeting real-time processing requirements.

[0026] Step S104, 4: A multi-source data synchronization algorithm with timestamp alignment is applied to different sensor sampling frequencies and data formats to fuse the real-time data sets and generate a fused data set in a unified format.

[0027] Raw data is collected from multiple sensors. Preliminary data normalization is performed to address sampling frequency differences and data format inconsistencies, resulting in a preliminary dataset. Based on this preliminary dataset, a timestamp alignment method is used to correct for time deviations in the multi-source data, ensuring temporal consistency across all data sources, and generating a time-aligned dataset. Data synchronization techniques are applied to this time-aligned dataset to address missing or redundant data caused by frequency differences. If data points are missing from a particular data source, they are supplemented through interpolation to generate a synchronized dataset. Based on this synchronized dataset, a multi-source data fusion operation is performed, integrating the real-time data streams batch by batch to generate a fused intermediate dataset. This fused intermediate dataset is formatted and adjusted to meet unified format requirements, resulting in a standardized formatted dataset. Data consistency checks are performed on this formatted dataset. If any data fields do not meet a preset threshold, they are corrected or marked to generate a final fused dataset. Data output results that meet business requirements are generated for this final fused dataset.

[0028] For example, the following method applies a timestamp-aligned multi-source data synchronization algorithm to different sensor sampling frequencies and data formats, fuses real-time datasets, and generates a unified fused dataset. First, assume we have two sensors. Sensor A collects data at a 10Hz frequency, generating 10 data points per second in the format of {timestamp, temperature}, such as {1634567890.1, 25.5}; sensor B collects data at a 5Hz frequency, generating 5 data points per second in the format of {timestamp, humidity}, such as {1634567890.2, 60.0}. During the data synchronization phase, a timestamp alignment algorithm is used to match the timestamps of the two sets of data. A time window of 0.05 seconds is set. If the difference in the timestamps of the two sensors is less than this window, the data are considered synchronized. For example, the time difference between the data points {1634567890.1, 25.5} of sensor A and {1634567890.2, 60.0} of sensor B is 0.1 seconds, which is greater than the window value. Therefore, a linear interpolation algorithm is used to generate a humidity value for sensor B at 1634567890.1. The calculation formula is: Humidity value = 60.0 + (61.0 - 60.0) * (1634567890.1 - 1634567890.2) / (1634567890.3 - 1634567890.2). Assume that the calculated result is 60.5. Next, during the data fusion phase, the synchronized data points are integrated into a unified format of {timestamp, temperature, humidity}, such as {1634567890.1, 25.5, 60.5}. The data is then smoothed using a moving average algorithm with a window size of 3. The average of the current point and the points before and after it is calculated. For example, if the current temperature is 25.5, the previous point is 25.0, and the next point is 26.0, the smoothed temperature value is (25.0 + 25.5 + 26.0) / 3 = 25.5. Finally, a fused dataset is generated and stored in a standard CSV format, containing three columns: timestamp, temperature, and humidity, ensuring direct access by subsequent analysis systems. This method, from timestamp alignment to data smoothing, forms a complete data processing chain, ensuring the accuracy and consistency of multi-source data and providing reliable data support for subsequent business operations such as environmental monitoring. This logically interconnected process ensures the accuracy and consistency of multi-source data.

[0029] Step S105, 5: Based on the fused data set, a dynamic visualization interface is constructed to render the melt temperature distribution heat map, component concentration three-dimensional map and flow velocity vector map in real time to present the melt state change trend.

[0030] Raw data on melt temperature, component concentration, and flow velocity are obtained from the fused dataset. Preprocessing methods are used to clean and format this raw data to generate a structured basic data set. Interpolation methods are applied to fill missing data within this structured basic data set, and time series smoothing is performed to determine a continuous melt state data set. Based on this continuous melt state data set, a dynamic visualization model is constructed to plot a melt temperature distribution heat map in real time, generating a corresponding heat map rendering result. This heat map rendering result is combined with the component concentration data to generate a three-dimensional graphical representation. Using volumetric data rendering techniques, a spatial distribution view of the component concentration is obtained. A vector graphic representation is then overlaid on this spatial distribution view, along with the flow velocity data. If the vector field data exceeds a preset threshold, anomalous regions are marked to identify areas with abnormal flow velocity distribution. Based on these abnormal distribution regions and the time series data, regression analysis is used to analyze melt state change trends and generate a predicted state change curve. This predicted curve is used to update the dynamic visualization interface, adjusting the display of the distribution heat map, three-dimensional graph, and vector graph in real time to determine the final melt state visualization.

[0031] For example, a dynamic visualization interface is constructed based on a fused data set to render a melt temperature distribution heat map, a component concentration three-dimensional map, and a flow velocity vector map in real time to present the melt state change trend. The specific implementation method is as follows. First, using the temperature data in the fused data set, assuming that the data set contains the temperature values ​​of the melt at different positions, for example, the temperature at the coordinate point (1,1,1) is 950.5 degrees Celsius, the discrete temperature data is converted into a continuous distribution through an interpolation algorithm such as trilinear interpolation, and the temperature value of each point in the grid is calculated with an error within 0.1 degrees Celsius. Then, a thermal map rendering technology is used to map the temperature value to a gradient color from red (high temperature) to blue (low temperature) using a color mapping function. The dynamic update frequency is set to 5 times per second to ensure real-time performance. The thermal map is rendered in the browser in combination with WebGL technology to present the temperature change trend from 1500.5 degrees Celsius to 1498.2 degrees Celsius over time. Secondly, for the three-dimensional graph of component concentration, assume that the data set provides concentration data of three components A, B, and C. For example, at a certain point, the concentration of A is 0.4, B is 0.3, and C is 0.3. Through volume data visualization algorithms such as volume rendering, the concentration data is projected into three-dimensional space. Transparency mapping is used to make high-concentration areas opaque and low-concentration areas transparent. The concentration gradient is calculated to analyze the component diffusion rate. It is found that the concentration of component A decreases from 0.4 to 0.35 at a rate of 0.01 per second, indicating a diffusion trend. The dynamic three-dimensional graph is rendered through OpenGL, and the update frequency is also 5 times per second. Finally, for the flow velocity vector diagram, assuming that the data set contains velocity component data, such as the velocity vector at a certain point is (2.5, 1.8, 0.9) meters per second, the arrow representation method is used to draw the vector field. By calculating the velocity magnitude and direction and analyzing the flow trend, it is found that the velocity magnitude increases from 2.5 meters per second to 3.0 meters per second, indicating accelerated flow. The particle tracking algorithm is used to simulate the melt flow path, and the flow trajectory within the next 5 seconds is predicted with an error control within 0.05 meters. The vector changes are dynamically displayed through the real-time rendering engine, and the influence of the flow on the heat and composition distribution is analyzed in combination with the temperature and concentration data, forming a closed-loop analysis logic to ensure that the three data are linked to present a full picture of the melt state changes.

[0032] Step S106,6: Based on the fused data set and the initial state feature set, a long short-term memory network time series prediction algorithm is applied to generate future change trend data of melt composition distribution and flow state.

[0033] Obtain a fusion data set and an initial state feature set, use a data cleaning method to remove outliers and missing values, and obtain a standardized feature set. Through the standardized feature set, apply a feature selection algorithm to extract high-weight features related to the melt composition and flow state, and determine a key feature subset. If the key feature subset meets the preset feature integrity threshold, a long short-term memory network model is used for training to obtain a trained prediction model; if not, return to the data processing step to supplement the features. Through the trained prediction model, input time series data to generate a preliminary prediction result of the melt composition distribution. Through the preliminary prediction result, combined with the historical data of the flow state, a time series analysis method is applied to obtain the change trend of the flow state. Through the change trend of the flow state, data fusion technology is used to integrate the melt composition distribution prediction result and the flow state change trend to generate a comprehensive change trend. Through the comprehensive change trend, visualization technology is applied to output future change trend data of the melt composition distribution and the flow state.

[0034] As an example, the specific implementation method for applying the long short-term memory network (LSTM) time series prediction algorithm to generate data on melt composition distribution and future flow state trends based on a fused dataset and an initial state feature set is as follows: First, assume a fused dataset containing melt composition data (e.g., iron content of 52.3% and silicon content of 25.6%) and flow state parameters (e.g., flow velocity of 1.2 m / s and temperature of 950°C). This data is collected once per minute in a time series, totaling 1000 time points. The initial state feature set extracts key features from the first 10 minutes, such as the average composition ratio and flow velocity fluctuation range (flow velocity fluctuation of 0.1 m / s). Next, a long short-term memory network model is constructed. The input layer receives the time series data, and the hidden layer is set to two layers, each with 64 neurons. A forget gate, input gate, and output gate mechanism are used to capture long-term dependencies. The Adam optimizer is selected with a learning rate of 0.001 and a mean squared error (MSE) loss function. During the training process, the dataset was divided into 80% training and 20% testing. After 100 training cycles, the mean square error (MSE) on the validation set converged to 0.0025, indicating high model prediction accuracy. The trained model was then used to predict the melt composition distribution and flow state for the next 30 minutes. Using the data from the last 10 minutes as input, the output predicted a potential decrease in iron content from 52.3% to 51.8% and an increase in flow rate from 1.2 m / s to 1.3 m / s. The predicted results were further visualized, plotting time trends of composition and flow rate. The decrease in iron content was found to correlate with an increase in temperature (predicted to be 1010°C), suggesting a melt reaction caused by increased heat input. To verify the reliability of the predictions, a retrospective analysis was conducted using historical data to calculate the deviation between the predicted and actual values. The deviation was found to be less than 5%, confirming the model's reliability. Finally, the predicted data was integrated with the production scheduling system. If the flow rate exceeds 1.5 m / s, an automatic warning mechanism is triggered to ensure production safety. Through the above method, a complete logical chain is formed from data processing to model prediction to result analysis, ensuring that the prediction results are both accurate and have practical application value.

[0035] Step S107,7: Construct a refining parameter optimization model based on gradient descent according to the change trend data and the initial state feature set, generate an optimized refining process parameter set, and adjust the melt state.

[0036] Characteristic data is acquired from the initial melt state. Sensors collect temperature, viscosity, and chemical composition to produce an initial state feature set. Based on this initial state feature set and historical change data, a preset time series analysis method is used to extract trend data and determine the melt state change trend. If the trend data exceeds a preset threshold, a parameter optimization model is constructed using a gradient descent algorithm to calculate the optimization direction and step size, resulting in a preliminary optimized parameter set. This preliminary optimized parameter set is then screened using process parameter constraints to eliminate parameters that do not meet control requirements, resulting in a refining process parameter set. The melt temperature and stirring rate are adjusted based on this refining process parameter set to generate adjusted state parameters and obtain an optimized melt state. Real-time feedback data is obtained from the optimized melt state and compared with the initial state feature set to calculate the state change prediction error and obtain a prediction error value. If the prediction error value exceeds a preset threshold, the error value is fed back to the gradient descent algorithm to update the number of iterations and adjustment range, resulting in an updated refining process parameter set.

[0037] For example, the construction of a gradient descent-based refining parameter optimization model requires starting with trend data and an initial state feature set to generate an optimized refining process parameter set to adjust the melt state. First, assume that the initial state feature set includes a melt temperature of 900°C, an oxygen content of 0.02%, and a stirring speed of 200 rpm. The trend data is the oxygen content decrease rate over time of -0.005% / min and the temperature drop rate of -2°C / min. During the data preprocessing stage, the feature set is normalized to map the temperature, oxygen content, and stirring speed to the [0,1] interval using the formula x'=(x-x_min) / (x_max-x_min). For example, the temperature is normalized to (1800-1700) / (1900-1700)=0.8. Next, an objective function is constructed, defined as maximizing melt purity. Purity P is negatively correlated with oxygen content O, P=1-10*O, and the initial P=1-10*0.02=0.8. The optimization model uses a gradient descent algorithm. The partial derivatives of the objective function with respect to the parameters (stirring speed S, refining time T) are ∂P / ∂S = -10*∂O / ∂S and ∂P / ∂T = -10*∂O / ∂T, where ∂O / ∂S = -0.0001% / rpm and ∂O / ∂T = -0.005% / min (derived from trend data). A learning rate α was set to 0.01, and 100 iterations were performed. The update formulas were: S_new = S_old - α*∂P / ∂S, and T_new = T_old - α*∂P / ∂T. Initially, S = 200 rpm and T = 10 minutes. After iterations, S converged to 220 rpm and T converged to 12 minutes. Analysis showed that increasing the stirring speed accelerated the decrease in oxygen content and increased the purity P to 0.85. Ultimately, the optimized parameter set was {temperature: 1800°C, stirring speed: 220 rpm, refining time: 12 minutes}. The control system automatically adjusted the melt state, reducing the oxygen content to 0.015% and achieving a purity of 0.85. Logically, the optimized parameters improved the melt purity, consistent with trend data predictions. Process stability was verified through closed-loop feedback, further improving related services such as subsequent casting quality.

[0038] Step S108,8 If the optimized refining process parameter set does not make the melt state meet the preset composition uniformity threshold, the real-time data set is re-collected, the long short-term memory network prediction model is updated, and a new refining process parameter set is generated.

[0039] If the melt composition uniformity does not reach the preset threshold, real-time data is collected through sensors to obtain a multidimensional process data set. Feature vectors are extracted from the multidimensional process data set, and dimensionality reduction is performed using principal component analysis to obtain a low-dimensional feature set. If the deviation between the low-dimensional feature set and the historical data exceeds the preset range, the long-short-term memory network model is updated to obtain an updated prediction model. The low-dimensional feature set is processed using the updated prediction model to generate an optimized refining process parameter set. The optimized refining process parameter set is used to adjust the equipment operating status to obtain new melt composition data. If the uniformity of the new melt composition data still does not reach the preset threshold, real-time data is repeatedly collected to obtain an updated data set. The process optimization effect is verified using the updated data set, and the final refining process parameter set is determined.

[0040] For example, if the optimized refining process parameter set fails to make the melt state meet the preset composition uniformity threshold (for example, the uniformity index needs to reach 0.95, but the actual value is only 0.87), the system will automatically trigger the real-time data acquisition module to collect key parameters such as melt temperature, stirring speed, and gas flow rate every 5 seconds through the sensor network, forming a new data set containing 1,000 data points. The system then uses this data set to update the long-short-term memory network prediction model. The specific method is to divide the new data into training and validation sets in an 8:2 ratio, use the Adam optimization algorithm, set the learning rate to 0.001, and train 50 rounds. The model accuracy is evaluated by calculating the mean square error (MSE) to ensure that the error is less than 0.01. If the error is higher, the learning rate is automatically adjusted to 0.0005 and retraining is carried out. The updated model can more accurately predict the mapping relationship between the melt state and process parameters. Next, the system generates a new set of refining process parameters based on the updated model. For example, through genetic algorithm optimization, the population size is set to 100, the number of iterations is set to 200, and the objective function is to maximize the composition uniformity index. The final output is a new parameter set with a stirring speed of 300 rpm, a gas flow rate of 50 liters / minute, and a temperature control of 1650 degrees Celsius. This process forms a closed-loop logic. If the new parameter set still does not meet the threshold, the system automatically enters the next round of data collection and model update. During this period, the melt composition analysis module can be linked to the spectrometer to detect element distribution deviations in real time (for example, the iron content deviation must be controlled within 0.5%). This provides additional constraints for parameter optimization and ensures the accuracy and stability of process adjustments.

[0041] Step S109,9 Iteratively update the long short-term memory network prediction model and the refining process parameter set according to the new refining process parameter set until the melt state meets the preset composition uniformity threshold and the quality of the magnesium alloy product is stabilized.

[0042] Acquire historical data and real-time monitoring data, construct an initial set of process parameters and melt state data sets, and obtain preliminary input information. Based on the preliminary input information, use the long short-term memory network model to predict the melt state and determine the deviation value between the predicted result and the actual state. If the deviation value of the predicted result exceeds the preset threshold, adjust the process parameter set according to the deviation value to obtain an updated parameter configuration. Re-run the long short-term memory network model based on the updated parameter configuration to determine whether the melt state meets the preset composition uniformity standard. If the melt state does not meet the standard, extract key variables from the adjusted parameter configuration to obtain a new optimization direction. According to the new optimization direction, iteratively update the process parameter set and prediction model to determine whether the final melt state meets the preset threshold. If the final melt state meets the preset threshold, save the current process parameter set to obtain a stable quality control solution.

[0043] For example, the technical content of updating the prediction model and iteratively optimizing process parameters based on a long-short-term memory network in the magnesium alloy refining process is specifically implemented as follows: First, the system uses sensors to collect real-time melt state data. For example, the melt temperature is set at 680°C, the stirring speed is 300 rpm, and the gas flow rate is 5 liters / minute. It also records key components in the melt, such as aluminum content deviation of 0.5% and zinc content deviation of 0.3%. This data is input into a pre-trained long-short-term memory network model using the Adam optimization algorithm, a learning rate of 0.001, and 128 hidden layer units. The model calculates the predicted compositional uniformity through forward propagation, assuming a predicted value of 0.85 and a preset uniformity threshold of 0.90. The system analyzes the difference between the predicted value and the threshold and finds that it is 0.05. It determines that the current melt state does not meet the standard and triggers the process parameter adjustment mechanism. Next, the system calls an optimization algorithm based on gradient descent, combines historical data and current prediction errors, and calculates a new set of process parameters. For example, the temperature is adjusted to 685°C, the stirring speed is increased to 320 rpm, and the gas flow rate is increased to 5.5 liters / minute. These parameters are automatically sent to the refining equipment control system. After the equipment operates according to the new parameters, the system collects melt state data again. Assuming that the deviation of aluminum content is reduced to 0.2% and the deviation of zinc content is reduced to 0.1%, the long short-term memory network model is re-input for prediction, and the new uniformity value is 0.91, which exceeds the threshold of 0.90 and is judged to be qualified. If the standard is not met, the parameters are adjusted and the model weights are updated iteratively until the requirements are met. The entire process forms a closed loop through data-driven and algorithm optimization to ensure the stable quality of magnesium alloy products. At the same time, the system records the parameters and prediction results of each iteration to form a traceable database to provide data support for subsequent process improvements.

[0044] The above embodiment is only one of the preferred implementation methods of the present invention and should not be used to limit the scope of protection of the present invention. Any changes or modifications that have no substantive meaning made to the main design concept and spirit of the present invention, as long as the technical problems they solve are still consistent with the present invention, should be included in the scope of protection of the present invention.

Claims

1. A new type of magnesium alloy refining control system, characterized in that: include: The data acquisition module is configured to deploy high-temperature resistant and anti-interference sensors in a high-temperature and high-activity environment to collect melt temperature and composition distribution data in real time and generate high-precision real-time data sets; a fluid dynamics modeling module configured to construct a melt flow dynamics model based on computational fluid dynamics according to the real-time data set and melt fluid characteristics, and generate melt velocity field and concentration field data; a feature extraction module configured to extract melt state feature vectors based on the velocity field and concentration field data in combination with a real-time data set to generate an initial state feature set; A multi-source data fusion module is configured to synchronize sensor data of different sampling frequencies and data formats through a timestamp alignment algorithm to generate a fused dataset in a unified format; a visualization module configured to render in real time a melt temperature distribution thermal map, a component concentration three-dimensional map, and a flow velocity vector map based on the fused data set; a prediction module configured to apply a long short-term memory network time series prediction algorithm to generate future change trend data of melt composition distribution and flow state based on the fused data set and the initial state feature set; a parameter optimization module configured to construct a refining parameter optimization model based on a gradient descent algorithm, and generate an optimized refining process parameter set according to the change trend data and the initial state feature set; The closed-loop control module is configured to re-collect data and iteratively update the prediction model and process parameter set when the melt state does not meet the preset composition uniformity threshold until the melt state meets the standard.

2. The system according to claim 1, wherein: The data acquisition module further comprises: Signal filtering unit, used to perform denoising on the original data set; a Kalman filter unit configured to perform secondary optimization when the signal-to-noise ratio of the denoised data set is below a threshold; The feature classification unit uses the support vector machine algorithm to determine the dynamic change trend of melt temperature and composition distribution.

3. The system according to claim 1, wherein: The fluid dynamics modeling module generates velocity field and concentration field data through the following steps: Perform standardization preprocessing on the raw data; Meshing and numerical calculation based on finite element analysis; When the flow field data exceeds the threshold, a secondary correction is performed; Dynamically update model parameters based on real-time temperature and composition changes.

4. The system according to claim 1, wherein: The multi-source data fusion module includes: Timestamp alignment unit, used to correct time deviations of multi-source data; a data interpolation unit configured to fill in missing time point data by linear interpolation; The format conversion unit converts the fused data into a structured dataset.

5. The system according to claim 1, wherein: The visualization module achieves dynamic rendering in the following ways: Trilinear interpolation was used to handle missing data; The temperature value is mapped to a red-blue gradient using the heatmap color mapping function; Apply transparency mapping to characterize concentration gradients in three-dimensional concentration maps; Superimposed vector arrows mark areas of abnormal flow.

6. The system according to claim 1, wherein: The prediction module performs the following operations: Reduce the dimension of the feature set through principal component analysis; Return to supplementary feature extraction when feature completeness is insufficient; Integrate component distribution predictions and flow state trends to generate comprehensive change trend data.

7. The system according to claim 1, wherein: The parameter optimization module includes: Gradient descent calculation unit to determine the optimization direction and step size of process parameters; Constraint screening unit to eliminate parameters that do not meet process requirements; Error feedback unit, updates the iteration parameters when the prediction error exceeds the threshold.

8. The system according to claim 1, wherein: The closed-loop control module achieves iterative optimization in the following ways: When uniformity does not meet the standards, principal component analysis is used to extract multidimensional process data features; Update the long short-term memory network model weights; Generate new process parameter sets through genetic algorithms; The verification is repeated until the compositional uniformity threshold is met.

9. The system according to claim 8, characterized in that The genetic algorithm sets the population size to 100, the number of iterations to 200, and the objective function to maximize the component uniformity index.

10. The system according to any one of claims 1 to 9, characterized in that: The high-temperature resistant and anti-interference sensor is a zirconium oxide-based sensor with an operating temperature of ≥700°C, a sampling frequency of ≥10Hz, a temperature error of ≤±0.5°C, and a component analysis accuracy of ≤±0.1%.

Images