Method for optimizing a slag metal recovery process based on digital twinning

Abstract

The present application relates to the technical field of computing, and discloses a slag metal recovery process optimization method based on digital twinning, which comprises a data acquisition module, a data-driven model construction module, a digital twin construction module, a synchronous operation module, an online calibration module, a simulation optimization module, an online operation module and a model self-learning module; the present application directly pushes the recovery rate of each batch of materials to the theoretical limit through forward-looking simulation optimization; cost reduction and efficiency improvement are realized from the system level through mixed models and multi-objective optimization; process standardization and intelligent control are realized through digital twinning and automatic execution, significantly improving the stability and consistency of product quality; by adjusting model parameters and process constraints, the smelting plant of different scales and different raw material sources can be quickly adapted, which has strong universality, is easy to standardize and productize, has high commercial expansion potential and huge market space.

Description

Technical Field

[0001] This invention relates to the field of computing technology, and more specifically to an optimization method for slag metal recovery processes based on digital twins. Background Technology

[0002] Slag is a typical solid waste generated in the metallurgical industry. It often contains a considerable proportion of valuable metals. Efficient recycling of slag is not only an important means of tapping potential economic value, but also an inevitable requirement for realizing resource recycling and reducing environmental burden.

[0003] Production process control and optimization in this field mainly rely on the following two modes: 1. Offline optimization mode based on human experience: Currently, the setting of core process parameters in many production lines heavily depends on the experience of operators. Operators adjust parameters based on the apparent characteristics of the feed slag and limited offline test data. 2. Online control mode based on traditional models: To improve the level of automation, some advanced enterprises have tried to adopt process control systems based on traditional mechanistic models or simple statistical models.

[0004] The existing technology also has the following drawbacks:

[0005] Relying on offline testing results are severely delayed, and adjustments are mostly made after the fact. Real-time intervention during the reaction process is impossible, leading to missed optimization windows and resulting in large fluctuations in recovery rates that are difficult to achieve optimal levels.

[0006] Typically, only a single parameter is simply controlled, ignoring the complex coupling relationship between multiple parameters such as temperature, acid concentration, stirring, and solid content. Often, in order to ensure the recovery rate, the acid concentration or temperature is blindly increased, resulting in huge waste.

[0007] Production quality is highly dependent on the individual condition and experience of experienced workers. Differences in operation between different shifts and different personnel lead to large fluctuations and poor stability in the production process and results, making it difficult to achieve standardization.

[0008] Traditional optimization methods have poor universality. A single model is usually only applicable to a specific factory, specific raw materials, or even specific equipment conditions, making it difficult to replicate and promote.

[0009] Therefore, methods with high recovery rates, reduced consumption, improved stability, and universality are needed to solve the above problems. Summary of the Invention

[0010] In order to overcome the above-mentioned defects of the prior art, the present invention provides an optimization method for slag metal recovery process based on digital twins to solve the problems existing in the background art.

[0011] To achieve the above objectives, the present invention provides the following technical solution: an optimization method for slag metal recovery process based on digital twins, comprising:

[0012] S1. Collect historical data on slag metal recovery through the data acquisition module, including data before, during, and after recovery, and preprocess the data.

[0013] The recovery process data includes reaction temperature, pH value, stirring rate, oxidation-reduction potential (ORP), reaction time, and image data during the process, wherein the image data includes solution color and turbidity;

[0014] S2. Define the input vector and output vector, and use machine learning algorithms based on historical data to learn the complex nonlinear relationship between them. Build a data-driven model through the data-driven model building module.

[0015] S3. Construct a mechanism model and fuse it with the data-driven model. Construct a high-fidelity digital twin of the slag metal recovery process through the digital twin construction module.

[0016] The mechanistic model incorporates leaching reaction kinetic equations as constraints and guidelines. The fusion methods specifically include: a physical constraint loss function: adding a physical regularization term to the loss function of the neural network to penalize predictions that violate the mechanistic equations; hybrid modeling: using the output of the mechanistic model as one or more features, inputting them along with the original features X into the data-driven model for training; symbolic regression: using genetic programming to directly discover equation forms that conform to physical laws from the data; and the learning process of the data-driven model is constrained by the form of the reaction kinetic equations.

[0017] S4. Perform physical entity operation, collect data from the physical reaction vessel in real time through the synchronous operation module, and input the collected data into the digital twin in real time to drive the twin to run synchronously;

[0018] S5. Synchronously collect the measured values ​​of key parameters in the physical entity and the predicted values ​​of the twin, compare them through the online calibration module, and determine the threshold. If the error exceeds the threshold, start the parameter identification algorithm and update the model.

[0019] The startup parameter identification algorithm and model update are specifically as follows:

[0020] J(θ) = Σ[r(t) - y(t;θ)]²;

[0021] Where J(θ) is the loss function, a new set of mechanism parameters θ is found such that the overall error between the predicted sequence of the digital twin and the measured sequence is minimized in the entire recent time window [t0, tc]. y(t;θ) indicates that the predicted value depends on the parameter θ to be adjusted. The extended Kalman filter algorithm is used for calculation, and the parameter identification algorithm outputs a new and optimal set of parameter estimates θ1. The system immediately uses θ1 to back-optimize and adjust the key parameters in the internal mechanism model of the digital twin to update the digital twin. In each batch of processing, the digital twin is trained and corrected in real time. After the update, the digital twin will use this new set of parameters for prediction starting from the current moment.

[0022] S6. Before feeding physical entities, set optimization goals, and use the simulation optimization module to perform multiple simulations within the set operation parameter space to find a set of optimal operation parameters.

[0023] S7. The system directly sends the parameter to the distributed control system, automatically executes the optimal operating parameters of the equipment through the online operation module, and collects the actual results and predicted results for comparison to verify the accuracy and effectiveness of the simulation optimization.

[0024] S8. Store the complete data of this batch as a new sample in the historical database, and conduct the next round of model training through the model self-learning module to enable the digital twin to continue to evolve.

[0025] The technical effects and advantages of this invention are as follows:

[0026] 1. This invention uses forward-looking simulation optimization to predict the outcome and find the optimal solution before the problem occurs, changing "post-event control" to "pre-event optimization" and directly pushing the recovery rate of each batch of materials to the theoretical limit.

[0027] 2. This invention, through hybrid models and multi-objective optimization, can accurately find the Pareto optimal solution among thousands of combinations that can guarantee high recovery rate while simultaneously achieving low acid consumption and low energy consumption, thereby reducing costs and increasing efficiency at the system level.

[0028] 3. This invention, through digital twins and automated execution, solidifies the optimal process into the system, so that the production process no longer depends on personal experience, but is driven by high-fidelity models and algorithms, realizing process standardization and intelligent control, and significantly improving the stability and consistency of product quality.

[0029] 4. By adjusting model parameters and process constraints, this invention can be quickly adapted to smelters of different sizes and with different raw material sources. It is highly versatile, easy to standardize and commercialize, and has high commercial expansion potential and a huge market space. Attached Figure Description

[0030] Figure 1This is a diagram illustrating the method steps of the present invention.

[0031] Figure 2 This is the optimized decision logic diagram of the present invention. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The slag metal recovery process optimization method based on digital twins involved in the present invention is not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Reference Figure 1 This invention provides an optimization method for slag metal recovery process based on digital twins, including a data acquisition module, a data-driven model construction module, a digital twin construction module, a synchronous operation module, an online calibration module, a simulation optimization module, an online operation module, and a model self-learning module.

[0034] The specific implementation steps of this invention include the following steps:

[0035] S1. Collect historical data on slag metal recovery through the data acquisition module, including data before, during, and after recovery, and preprocess the data.

[0036] It should be specifically noted that the historical data on slag metal recovery includes:

[0037] Data prior to recovery: slag composition, slag weight, initial acid concentration, and acid volume, where the slag composition is the content of Cu, Zn, and Fe, obtained from an XRF fluorescence analyzer;

[0038] Recovery process data: reaction temperature, pH value, stirring rate, redox potential (ORP), reaction time, and image data during the process, including solution color and turbidity;

[0039] Post-recovery data: final metal yield, final pH value, tailings composition, and energy consumption data.

[0040] The data preprocessing specifically includes:

[0041] Data cleaning: handling missing values ​​and removing outliers. Interpolation is used to handle missing values, and the 3σ criterion or interquartile range method is used to remove outliers.

[0042] Data normalization / standardization: Scaling feature data of different dimensions and orders of magnitude to the same interval [0, 1] or with a mean of 0 and a variance of 1, using Min-Max normalization or Z-Score standardization to accelerate model convergence and improve model accuracy.

[0043] S2. Define the input vector and output vector, and use machine learning algorithms based on historical data to learn the complex nonlinear relationship between them. Build a data-driven model through the data-driven model building module.

[0044] It should be specifically noted that the input vector is as follows:

[0045] X = [C, T, V, t];

[0046] Where X is the input vector, X is the slag composition, such as Cu: 5%, Zn: 3%, Fe: 35%, C is the initial acid concentration, T is the reaction temperature, V is the stirring rate, and t is the reaction time;

[0047] The output vector is specifically:

[0048] Y = [H, ZpH, N];

[0049] Where Y is the output vector, H is the recovery rate of the target metal, ZpH is the final pH value, and N is the energy consumption.

[0050] The machine learning algorithms mentioned employ gradient boosting decision trees, random forests, or neural networks. These algorithms excel at learning complex mapping relationships from high-dimensional and non-linear industrial data. Furthermore, gradient boosting decision trees and random forests are insensitive to feature loss and have strong noise resistance.

[0051] The data-driven model building module is trained using a training set containing 1,000 historical production batches of data to establish a model based purely on statistical data patterns, serving as the foundation for the hybrid model.

[0052] S3. Construct a mechanism model and integrate it with the data-driven model. Construct a high-fidelity digital twin of the slag metal recovery process through the digital twin construction module.

[0053] It should be specifically noted that the mechanistic model incorporates a leaching reaction kinetic equation as a constraint and guide; taking copper as an example, the leaching reaction kinetic equation is specifically as follows:

[0054] ;

[0055] Where d[Cu] / dt is the copper leaching rate, [Cu] is the copper ion concentration (from collected data), k is the reaction rate constant (from historical data fitting or literature), [H+] is the hydrogen ion concentration (calculated using pH value, [H+] = 10^(-pH), n is the reaction order (an empirical constant representing the sensitivity of the reaction rate to acid concentration), and S is the ore surface area (estimated as a relative surface area value using particle size analysis data of the feed slag and slurry density).

[0056] The model fusion method specifically includes:

[0057] Physical constraint loss function: Add a physical regularization term to the loss function of the neural network to penalize predictions that violate the mechanistic equations;

[0058] Hybrid modeling: The output of the mechanistic model is used as one or more features, which are then input together with the original features X into the data-driven model for training;

[0059] Symbolic regression: Using genetic programming methods to directly discover equations that conform to physical laws from data.

[0060] It should be explained that the learning process of the data-driven model is constrained by the above-mentioned mechanistic equations, so that its prediction results conform to both data patterns and physical and chemical laws, thereby improving the model's generalization ability and interpretability.

[0061] S4. Perform physical entity operation, collect data from the physical reaction vessel in real time through the synchronous operation module, and input the collected data into the digital twin in real time to drive the twin to run synchronously.

[0062] It should be specifically noted that the physical reaction vessel's status data is collected in real time using sensors (temperature, pH, ORP sensors), actuators (valve, frequency converter status), and a machine vision system (high-definition camera, spectrometer); the real-time data collection from the physical reaction vessel specifically includes:

[0063] Feed data: slag weight, preliminary composition data (XRF), and acid concentration and volume;

[0064] Dissolution / leaching reaction data: temperature, pH, stirring rate, redox potential (ORP), and reaction time;

[0065] Solid-liquid separation data: filtration rate and filter cake appearance;

[0066] Metal recovery data: electrolysis current / voltage, amount of precipitant added, final metal yield, and tailings composition.

[0067] It is important to note that the dissolution / leaching reaction data is obtained by capturing changes in solution color, turbidity, and precipitation in real time, and extracting image features into numerical vectors; the solid-liquid separation data is obtained by scanning the filter residue, analyzing its color and texture, and assessing whether the leaching is complete.

[0068] The collected real-time data, including numerical feature vectors extracted from images, is input into the digital twin, driving it to run in parallel and synchronously with the physical entity, thus achieving virtual-real mapping.

[0069] Reference Figure 2 S5. Synchronously collect the measured values ​​of key parameters in the physical entity and the predicted values ​​of the twin, compare them through the online calibration module, and determine the threshold. If the error exceeds the threshold, the parameter identification algorithm is activated and the model is updated.

[0070] It should be noted that not all data is used for comparison. Only key state variables that are highly sensitive to the model and easy to measure in real time are selected for comparison. These key parameters include pH value, oxidation-reduction potential (ORP), and color feature values ​​extracted from machine vision.

[0071] The comparison and threshold determination via the online calibration module specifically involves:

[0072] At each sampling time t, the system calculates the error between the measured value and the predicted value, specifically:

[0073] e(t) = |r(t) - y(t)|;

[0074] Where e(t) is the prediction error, r(t) is the measured value of the key parameter, and y(t) is the predicted value of the key parameter.

[0075] During the initial stable operation period, the system calculates the mean μ and standard deviation σ of the prediction error of key parameters; the threshold is set to μ+n*σ, where n is an adjustable parameter (usually 2 or 3). When e(t) > μ+n*σ, the system considers the error to be significantly deviating from the normal fluctuation range and starts the parameter identification algorithm.

[0076] For example: if the mean value of e_pH is calculated to be μ=0.05 and the standard deviation is σ=0.03, then the dynamic threshold can be set to 0.05+2*0.03=0.11. This means that if e_pH>0.11 at a certain moment, the error exceeds the threshold.

[0077] It should be specifically explained that the startup parameter identification algorithm and model update are as follows:

[0078] J(θ) = Σ[r(t) - y(t;θ)]²;

[0079] Where J(θ) is the loss function, a new set of mechanistic parameters θ is found such that the overall error between the predicted sequence of the digital twin and the measured sequence is minimized within the entire recent time window [t0,tc]. y(t;θ) indicates that the predicted value depends on the parameter θ to be adjusted. The Extended Kalman Filter (EKF) algorithm is used for calculation. The Kalman filter can not only estimate the state, but also estimate the parameters of the model at the same time. It treats the parameter θ as a state variable with a rate of change of zero, and then performs optimal estimation through the filtering algorithm.

[0080] The parameter identification algorithm outputs a new set of optimal parameter estimates θ1. The system immediately uses θ1 to back-optimize and adjust the key parameters in the internal mechanism model of the digital twin, updating the digital twin. In each batch of processing, the digital twin is trained and corrected in real time, so that the prediction curve of the digital twin quickly approaches the real curve. After the update, the digital twin will use this new set of parameters to make predictions from the current moment, and its prediction results will be closer to the real behavior of the physical entity.

[0081] S6. Set optimization goals, and use the simulation optimization module to perform multiple simulations within the set operating parameter space to find a set of optimal operating parameters.

[0082] It should be specifically noted that the simulation optimization steps are as follows:

[0083] A1. Determine the decision variables, i.e. the process parameters to be optimized;

[0084] A2. Set the main optimization objective and constraints, and perform multi-objective optimization;

[0085] A3. Select an optimization algorithm, specifically including:

[0086] Bayesian optimization algorithm constructs a surrogate model to approximate the digital twin model, predicts the performance of unknown points based on the evaluated points, and uses a sampling function to determine the next parameter combination most likely to find the global optimum; it is suitable for simulations with high computational cost, can find the global optimum with the fewest number of simulations, and is highly efficient.

[0087] Genetic algorithms simulate natural selection by encoding parameters into chromosomes. Through operations such as selection, crossover, and mutation, they evolve to produce better combinations of parameters generation after generation. They have strong global search capabilities, are easy to parallelize, and can handle a large number of simulations simultaneously.

[0088] Particle swarm optimization (PSO) simulates bird flock behavior, where each particle represents a combination of parameters and updates its position (parameters) by tracking the individual's historical best and the group's historical best; it has fewer parameters and a faster convergence speed.

[0089] A4. Run the simulation. The algorithm randomly selects the first batch of parameter combinations (e.g., 10 groups). The digital twin performs a fast simulation on these 10 groups of parameters in parallel to obtain the optimization target result for each group.

[0090] Bayesian optimization updates the surrogate model based on these 10 sets of results and calculates the next most worthwhile parameter point for simulation using the acquisition function, repeating this process; the genetic algorithm performs selection, crossover, and mutation based on the performance (i.e., fitness function) of these 10 sets to generate the next generation of 10 sets of parameters, and repeats the simulation.

[0091] The optimization stops when the maximum number of simulations (e.g., 10,000) is reached, a satisfactory solution is found, or performance cannot be improved after multiple iterations.

[0092] A5. Generate the optimal operating parameters and output the optimal parameters and their prediction results.

[0093] S7. The system directly sends the parameter to the distributed control system, which automatically executes the equipment's set value through the online operation module, and collects the actual results and predicted results for comparison to verify the accuracy and effectiveness of the simulation optimization.

[0094] It should be specifically noted that, through industrial communication protocols, the optimized parameter set obtained is directly sent from the digital twin system to the distributed control system (DCS) or programmable logic controller (PLC). The DCS / PLC automatically executes the adjustments to actuators such as acid pump flow, heater power, and stirrer speed to form a closed-loop optimization control. After the batch is completed, the actual recovery rate is compared with the predicted recovery rate, the error is calculated, and the effect of this optimization is evaluated.

[0095] The specific steps for verifying the accuracy and effectiveness of the simulation optimization are as follows:

[0096] B1. After the batch ends, the system collects data from different sources, including:

[0097] The predicted data Yp of the digital twin is the predicted result vector obtained after simulating the optimal parameter set in step S6;

[0098] The actual data Ya from the on-site DSC / MES / laboratory is the final result vector after the actual operation of this batch.

[0099] B2. Calculate the absolute error and relative error for each key performance indicator. Absolute error = |predicted value - actual value|, relative error = (|predicted value - actual value| / actual value) × 100%.

[0100] B3. Set acceptance criteria and judgment results. Enterprises set acceptable error ranges based on their own process level, and the system will make automatic judgments based on this. If the errors of all key indicators are within the allowable range, the optimization is judged to be successful and effective. If any indicator (especially the main target) is seriously out of tolerance, it is marked as needing investigation.

[0101] B4. Root Cause Analysis: If verification reveals significant deviations, the system will assist engineers in conducting root cause analysis, which may include:

[0102] Model issues: Check the S5 model update log: During batch runs, are model updates frequently initiated? Have parameters been drastically adjusted? This could indicate significant bias in the model from the outset; Analyze outliers: Are there any new operating conditions that the model has never encountered? The current hybrid model may not be able to cover such new situations.

[0103] Data issues: Sensor failure: Check whether key sensors (such as pH meters and flow meters) are working properly during operation, and whether the data fluctuates or remains unchanged for a long time; Actuator error: Are the set values ​​issued by the DCS accurately executed? Deviation of the actuator is a common source of error.

[0104] Process and Raw Material Issues: Raw Material Fluctuations: Is the composition of the feed slag consistent with the assumptions made during the S6 simulation? Are there any other interfering components that suddenly increased, affecting the reaction, and which the model did not consider? Unmeasurable Interferences: Are there any interfering factors such as ambient temperature and humidity that were not included in the model?

[0105] B5. Generate verification reports and visualizations;

[0106] The report includes a review of the optimization goals, a comparison table of recommended parameters vs. actual parameters, a comparison table of predicted results vs. actual results, a conclusion of success or failure, and possible root causes if failure occurs.

[0107] The visualization charts include: radar charts: used to compare the predicted and actual values ​​of multiple indicators; trend comparison charts: plotting the predicted and actual change curves of important parameters throughout the entire batch process, clearly showing at which stage the discrepancy occurred.

[0108] S8. Store the complete data of this batch as a new sample in the historical database, and conduct the next round of model training through the model self-learning module to enable the digital twin to continue to evolve.

[0109] It should be specifically noted that the continuous evolution of the digital twin includes:

[0110] Data archiving: The complete "input-output" data of this batch (including optimization operations and final results) is stored in the historical database as a new high-quality sample.

[0111] Periodic retraining: After accumulating a certain amount of new data (e.g., 50 batches), retrain the models in S2 and S3 using the entire expanded dataset, and thoroughly update the model weights.

[0112] Online learning: For linear models or simple neural networks, stochastic gradient descent (SGD) and other methods can be used for online incremental updates.

[0113] It needs to be explained that this is the key to the system's long-term vitality. Through continuous learning, the digital twin can adapt to the slow changes in production processes (such as equipment wear and tear, and changes in raw material sources), becoming increasingly intelligent and precise.

[0114] Through the above description of the embodiments, those skilled in the art can clearly understand that the various embodiments of this application can be implemented by means of software or software combined with necessary general-purpose hardware platforms, and of course, they can also be implemented by hardware functions. Based on this understanding, the technical solution of this application, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. The software product is stored in a storage medium and includes several instructions to cause a computer device, such as including but not limited to a personal computer, server, or network device, to execute all or part of the steps of the method described in any embodiment of this application.

[0115] The foregoing has described exemplary embodiments of this application. It should be understood that the above exemplary embodiments are not restrictive but illustrative, and the scope of protection of this application is not limited thereto. It should be understood that those skilled in the art can make modifications and variations to the embodiments of this application without departing from the spirit and scope of this application, and such modifications and variations should be within the scope of protection of this application.

Claims

1. An optimization method for slag metal recovery process based on digital twins, characterized in that, Specifically, the following steps are included: S1. Collect historical data on slag metal recovery through the data acquisition module, including data before, during, and after recovery, and preprocess the data. The recovery process data includes reaction temperature, pH value, stirring rate, redox potential (ORP), reaction time, and image data during the process, wherein the image data includes solution color and turbidity; S2. Define the input vector and output vector, and use machine learning algorithms based on historical data to learn the complex nonlinear relationship between them. Build a data-driven model through the data-driven model building module. S3. Construct a mechanism model and fuse it with the data-driven model. Construct a high-fidelity digital twin of the slag metal recovery process through the digital twin construction module. The mechanistic model incorporates leaching reaction kinetic equations as constraints and guidelines. The fusion methods specifically include: a physical constraint loss function: adding a physical regularization term to the loss function of the neural network to penalize predictions that violate the mechanistic equations; hybrid modeling: using the output of the mechanistic model as one or more features, inputting them along with the original features X into the data-driven model for training; symbolic regression: using genetic programming to directly discover equation forms that conform to physical laws from the data; and the learning process of the data-driven model is constrained by the form of the reaction kinetic equations. S4. Perform physical entity operation, collect data from the physical reaction vessel in real time through the synchronous operation module, and input the collected data into the digital twin in real time to drive the twin to run synchronously; S5. Synchronously collect the measured values ​​of key parameters in the physical entity and the predicted values ​​of the twin, compare them through the online calibration module, and determine the threshold. If the error exceeds the threshold, start the parameter identification algorithm and update the model. The startup parameter identification algorithm and model update are specifically as follows: J(θ) = Σ[r(t) - y(t;θ)]²; Where J(θ) is the loss function, a new set of mechanism parameters θ is found such that the overall error between the predicted sequence of the digital twin and the measured sequence is minimized in the entire recent time window [t0, tc]. y(t;θ) indicates that the predicted value depends on the parameter θ to be adjusted. The extended Kalman filter algorithm is used for calculation, and the parameter identification algorithm outputs a new and optimal set of parameter estimates θ1. The system immediately uses θ1 to back-optimize and adjust the key parameters in the internal mechanism model of the digital twin to update the digital twin. In each batch of processing, the digital twin is trained and corrected in real time. After the update, the digital twin will use this new set of parameters for prediction starting from the current moment. S6. Before feeding physical entities, set optimization goals, and use the simulation optimization module to perform multiple simulations within the set operation parameter space to find a set of optimal operation parameters. S7. The system directly sends the parameter to the distributed control system, automatically executes the optimal operating parameters of the equipment through the online operation module, and collects the actual results and predicted results for comparison to verify the accuracy and effectiveness of the simulation optimization. S8. Store the complete data of this batch as a new sample in the historical database, and conduct the next round of model training through the model self-learning module to enable the digital twin to continue to evolve.

2. The method for optimizing the slag metal recovery process based on digital twins according to claim 1, characterized in that: The historical data on slag metal recovery specifically includes: Data before recovery: slag composition, slag weight, initial acid concentration, and acid volume, where the slag composition is the content of Cu, Zn, and Fe, obtained from an XRF fluorescence analyzer; Data during recovery: reaction temperature, pH value, stirring rate, redox potential, reaction time, and image data during the process, including solution color and turbidity; Data after recovery: final metal yield, final pH value, tailings composition, and energy consumption data.

3. The method for optimizing slag metal recovery process based on digital twins according to claim 1, characterized in that: The input vector is specifically: X = [C, T, V, t]; where X is the input vector, X is the slag composition, C is the initial acid concentration, T is the reaction temperature, V is the stirring rate, and t is the reaction time; the output vector is specifically: Y = [H, ZpH, N]; where Y is the output vector, H is the recovery rate of the target metal, ZpH is the final pH value, and N is the energy consumption. The data-driven model building module is trained using a training set containing 1,000 historical production batches of data to establish a model based purely on statistical patterns in the data.

4. The method for optimizing the slag metal recovery process based on digital twins according to claim 1, characterized in that: The comparison and threshold determination via the online calibration module specifically involves: At each sampling time t, the system calculates the error between the measured and predicted values ​​of key parameters. These key parameters include pH value, redox potential, and color feature values ​​extracted from machine vision. The error calculation is as follows: e(t) = |r(t) - y(t)|; Where e(t) is the prediction error, r(t) is the measured value of the key parameter, and y(t) is the predicted value of the key parameter; during the initial stable operation period, the system calculates the mean μ and standard deviation σ of the prediction error of the key parameter; the threshold is set to μ+n*σ, where n is an adjustable parameter. When e(t) > μ+n*σ, the system considers the error to be significantly deviating from the normal fluctuation range, starts the parameter identification algorithm and updates the model.

5. The method for optimizing slag metal recovery process based on digital twins according to claim 1, characterized in that: The specific steps of the simulation optimization are as follows: A1. Determine the decision variables, i.e. the process parameters to be optimized; A2. Set the main optimization objective and constraints, and perform multi-objective optimization; A3. Select an optimization algorithm, including: Bayesian optimization algorithm, genetic algorithm and particle swarm optimization algorithm; A4. Run the simulation. The algorithm randomly selects the first batch of parameter combinations, and the digital twin performs rapid simulations on these 10 sets of parameters in parallel to obtain the optimization target result for each set. Bayesian optimization updates the surrogate model based on these 10 results and calculates the next most worthwhile parameter point to simulate through the acquisition function, repeating this process. The genetic algorithm selects, crossovers, and mutates based on the performance (fitness function) of these 10 sets to generate the next generation of 10 sets of parameters, and repeats the simulation. Optimization stops when the maximum number of simulations is reached, a satisfactory solution is found, or performance cannot be improved after multiple consecutive iterations. A5. Generate the optimal operating parameters and output the optimal parameters and their prediction results.

6. The method for optimizing slag metal recovery process based on digital twins according to claim 1, characterized in that: The specific steps for verifying the accuracy and effectiveness of the simulation optimization are as follows: B1. After the batch is completed, the system collects data from different sources, including: the predicted data Yp of the digital twin, which is the predicted result vector obtained after simulating the optimal parameter set in step S6; and the actual data Ya from the field DSC / MES / laboratory, which is the final result vector after the actual operation of the batch. B2. Calculate the absolute error and relative error for each key performance indicator. Absolute error = |predicted value - actual value|, relative error = (|predicted value - actual value| / actual value) × 100%. B3. Set acceptance criteria and judgment results. Enterprises set acceptable error ranges based on their own process level, and the system will make automatic judgments based on this. If the errors of all key indicators are within the allowable range, the optimization is judged to be successful and effective. If any indicator is seriously out of tolerance, it is marked as needing investigation. B4. Root cause analysis: If significant deviations are found during verification, the system will assist engineers in conducting root cause analysis. B5. Generate verification reports and visualizations; The report includes a review of optimization goals, a comparison table of recommended parameters vs. actual parameters, a comparison table of predicted results vs. actual results, a conclusion of success or failure, and possible root causes if failure occurs. Visual charts include: a radar chart for comparing predicted and actual values ​​of multiple indicators; and a trend comparison chart that plots the predicted and actual change curves of important parameters throughout the entire batch process, clearly showing at which stage the discrepancy occurred.

Images