A method and system for generating visual representation of predicted guided thermal state of a rare earth molten salt electrolysis furnace

Abstract

The application discloses a kind of rare earth molten salt electrolytic furnace heat state generation formula visual expression method and system of prediction guidance.Its method includes: collecting relevant data in the operation process of rare earth molten salt electrolytic furnace;Data processing;Temperature hysteresis, power input, thermal momentum, material load and statistical combination characteristics are constructed;Estimation temperature, prediction residual and feature importance are obtained by temperature state estimation model;Estimation temperature, prediction residual, power input, thermal momentum and material load are mapped into visual semantic variables;Parameterized thermal field map and structured prompt word are generated based on visual semantic variables, and auxiliary concept image is formed by generative image model;Curve, thermal field map and concept image are output to digital twin display layer.The application can improve the perceptibility of temperature estimation result and prediction residual signal, and provide auxiliary visual expression for high-temperature industrial process.

Description

Technical Field

[0001] This invention relates to the fields of industrial artificial intelligence, digital twins, industrial soft measurement and visualization technology, specifically to a prediction-guided method and system for generating a visual representation of the thermal state of a rare earth molten salt electrolysis furnace. Background Technology

[0002] Rare earth molten salt electrolysis furnaces typically operate in high-temperature, highly corrosive, and highly disturbed environments. Their temperature status is influenced by multiple factors, including voltage, current, power input, material load, and historical thermal state, exhibiting nonlinearity, strong coupling, and hysteresis. Temperature status directly affects electrolysis efficiency, product quality, furnace lifespan, and production safety; therefore, effective sensing and visualization are essential.

[0003] In existing technologies, temperature monitoring of rare earth molten salt electrolysis furnaces mainly relies on thermocouples, infrared thermometry, image recognition, manual judgment, and conventional monitoring curves. While these methods can provide some temperature information, they have limited ability to express thermal inertia, high-temperature risks, prediction uncertainties, and abnormal disturbances under conditions such as high-temperature corrosion, on-site obstruction, smoke and dust interference, sensor aging, and operating condition fluctuations. They also suffer from problems such as insufficient stability, limited real-time performance, or a single expression method.

[0004] With the development of industrial soft sensing and machine learning technologies, temperature state estimation models can be established using industrial time-series data such as voltage, current, power input, historical temperature, and material load to obtain temperature estimates and prediction residuals. However, existing methods focus primarily on prediction accuracy and error metrics, lacking a mechanism to further transform temperature estimation results, prediction residuals, and key operating variables into perceptible visual information for a digital twin display layer. Furthermore, existing digital twin display interfaces mainly consist of equipment models, curves, dashboards, and alarm values, making it difficult to intuitively express thermal inertia, high-temperature risks, model uncertainties, and abnormal residuals. While generative artificial intelligence can generate visual representations, without industrial data constraints, it can easily produce visual results inconsistent with actual operating conditions or misleading ones.

[0005] Therefore, there is an urgent need for a method that combines temperature state estimation, predictive residual analysis, thermal state coding, and generative visual representation to improve the perceptibility, interpretability, and controllability of the thermal state of rare earth molten salt electrolysis furnaces in the digital twin display layer. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a prediction-guided method and system for generating a visual representation of the thermal state of a rare earth molten salt electrolysis furnace. This invention transforms temperature state estimation results, prediction residuals, and key operating variables such as power input, thermal momentum, and material load into controllable visual semantics, and further generates a visual representation of the thermal state suitable for digital twin display layers.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a prediction-guided method for generating a visual representation of the thermal state of a rare earth molten salt electrolysis furnace, specifically comprising the following steps: S1. By collecting multi-source industrial time-series operation data of rare earth molten salt electrolysis furnaces, a raw industrial time-series dataset with timestamps is constructed. S1 specifically includes the following steps: S1.1 Collect multi-source industrial time-series operation data, including temperature, voltage, current, material load, feed rate, output, equipment operating status, and historical operating conditions. Temperature data is obtained from thermocouples / infrared temperature measuring devices, electrical parameters are obtained from electrolytic power supply systems, and material data is obtained from weighing, feeding ledgers, and production management systems. S1.2 Sampling period setting: Sampling interval from 1 minute to 60 minutes; a fixed sampling period of 1 hour can be selected for scenarios with smooth operating conditions. S1.3. All collected data are bound to timestamps and aggregated to form an unwashed raw time-series dataset; S2. Perform preprocessing operations on the original industrial time series dataset to obtain a standardized time series dataset; The preprocessing operations include time series normalization, resampling, missing value imputation, and median absolute deviation anomaly removal; S2 specifically includes the following steps: S2.1 Sort data by timestamp and align multi-frequency data to a uniform sampling interval; aggregate continuous variables using window mean / median / value, and aggregate cumulative variables such as material input and output using window summation; S2.2 Missing values ​​are filled using forward filling, backward filling, or linear / median interpolation; S2.3, Use the rolling median and median absolute deviation algorithm to remove outliers: Calculate the rolling median of the temperature series within a rolling window of length w. And calculate the measured temperature Deviation from the rolling median: Calculate the median absolute deviation within the rolling window. And construct the corrected standard score: in, The absolute deviation of the median within the scrolling window. To prevent extremely small positive numbers with a denominator of zero; when Greater than the preset threshold When this occurs, the point is identified as an outlier; the scrolling window length w ranges from 5 to 60 sampling points, and the threshold... Take a value between 2.5 and 5.0.

[0008] Step S1 comprehensively covers multi-dimensional operational information of the furnace body, including electrical, temperature, and material data. This avoids the limitations of a single data source and solves the problem of incomplete data from single-point sensors in high-temperature and corrosive environments, providing complete raw data support for end-to-end soft measurement and visualization. Step S2 eliminates data distortion caused by sampling timing misalignment, sensor jumps, data gaps, and on-site dust disturbances, improving the stability of input data and preventing dirty data from causing prediction errors in subsequent models.

[0009] S3. Based on the standardized time-series dataset, the working condition-derived features are constructed by converting multiple feature formulas to obtain a set of combined feature vectors for temperature modeling. S3 specifically includes the following steps: S3.1, Power Input According to voltage and current Calculations show that when multiple material load variables exist, they are combined into a total material load. The calculation formulas are as follows: in, , to These represent material loads from different sources or of different types at time t, where t is the time index. S3.2 Hysteresis characteristics: Extract historical data from 1 to 24 sampling periods to construct hysteresis characteristics to characterize thermal inertia; S3.3, Constructing temperature gradient features: Where k is the preset time interval; S3.4 Constructing cumulative energy characteristics : in, This represents the cumulative window length. S3.5, Constructing thermal momentum characteristics : in, A value greater than zero indicates an overall upward trend in temperature, a value less than zero indicates an overall downward trend in temperature, and a value close to zero indicates a relatively stable temperature. S3.6 Constructing transient heat-to-mass ratio characteristics : in, To prevent extremely small positive numbers with a denominator of zero.

[0010] S4. Based on the combined feature vector set, the time-by-time temperature estimate is obtained through regression machine learning model calculation. ; S4 specifically includes the following steps: S4.1 Select any one or a combination of the following models: gradient boosting tree model, random forest model, linear regression model, ridge regression model, support vector regression model, and neural network model; S4.2, Let the eigenvector of the t-th hour be... The temperature estimate output by the model can be expressed as: in, This represents a temperature state estimation model; S4.3 The dataset is split into training and test sets according to time sequence, and time-series cross-validation is used to select the optimal model by mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). S4.4 Optimal model online inference outputs temperature estimates at each time point. ; S5, Based on temperature estimate Compared with the measured temperature By using the residual calculation formula and the double threshold judgment rule, the predicted residual value and the classification results of the three types of thermal states are obtained. Specifically, the following steps are included: S5.1, based on actual temperature and temperature estimates Calculate the predicted residual Simultaneously calculate the absolute value of the predicted residuals. The calculation formulas are as follows: S5.2, Set the high temperature threshold T h and residual threshold E h ; S5.3, State determination: When <T h And| | <E h When, it is determined to be a normal thermal state; when >T h And| | <E h When | |>E hWhen this occurs, it is determined to be an abnormal residual state.

[0011] Step S4 uses soft sensing to solve for the temperature values ​​of areas within the high-temperature sealed furnace that cannot be directly measured, solving the problems of limited sensor deployment and temperature measurement failure in high-temperature corrosive environments, and obtaining quantitative temperature data without physical temperature measurement points. Step S5 separates two types of problems: actual over-temperature faults and model prediction inaccuracies, distinguishing between high-temperature risks of equipment and uncertainties in data disturbances, thus making up for the shortcomings of traditional monitoring in quantifying model errors.

[0012] S6. Based on the thermal state classification results, temperature estimates, and power / thermal momentum / material load parameters, a complete set of structured visual semantic parameters is obtained through variable normalization and visual mapping formula encoding. S6 specifically includes the following steps: S6.1, for any variable Its normalization result can be expressed as: in, and These are the minimum and maximum values ​​of the variable within a preset time window or historical samples, respectively. To prevent extremely small positive numbers with a denominator of zero; S6.2 Visual disturbance intensity Calculated based on the absolute value of the predicted residual: in, The closer it is to 1, the larger the model residual, and the more obvious the local perturbation, red warning, crack texture or stripe misalignment effect in the generated image; S6.3, Particle Density Calculated based on the normalized power input value: in, and These are the preset minimum particle density and maximum particle density, respectively. This represents the normalized power input; S6.4, Estimated temperature value Mapped to visual hue, brightness, and thermal field intensity; predicting residuals Mapped to visual perturbation intensity, local distortion degree, crack texture intensity, or red alert level; power input Mapped to luminescence intensity and particle density; thermal momentum Mapping to heat flow direction, flow velocity, and motion trend; material load This is mapped to texture thickness, image density, and perceived heat capacity. S6.5 Summarize all mapping parameters to generate a visual semantic set.

[0013] S7. Based on the visual semantic parameter set and the classification results of three types of thermal states, structured prompt words constrained by industrial data are generated by filling in fixed structure prompt word templates; S7.1 Constructing a basic template for structured prompt words: Prompt = A + B + C + D + E + F + G + H; Wherein, A represents the description of the industrial object, B represents the description of the thermal state category, C represents the description of the thermal field color, D represents the description of the thermal flow pattern, E represents the description of the particle density, F represents the description of the residual perturbation, G represents the description of the digital twin interface element, and H represents the conceptual auxiliary expression statement; the constructed structured prompts include the description of the industrial object, the description of the state category, the description of the thermal field color, the description of the thermal flow pattern, the description of the particle density, the description of the residual perturbation, the description of the digital twin interface element, and the conceptual expression statement; S7.2 When the state category is normal thermal state, the prompt words include descriptions of stable orange-red thermal field, smooth heat flow, low disturbance, and low residual; when the state category is high-temperature thermal state, the prompt words include descriptions of yellow-white bright core, strong thermal radiation, high particle density, and high temperature; when the state category is abnormal residual state, the prompt words include descriptions of red warning, local visual distortion, crack texture, increased residual, and unstable heat flow. S7.3 Output a structured Prompt containing bound operating condition data.

[0014] Step S7 abandons the arbitrary creation of prompts by humans and uses measured industrial data to constrain the AI ​​drawing logic, thus solving the defect of generative AI generating images out of thin air that are detached from the actual working conditions and ensuring that the images correspond one-to-one with the actual furnace operating status.

[0015] S8. Based on visual semantic parameters and structured prompts, two types of visualization results are obtained by performing operations along two generation paths: parameterized thermal field map and generative thermal state concept image. Specifically, the following steps are included: S8.1 Parametric thermal field diagram: Based on normalized parameters, measured / estimated temperature, and residual data, plot temperature curves, residual curves, thermal field intensity diagrams, and residual perturbation quantification views; S8.2 Generative Image: Input the prompt from step S7 into the generative model to automatically generate thermal field effect diagrams corresponding to normal / high temperature / abnormal operating conditions.

[0016] S9. Based on parametric thermal field diagrams, generative thermal state concept images, thermal state classification results, and full-process numerical data, an integrated thermal state visualization display interface is obtained and deployed through the integration and configuration of a digital twin platform.

[0017] S9.1 Interface Integration Content: Measured / estimated temperature curves, residual change curves, thermal state classification labels, anomaly pop-up prompts, parameterized thermal field diagrams, and AI-generated conceptual image partitioning windows; S9.2 Deployment carriers: industrial computers, servers, edge devices, and digital twin platforms; S9.3 Display Specifications: The generated images are for auxiliary reference only. Production control shall be based on actual measured sensor data and process thresholds.

[0018] Step S8: Dual-channel image complementarity: The parameter graph retains quantitative engineering data to meet process verification needs, while the AI ​​concept graph enhances the intuitive understanding of the thermal state, balancing data rigor with visual readability. Step S9: Breaking away from the limitations of traditional twin platforms that rely solely on dashboards and curves, operators can simultaneously view quantitative data and intuitive thermal field images, quickly identifying potential high-temperature hazards and abnormal disturbances, thus improving on-site operation and maintenance efficiency.

[0019] A prediction-guided visual representation system for generating the thermal state of a rare earth molten salt electrolytic furnace, used to apply the above-mentioned method for representing the thermal state of an electrolytic furnace, includes: The data acquisition module is used to collect industrial time-series data such as temperature, voltage, current, material load, and production records; the data preprocessing module is used to perform time alignment, missing value completion, and outlier handling. The feature construction module is used to construct features such as power input, historical lag, temperature gradient, thermal momentum, cumulative energy, and transient heat mass ratio; Temperature state estimation module, used to output temperature estimate values; The residual calculation module is used to calculate the prediction residual between the actual temperature and the temperature estimate; The thermal state coding module is used to map temperature estimates, prediction residuals, and key operating condition variables into visual semantics. The predictive prompt word generation module is used to generate structured prompt words; The visual representation generation module is used to generate parametric thermal field maps and generative thermal state images; The digital twin display module is used to display temperature status, residual signals, status categories, and generative visual representation results.

[0020] The present invention provides an electronic device, including a processor and a memory, wherein executable code is stored in the memory, and when the executable code is executed by the processor, the processor executes the above-described method for expressing the thermal state of an electrolytic furnace.

[0021] The present invention provides a computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to execute the above-described method for expressing the thermal state of an electrolytic furnace.

[0022] Compared with the prior art, the present invention has the following beneficial effects: (1) By introducing prediction residual calculation and dual threshold judgment rules, the present invention divides the thermal state into three categories: normal, high temperature and abnormal residual, which solves the problem that traditional monitoring is difficult to quantify model uncertainty and abnormal disturbance. (2) This invention maps temperature, residual, power, thermal momentum and material load into visual semantics such as color, disturbance, particle density, flow direction and texture through thermal state encoding, which solves the problem of the single information expression of the digital twin display layer. (3) This invention uses a predictive guidance prompt generation mechanism to make the generated images subject to industrial data constraints, thus solving the problem of unclear correspondence between the generated content and the working conditions; (4) This invention outputs parameterized thermal field diagrams and generative conceptual images through dual channels and integrates them into the digital twin display layer, taking into account both numerical interpretability and visual intuitiveness. It realizes intuitive perception of the thermal state of the furnace body, identification of high temperature risks and prompts of abnormal disturbances, effectively improving the efficiency of operators in making comprehensive judgments on thermal state and abnormal working conditions. Attached Figure Description

[0023] Figure 1 A roadmap for visual representation methods; Figure 2 A schematic diagram of the data preprocessing and feature construction process; Figure 3 This is a schematic diagram of the temperature state estimation model and its output. Figure 4 This is a schematic diagram of hot state encoding and visual semantic mapping; Figure 5 This is a schematic diagram illustrating the process of generating predictive prompts. Figure 6 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] Example 1 like Figure 1As shown, a prediction-guided method for generating a visual representation of the thermal state of a rare earth molten salt electrolysis furnace is presented. This method takes industrial time-series data from the furnace's operation as input and, through steps such as data preprocessing, feature construction, temperature state estimation, prediction residual calculation, thermal state encoding, prediction-guided prompt generation, and digital twin display layer output, transforms the temperature estimation results and key operating condition variables into controllable visual semantics. It further generates a visual representation of the thermal state for use in the digital twin display layer. The method includes the following steps: S1. Collect multi-source industrial data during the operation of rare earth molten salt electrolysis furnaces. Industrial time-series data is collected during the operation of rare earth molten salt electrolysis furnaces. This data includes, but is not limited to, temperature, voltage, current, material load, feed rate, output, equipment operating status, timestamps, and historical operating condition records. Temperature data can be obtained from thermocouples, infrared temperature measuring devices, furnace surface temperature detection devices, or existing monitoring systems; voltage and current data come from the electrolysis power supply system; and material load data can come from feed records, weighing devices, inventory records, or production management systems. The data sampling period is from 1 minute to 60 minutes; in cases where the data recording frequency is low or data fluctuations are minimal in the industrial field, a uniform sampling period of 1 hour can be used. The collected data is aligned according to timestamps to form a unified industrial time-series dataset.

[0026] S2. Preprocessing industrial time-series data The collected industrial time-series data is preprocessed, including time sorting, time alignment, resampling, missing value completion, and outlier handling.

[0027] Specifically, the raw data is first sorted according to timestamps, and data from different sampling frequencies are aligned to a preset time interval. For continuous variables such as voltage, current, and temperature, the mean, median, or last value within the time window can be used as the representative value for that time interval; for cumulative variables such as feed amount and output, the summation or cumulative value within the time window can be used as the representative value for that time interval.

[0028] For missing values, methods such as forward imputation, backward imputation, linear interpolation, median interpolation, or interpolation based on adjacent time windows can be used to complete them. For data points in the temperature curve that have local peaks, sudden drops, or significant deviations from the normal operating trend, outlier detection methods based on the rolling median and median absolute deviation can be used for processing.

[0029] Specifically, the rolling median of the temperature series is calculated within a rolling window of length w. And calculate the temperature value. Deviation from the rolling median: Further calculate the median absolute deviation within the rolling window. And construct the corrected standard score: in, The absolute deviation of the median within the scrolling window. To prevent extremely small positive numbers with a denominator of zero. When Greater than the preset threshold If this occurs, the point is identified as an outlier.

[0030] Preferably, the scrolling window length w is 5 to 60 sampling points, and the threshold... Use values ​​between 2.5 and 5.0. Data points identified as outliers can be marked as missing values ​​and filled using interpolation.

[0031] S3. Constructing thermal state estimation features After data preprocessing, thermal state estimation features are constructed based on temperature, voltage, current, material load, and historical operating condition data. These features include basic operating condition features, historical lag features, energy input features, thermal inertia features, and material load features, such as... Figure 2 As shown.

[0032] Among them, power input According to voltage and current The calculation yielded: When multiple material load variables exist, the individual material variables are combined into a total material load. : in, , to These represent material loads from different sources or of different types at time t.

[0033] To characterize the thermal inertia of the rare earth molten salt electrolysis furnace, historical hysteresis features of temperature, power input, and material load are constructed. Preferably, the hysteresis order is 1 to 24 sampling periods. For example, the historical hysteresis features of the previous time step are constructed. -1. Temperature of the first two cycles -2. Power input in the first 3 cycles -3 and material load of the first 6 cycles Features such as -6.

[0034] To characterize the temperature change trend, a temperature gradient feature is constructed: Where k is a preset time interval, preferably 1 to 6 sampling periods.

[0035] To characterize the heat accumulation effect, a cumulative energy feature is constructed: in, The cumulative window length is preferably 3 to 24 sampling periods.

[0036] To characterize the rising and falling trends of the furnace body's thermal state, thermal momentum features are constructed: in, A value greater than zero indicates an overall upward trend in temperature, a value less than zero indicates an overall downward trend in temperature, and a value close to zero indicates a relatively stable temperature.

[0037] Preferably, the length of the thermal momentum window Take 3 to 12 sampling periods.

[0038] To characterize the transient thermal response under unit material load, a transient heat-to-mass ratio feature is constructed: in, To prevent extremely small positive numbers with a denominator of zero.

[0039] By constructing the above features, the temperature state estimation model can utilize not only the voltage, current, and temperature data at the current moment, but also historical thermal state, energy input, thermal inertia, and material load information.

[0040] S4. Establish a temperature state estimation model and output the temperature estimate. The feature vector constructed in step S3 is used as input to the temperature state estimation model to obtain the temperature estimate at the current time. Let the feature vector at hour t be... The temperature estimate output by the model can be expressed as: in, This represents the temperature state estimation model. The temperature state estimation model can employ gradient boosting tree models, random forest models, linear regression models, ridge regression models, support vector regression models, neural network models, or combinations of these models. Preferably, LightGBM, XGBoost, Random Forest, or Ridge Regression are used as the temperature state estimation model. During model training, time series cross-validation can be used to evaluate model performance, with evaluation metrics including mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). To avoid leakage of temporal information, the training and test sets are divided according to chronological order, without random shuffling.

[0041] In specific experiments, models such as Naive Baseline, Ridge Regression, Random Forest, and LightGBM can be used for comparison. Naive Baseline uses the temperature of the previous moment as the current temperature estimate to test whether the model depends solely on temperature inertia; Ridge Regression is used to characterize linear relationships; and Random Forest and LightGBM are used to capture nonlinear relationships.

[0042] In a set of experimental data, the cross-validation MAE, RMSE, and R² of Naive Baseline were 10.6970, 19.6170, and 0.2624, respectively; those of LightGBM were 6.9207, 11.7025, and 0.6770; those of Ridge Regression were 7.3827, 11.0920, and 0.6895; and those of Random Forest were 7.3851, 11.8535, and 0.6702. The experimental results show that, compared to baseline methods that rely solely on the temperature of the previous time step, machine learning models can more effectively utilize variables such as voltage, current, power input, historical temperature, and material load for temperature state estimation. In the final test interval, the MAE, RMSE, and R² of the LightGBM model were 5.3979, 10.8669, and 0.8852, respectively, indicating that the model can obtain relatively stable temperature estimation results within the test interval. It should be noted that the above experimental data are used to demonstrate the feasibility of this embodiment and do not limit this embodiment to using only the LightGBM model or the aforementioned index ranges.

[0043] S5. Calculate the predicted residuals and determine the thermal state category. like Figure 3 As shown, based on the actual temperature and temperature estimates Calculate the predicted residuals: Simultaneously calculate the absolute value of the prediction residuals: The prediction residual is used to characterize the estimation bias and uncertainty of the temperature state estimation model. When the absolute value of the prediction residual is small, it indicates that the model estimation result is close to the actual temperature; when the absolute value of the prediction residual is large, it indicates that there may be sudden high temperature changes, sensor anomalies, operating condition disturbances, or a decrease in model applicability.

[0044] Based on the temperature estimate and predicted residual, the current thermal state is divided into normal thermal state, high-temperature thermal state, and abnormal residual state. Preferably, a high-temperature threshold T is set. h and residual threshold E h The status is determined according to the following rules: when( <T h And (| | <E h When this condition is met, it is considered a normal thermal state. when( >T h And (| | <E h When ), it is determined to be in a high-temperature thermal state; When (| |>E h When ), it is determined to be an abnormal residual state.

[0045] In a preferred embodiment, if both a high-temperature state and an abnormal residual state occur simultaneously, the abnormal residual state is prioritized, and both a high-temperature warning and a residual perturbation warning are overlaid in the visual representation to avoid underestimating the high-temperature risk due to significant model uncertainty. Preferably, the high-temperature threshold (T) h The residual threshold (E) can be set according to process requirements, historical operating data, or expert experience. h The temperature can be set from 5 ℃ to 50 ℃, or it can be adaptively determined based on the 80% to 99% quantile of the residual distribution of the training set.

[0046] S6. Encode the thermal state estimation results and operating condition variables into visual semantics. As shown in Table 1, thermal state coding rules are constructed based on temperature estimates, prediction residuals, power inputs, thermal momentum, and material loads to transform industrial variables into the visual semantics required for generative visual representation. This process is called thermal state coding.

[0047] Table 1: Thermal State Encoding Rules like Figure 4 As shown, specifically, the temperature estimate... Mapped to visual hue, brightness, and thermal field intensity; temperature estimate. The higher the value, the brighter or higher-temperature hue the thermal field color tends to be, with corresponding increases in brightness and thermal field intensity. This will predict the residual. Mapped to visual perturbation intensity, local distortion degree, crack texture intensity, or red alert level; predict residuals The larger the value, the stronger the visual disturbance. Input power... Mapped to luminescence intensity and particle density; power input The higher the value, the higher the luminescence intensity and particle density in the image. This relates to thermal momentum. Mapped to heat flow direction, flow velocity, and motion tendency; when A value greater than 0 indicates an upward temperature trend; when... When it is less than 0, it indicates a cooling trend; when When the heat flux approaches zero, a stable heat flux is displayed. The material load is then adjusted. This is mapped to texture thickness, image density, and thermal capacity perception intensity. The greater the material load, the thicker the image texture or the higher the image density.

[0048] Before mapping, each industrial variable can be normalized. For any variable... Its normalization result can be expressed as: in, and These are the minimum and maximum values ​​of the variable within a preset time window or historical samples, respectively. To prevent extremely small positive numbers with a denominator of zero.

[0049] In a preferred embodiment, the visual disturbance intensity Calculated based on the absolute value of the predicted residual: in, The closer to 1, the larger the model residual, and the more pronounced the local perturbations, red warnings, crack textures, or stripe misalignments in the generated image. When close to 0, it shows smooth heat flow; when When the value is close to 1, it displays obvious local distortion, warning colors, or crack textures.

[0050] In another preferred embodiment, particle density Calculated based on the normalized power input value: in, and These are the preset minimum particle density and maximum particle density, respectively. This represents the normalized power input.

[0051] Through the above encoding, a structured visual semantic set is formed, including thermal field color, thermal field brightness, heat flow pattern, particle density, texture thickness, residual perturbation, and status warning information. The system can transform the originally abstract industrial variables and model residual information into the control conditions required for visual expression.

[0052] S7. Generate structured prompts based on the predictive prompt generation framework. Based on the thermal state categories obtained in step S5 and the visual semantic set obtained in step S6, a prediction guidance prompt word generation framework is constructed. This framework converts the temperature state estimation results and the operating condition variable encoding results into structured prompt words, which are used to guide the generative artificial intelligence model to generate the corresponding visual expression results of the thermal state. These prompt words are not freely input by humans, but are automatically generated from the temperature state estimation results, prediction residuals, and thermal state encoding results.

[0053] The structured cue words include descriptions of industrial objects, state categories, thermal field colors, thermal flow patterns, particle density, residual perturbations, digital twin interface elements, and conceptual statements.

[0054] Its basic template is: Prompt = A + B + C + D + E + F + G + H; Wherein, (A) is the description of the industrial object, representing a rare earth molten salt electrolysis furnace or its digital twin object; (B) is the description of the thermal state category; (C) is the description of the thermal field color; (D) is the description of the heat flow pattern; (E) is the description of the particle density; (F) is the description of the residual perturbation; (G) is the description of the digital twin interface elements; and (H) is the conceptual auxiliary expression statement.

[0055] When the state category is normal thermal state, the prompts include descriptions such as stable orange-red thermal field, smooth heat flow, low disturbance, and low residual; when the state category is high-temperature thermal state, the prompts include descriptions such as yellow-white bright core, strong thermal radiation, high particle density, and high-temperature warning; when the state category is abnormal residual state, the prompts include descriptions such as red warning, local visual distortion, cracked texture, increased residual warning, and unstable heat flow. The overall flow of S3-S7 is as follows: Figure 5 As shown.

[0056] In this way, the input of the generative artificial intelligence model no longer depends solely on subjective human description, but is constrained by temperature estimates, prediction residuals, and key operating condition variables. This improves the correspondence between the generated results and the industrial thermal state, and reduces the inconsistency problem caused by relying entirely on subjective human description.

[0057] S8. Generate parameterized thermal field maps and generative visual representations. Two types of visual representation results are generated based on the thermal state encoding results.

[0058] The first type is parametric thermal field maps. These images are generated based on temperature estimates, actual temperatures, prediction residuals, and normalized visual parameters. They can include thermal field intensity maps, residual perturbation maps, temperature estimate curves, actual temperature curves, residual curves, and a comprehensive state display interface. Parametric thermal field maps are used to preserve numerical representation and engineering interpretability.

[0059] The second category is generative thermal state concept images. These images are generated by inputting structured prompts into a generative artificial intelligence model and are used to express the thermal field atmosphere, heat flow trend, residual perturbation, and digital twin display effects under normal thermal state, high-temperature thermal state, and abnormal residual state.

[0060] In practical applications, parametric thermal field maps and generative thermal state concept images can be simultaneously output to the digital twin display layer. The parametric thermal field map provides traceable numerical data, while the generative thermal state concept image enhances operators' intuitive understanding of thermal conditions, anomaly risks, and model uncertainties.

[0061] S9. Output the thermal state visual representation result in the digital twin presentation layer. The parameterized thermal field map, generative thermal state concept image, temperature estimate, actual temperature, prediction residual, thermal state category, and anomaly warning information obtained in step S8 are output to the digital twin display layer.

[0062] The digital twin display layer can include temperature state estimation curves, prediction residual curves, thermal state indices, thermal field diagrams, state category labels, anomaly warning labels, and generative visual representation windows. Operators can simultaneously observe temperature numerical changes, model residual changes, thermal state trends, and anomaly risk warnings through this display layer.

[0063] It should be noted that the generative visual representation results in this embodiment are used for conceptual auxiliary representation in the digital twin display layer and do not replace real equipment structure diagrams, engineering simulation results, actual monitoring screenshots, temperature measurement devices, or production control decisions. Production control should still be based on real sensor data, deterministic model outputs, alarm thresholds, and on-site operating procedures.

[0064] It should be noted that the generative thermal state images in this embodiment are mainly used for conceptual auxiliary expression in the digital twin display layer, and do not replace real equipment structure diagrams, engineering simulation results, actual monitoring screenshots, temperature measurement devices, or production control decisions.

[0065] The key parameters and preferred ranges of this embodiment are shown in Table 2: Table 2: Key Parameters and Optimal Range Through the above solution, this embodiment can achieve the following technical effects. First, by using industrial time-series data and temperature state estimation models, temperature estimates can be obtained when it is difficult to directly and stably observe the temperature inside the furnace, thus providing a data foundation for expressing the thermal state. Second, by introducing predictive residuals, model errors and uncertainties can be transformed into abnormal residual states and visual disturbance signals, making residual information that is not easily perceived in traditional monitoring more intuitive. Third, through thermal state coding rules, temperature estimates, prediction residuals, power inputs, thermal momentum, and material loads can be transformed into visual semantics such as color, brightness, particle density, flow trends, texture thickness, and visual disturbances, thereby improving the information expression capabilities of the digital twin display layer. Fourth, through the predictive guidance prompt generation mechanism, generative images no longer rely solely on subjective human descriptions, but are constrained by industrial data and model outputs, thereby improving the correspondence between generative visual expressions and actual working conditions.

[0066] Fifth, the dual-channel output of parametric thermal field diagrams and generative thermal state images can balance numerical interpretability and visual perception, providing operators with auxiliary references for understanding the thermal state, high-temperature risks, and abnormal residuals of rare earth molten salt electrolysis furnaces.

[0067] Example 2 This embodiment provides a system for expressing the thermal state of an electrolytic furnace, used in applying the method for expressing the thermal state of an electrolytic furnace in Embodiment 1. For example... Figure 6 As shown, the system includes a data acquisition module, a data preprocessing module, a feature construction module, a temperature state estimation module, a residual calculation module, a thermal state encoding module, a prediction guidance prompt word generation module, a visual expression generation module, and a digital twin display module.

[0068] The system comprises the following modules: a data acquisition module for collecting industrial time-series data such as temperature, voltage, current, material load, and production records; a data preprocessing module for performing time alignment, missing value completion, and outlier handling; a feature construction module for constructing features such as power input, historical lag, temperature gradient, thermal momentum, cumulative energy, and transient heat mass ratio; a temperature state estimation module for outputting temperature estimates; a residual calculation module for calculating the prediction residual between the actual temperature and the temperature estimate; a thermal state encoding module for mapping temperature estimates, prediction residuals, and key operating condition variables to visual semantics; a prediction guidance prompt generation module for generating structured prompts; a visual representation generation module for generating parametric thermal field maps and generative thermal state images; and a digital twin display module for displaying temperature states, residual signals, state categories, and generative visual representation results.

[0069] Example 3 This embodiment provides an electronic device, including a processor and a memory. The memory stores executable code. When the executable code is executed by the processor, the processor executes the electrolytic furnace thermal state expression method of Embodiment 1.

[0070] Example 4 This embodiment provides a computer-readable storage medium storing executable code thereon. When the executable code is executed by a processor of an electronic device, the processor executes the electrolytic furnace thermal state expression method of Embodiment 1.

[0071] 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 prediction-guided method for generating a visual representation of the thermal state of a rare earth molten salt electrolysis furnace, characterized in that, Includes the following steps: S1. By collecting multi-source industrial time-series operation data of rare earth molten salt electrolysis furnaces, a raw industrial time-series dataset with timestamps is constructed. S2. Perform preprocessing operations on the original industrial time series dataset to obtain a standardized time series dataset; S3. Based on the standardized time-series dataset, the working condition-derived features are constructed by converting multiple feature formulas to obtain a set of combined feature vectors for temperature modeling. S4. Based on the combined feature vector set, the time-by-time temperature estimate is obtained through regression machine learning model calculation. ; S5. Estimated temperature value Compared with the measured temperature Perform residual calculation formula and double threshold judgment rule, and output the predicted residual value and the three thermal state classification results; S6. Based on the thermal state classification results, temperature estimates, and load parameters, a complete set of structured visual semantic parameters is obtained through variable normalization and visual mapping formula encoding. S7. Based on the visual semantic parameter set and the classification results of three types of thermal states, structured prompt words constrained by industrial data are generated by filling in fixed structure prompt word templates; S8. Based on visual semantic parameters and structured prompts, two types of visualization results are obtained by performing operations along two generation paths: parameterized thermal field map and generative thermal state concept image. S9. Based on parametric thermal field diagrams, generative thermal state concept images, thermal state classification results, and full-process numerical data, an integrated thermal state visualization display interface is obtained and deployed through the integration and configuration of a digital twin platform.

2. The method for predictive-guided generation of a visual representation of the thermal state of a rare earth molten salt electrolysis furnace according to claim 1, characterized in that, In step S2, the preprocessing operation specifically includes the following steps: S2.1 Sort the data by timestamp and align the original industrial time series dataset to a uniform sampling interval; aggregate continuous variables using window mean / median / value, and aggregate cumulative variables such as material input and output using window summation; S2.2 Missing values ​​are filled using forward filling, backward filling, or linear / median interpolation; S2.

3. Use the rolling median and median absolute deviation algorithm to remove outliers; Calculate the rolling median of the temperature series within a rolling window of a preset length. And calculate the measured temperature Deviation from the rolling median: ； Calculate the median absolute deviation within the rolling window. And construct the corrected standard score: ； in, The absolute deviation of the median within the scrolling window. To prevent extremely small positive numbers with a denominator of zero; when Greater than the preset threshold If this occurs, the point is identified as an anomaly.

3. The method for predictive-guided generation of a visual representation of the thermal state of a rare earth molten salt electrolysis furnace according to claim 1, characterized in that, S3 specifically includes the following steps: S3.1, Power Input According to voltage and current Calculations show that when multiple material load variables exist, they are combined into a total material load. The calculation formulas are as follows: ； ； in, , to These represent material loads from different sources or of different types at time t, where t is the time index. S3.2 Hysteresis characteristics: Extract historical data from 1 to 24 sampling periods to construct hysteresis characteristics to characterize thermal inertia; S3.3, Constructing temperature gradient features: ； Where k is the preset time interval; S3.4 Constructing cumulative energy characteristics : ； in, This represents the cumulative window length. S3.5, Constructing thermal momentum characteristics : ； in, A value greater than zero indicates an overall upward trend in temperature, a value less than zero indicates an overall downward trend in temperature, and a value close to zero indicates a relatively stable temperature. S3.6 Constructing transient heat-to-mass ratio characteristics : ； in, To prevent extremely small positive numbers with a denominator of zero.

4. The method for predictive-guided generation of a visual representation of the thermal state of a rare earth molten salt electrolysis furnace according to claim 1, characterized in that, S4 specifically includes the following steps: S4.2 Based on the regression machine learning model, let the feature vector of hour t be... The temperature estimate output by the model can be expressed as: ； in, This represents a temperature state estimation model; S4.

3. Use time-series cross-validation to select the optimal model based on performance metrics such as mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). S4.4 Optimal model online inference outputs temperature estimates at each time point. .

5. The predictive-guided visual representation method for generating the thermal state of a rare earth molten salt electrolysis furnace according to claim 1, characterized in that, S5 specifically includes the following steps: S5.1, based on actual temperature and temperature estimates Calculate the predicted residual Simultaneously calculate the absolute value of the predicted residuals. The calculation formulas are as follows: ； ； S5.2, Set the high temperature threshold T h and residual threshold E h ; S5.3, State determination: When <T h And| | <E h When, it is determined to be a normal thermal state; when >T h And| | <E h When | |>E h When this occurs, it is determined to be an abnormal residual state.

6. The method for predictive-guided generation of a visual representation of the thermal state of a rare earth molten salt electrolytic furnace according to claim 1, characterized in that, S6 specifically includes the following steps: S6.1, for any variable Its normalization result It can be represented as: ； in, and These are the minimum and maximum values ​​of the variable within a preset time window or historical samples, respectively. To prevent extremely small positive numbers with a denominator of zero; S6.2 Visual disturbance intensity Calculated based on the absolute value of the predicted residual: ； in, The closer it is to 1, the larger the model residual, and the more obvious the local perturbation, red warning, crack texture or stripe misalignment effect in the generated image; S6.3, Particle Density Calculated based on the normalized power input value: ； in, and These are the preset minimum particle density and maximum particle density, respectively. This represents the normalized power input; S6.4, Estimated temperature value Mapped to visual hue, brightness, and thermal field intensity; predicting residuals Mapped to visual perturbation intensity, local distortion degree, crack texture intensity, or red alert level; power input Mapped to luminescence intensity and particle density; thermal momentum Mapping to heat flow direction, flow velocity, and motion trend; material load This is mapped to texture thickness, image density, and perceived heat capacity. S6.5 Summarize all mapping parameters to generate a visual semantic set.

7. The method for predictive-guided generation of a visual representation of the thermal state of a rare earth molten salt electrolysis furnace according to claim 1, characterized in that, S7 specifically includes the following steps: S7.1 Constructing a basic template for structured prompt words: Prompt = A + B + C + D + E + F + G + H; Where A is the description of the industrial object, B is the description of the thermal state category, C is the description of the thermal field color, D is the description of the thermal flow pattern, E is the description of the particle density, F is the description of the residual perturbation, G is the description of the digital twin interface element, and H is the conceptual auxiliary expression statement. The constructed structured cue words include descriptions of industrial objects, state categories, thermal field colors, heat flow patterns, particle density, residual perturbations, digital twin interface elements, and conceptual expression statements. S7.2 When the state category is normal thermal state, the prompt words include descriptions of stable orange-red thermal field, smooth heat flow, low disturbance, and low residual; when the state category is high-temperature thermal state, the prompt words include descriptions of yellow-white bright core, strong thermal radiation, high particle density, and high temperature; when the state category is abnormal residual state, the prompt words include descriptions of red warning, local visual distortion, crack texture, increased residual, and unstable heat flow. S7.3 Output a structured Prompt containing bound operating condition data.

8. The predictive-guided visual representation method for generating the thermal state of a rare earth molten salt electrolysis furnace according to claim 1, characterized in that, S8 specifically includes the following steps: S8.1 Parametric thermal field diagram: Based on normalized parameters, measured / estimated temperature, and residual data, plot temperature curves, residual curves, thermal field intensity diagrams, and residual perturbation quantification views; S8.2 Generative Images: Input the prompts into the generative artificial intelligence model to automatically generate thermal field effect diagrams corresponding to normal / high temperature / abnormal operating conditions.

9. The predictive-guided visual representation method for generating the thermal state of a rare earth molten salt electrolytic furnace according to claim 1, characterized in that, S9 specifically includes the following steps: S9.1 Interface Integration Content: Measured / estimated temperature curves, residual change curves, thermal state classification labels, anomaly pop-up prompts, parameterized thermal field diagrams, and AI-generated conceptual image partitioning windows; S9.2 Deployment carriers: industrial computers, servers, edge devices, and digital twin platforms; S9.3 Display Specifications: The generated images are for auxiliary reference only. Production control shall be based on actual measured sensor data and process thresholds.

10. A prediction-guided visual representation system for generating the thermal state of a rare earth molten salt electrolytic furnace according to claim 1, used to execute the prediction-guided visual representation method for generating the thermal state of a rare earth molten salt electrolytic furnace according to any one of claims 1-9, characterized in that, Specifically, it includes: The data acquisition module is used to collect industrial time-series data such as temperature, voltage, current, material load, and production records. The data preprocessing module is used to perform time alignment, missing value completion, and outlier handling. The feature construction module is used to construct features such as power input, historical lag, temperature gradient, thermal momentum, cumulative energy, and transient heat mass ratio; Temperature state estimation module, used to output temperature estimate values; The residual calculation module is used to calculate the prediction residual between the actual temperature and the temperature estimate. The thermal state coding module is used to map temperature estimates, prediction residuals, and key operating condition variables into visual semantics. The predictive prompt word generation module is used to generate structured prompt words; The visual representation generation module is used to generate parametric thermal field maps and generative thermal state images; The digital twin display module is used to display temperature status, residual signals, status categories, and generative visual representation results.

Images