Dynamic working condition accurate control method in gallium enrichment and impurity removal process

Abstract

The invention discloses a gallium enrichment and impurity removal process dynamic working condition accurate control method, which comprises the following steps: firstly, constructing an initial Gaussian prediction model based on historical working condition data for model prediction control and working condition monitoring; and when working condition switching is monitored, starting to collect data for model updating. In a transition stage, model mismatch causes concept drift of a prediction result and model prediction control performance reduction, and in order to reduce transition period control fluctuation, regulation and control variable input is subjected to tightening constraint according to prediction uncertainty estimation. After a small number of samples are collected, the model updating module obtains a drift matrix through a concept drift correction method, and a new working condition data set with pseudo labels is generated in combination with an original sample set. And then, a model is reconstructed based on the new data set, the working condition is monitored again, and a model prediction controller adaptively relaxes a regulation variable input constraint boundary to ensure accurate control. According to the method, the stability and the control precision of the gallium enrichment and impurity removal process can be ensured under the condition that the working conditions change frequently.

Description

Technical Field

[0001] This invention belongs to the field of industrial control technology, specifically relating to a method for precise control of dynamic operating conditions in a gallium enrichment and impurity removal process. Background Technology

[0002] Gallium in nature mainly exists as a by-product mineral, and currently, the vast majority of gallium globally is recovered and prepared from the alumina mother liquor of the aluminum smelting industry. Typically, a gallium recovery production line mainly includes three core processes: resin adsorption and desorption, enrichment and impurity removal, and electrolytic deposition. The resin adsorption and desorption process uses special resins to selectively adsorb gallium ions from the alumina mother liquor, and then separates gallium from impurities through water washing and acid desorption steps, converting the low-concentration mother liquor into a higher-concentration gallium-containing desorption solution, completing the initial extraction of gallium resources. The enrichment and impurity removal process uses a series of purification methods such as neutralization precipitation, flocculation separation, and oxidation to convert the gallium-containing desorption solution into a clean sodium gallate electrolytic solution. Finally, in the electrolytic deposition process, the prepared sodium gallate solution is used as the electrolyte and passed into the electrolytic cell. Under the action of direct current, a specific current density and cell voltage are controlled, causing gallium ions to be reduced and deposited as liquid crude gallium at the cathode. Among them, the enrichment and impurity removal process is a key link connecting crude extraction and fine electrolysis, and its control performance directly determines the purity of the electrolyte and the efficiency of subsequent electrolysis.

[0003] However, due to fluctuations in the composition of upstream mother liquor, changes in reagent activity, and equipment aging, the enrichment and impurity removal process operates in a dynamic state for extended periods, with frequent occurrences of unknown operating conditions, making precise control of the impurity removal process a challenge. On one hand, the number of samples that can be collected in the early stages of changes in operating conditions is extremely limited, making it very difficult to achieve high-precision model updates quickly under small sample conditions. Traditional data-driven methods typically require accumulating a large amount of new operating condition data to retrain the model, which is insufficient to meet the real-time requirements of industrial sites. Existing transfer learning methods, while utilizing historical knowledge, often perform poorly during the "cold start" phase of new data accumulation. While recursive identification techniques can achieve online parameter estimation, they are sensitive to industrial site noise and converge slowly during sudden changes in operating conditions, easily generating large transient errors, making it difficult for the model to quickly adapt to new conditions. On the other hand, accurately identifying and triggering model updates during operating condition changes is difficult. Existing operating condition identification frameworks primarily rely on point estimates of prediction errors for judgment. This deterministic triggering mechanism is highly susceptible to instantaneous noise interference in the field, leading to false alarms. Furthermore, it may fail to identify operating condition drift in a timely manner when only slight shifts occur in the error distribution, resulting in delayed control strategy switching. In addition, control stability during the model update transition phase is not guaranteed. During the "transition period" when a new operating condition is identified but the model has not yet been updated, the control system is in a high-risk model mismatch state. Existing robust or stochastic model predictive control methods are often overly conservative or make overly ideal assumptions, lacking proactive management and dynamic constraints on model confidence risk during this specific phase. This can easily lead to drastic fluctuations in control inputs, actuator input saturation, or even system instability. Summary of the Invention

[0004] This invention provides a dynamic and precise control method for gallium enrichment and impurity removal processes, which can ensure system stability and control accuracy under frequently changing operating conditions.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solution: A method for precise dynamic control of gallium enrichment and impurity removal process includes: An initial Gaussian prediction model was constructed based on historical operating data of the gallium enrichment and impurity removal process, and it was used for model predictive control and operating condition monitoring. When a change in operating condition is detected, a concept drift correction method is used to obtain a correction matrix, and the correction matrix is ​​used to map the historical operating condition dataset to generate a dataset with pseudo-labels. Using sample data from the transition phase of the operating condition and a dataset with pseudo-labels, a Gaussian prediction model for the new operating condition is trained. A Gaussian prediction model for the new operating conditions is used to perform model prediction control and operating condition monitoring for the gallium enrichment and impurity removal process under the new operating conditions.

[0006] Furthermore, the method for using a Gaussian prediction model for operating condition monitoring is as follows: The control variables for the gallium enrichment and impurity removal process at a preset time delay and the quality indicators of the solution after impurity removal are obtained, input into the Gaussian prediction model, and the quality indicators at the current time are output. Predicted mean and prediction variance ; Based on predicted mean and prediction variance And based on the current moment of the gallium enrichment and impurity removal process Actual values ​​of quality indicators Calculate the probability of switching operating conditions at the current moment. : ; In the formula, The standard normal cumulative distribution function is... The standard normal distribution is at a significance level. The critical value below, For the current moment The standardized residuals of the quality indicators; Let be the confidence radius. ; Probability of operating condition switching With the judgment threshold Comparison, only when At that time, it was determined that the operating conditions of the gallium enrichment and impurity removal process had undergone a substantial change, thereby triggering subsequent model updates.

[0007] Furthermore, a concept drift correction method is used to obtain a correction matrix, and the correction matrix is ​​used to map the historical working condition dataset, including: Step 2.1, establish the prediction model based on the parameter evolution model of sparse affine transformation: ; In the formula, Let T be the parameter vector for the new operating condition. For the parameter vector of the old working condition T-1, Used to characterize global physical changes during the impurity removal process; It is a sparse incremental vector, and the number of its non-zero elements is much smaller than the dimension of the parameter vector. It is used to characterize the local characteristic drift during the impurity removal process; Step 2.2: Substitute the parameter evolution model into the prediction model of the impurity removal process to obtain the output relationship of the new operating condition T: ; In the formula, For the impurity removal process, a predictive model is provided under the new operating condition T. For the basis function vector, For the dataset of the new working condition T; time step The basis function vector is represented as , To regulate the variable at time 1 Forward The nonlinear mapping basis function corresponding to each time delay For quality indicators at any time Forward The nonlinear mapping basis function corresponding to each time delay and These are the time delay orders for the control variables and quality indicators, respectively; Step 2.3, considering the prediction model of the old working condition T-1 for the dataset of the new working condition T. The predicted output is , The prediction model for the old working condition T-1 will be used to predict the actual quality indicators of the new working condition T. Rewritten as about the correction vector The linear regression form: ; In the formula, ; Step 2.4, using Based on the sparsity property, the sparse regression algorithm is used to solve for the optimal correction matrix. : ; in, For regularization parameters; Step 2.5, using the optimal correction matrix The large amount of data accumulated from the old operating condition T-1 Perform feature mapping to generate pseudo-labels that fit the dynamic characteristics of the current operating condition T. : ; Step 2.6: Combine the historical dataset with pseudo-labels with the collected small sample dataset of the real new working conditions T. Merge and build an enhanced dataset .

[0008] Furthermore, during the transition phase of the gallium enrichment and impurity removal process, the constraint boundaries of the control variables are dynamically adjusted during the model predictive control process.

[0009] Furthermore, the constraint boundaries of the control variables are dynamically adjusted, specifically as follows: Step A1: Express the constraint boundaries of the control variables in probabilistic form: ; In the formula, For a moment The regulatory variables, To regulate the variable at time 1 Boundary constraint coefficients, To regulate the variable at time 1 The upper limit of the constraint, Risk coefficient; Step A2: Based on the probability of operating condition switching And an affine transformation strategy is used to calculate the current time. risk coefficient : ; in, and For mapping parameters; Step A3: Use the prediction standard deviation output by the Gaussian prediction model Characterizing the uncertainty of the impurity removal process, combined with the inverse cumulative distribution function of the standard normal distribution. Calculate the dynamic contraction parameters of the control variables. : ; in, The adjustment parameters are used to map the uncertainty of the Gaussian prediction model output to the input domain; Step A4, using dynamic compression parameters The constraint boundaries of the control variables are dynamically adjusted to obtain the tightened constraint conditions for the control variables: .

[0010] Furthermore, by dynamically adjusting the constraint boundaries of the control variables, the gallium enrichment and impurity removal process is subjected to model predictive control, and the optimization problem is formulated as follows: ; In the formula, the optimization objective is... The aim is to find an optimal control sequence. This will make the future The cumulative tracking error within each prediction step and The control increment within each control step is minimized. For quality indicators at any time The set reference value, Gaussian prediction model Quality indicators at any time The predicted value, To regulate the variable at time 1 The control increment, and As a weighting factor, and For the time index of the sample point, and These are the prediction step size and the control step size, respectively. and These are the time delay orders for the control variables and quality indicators, respectively; and These are the upper and lower limits of hard constraints for the control variables.

[0011] Furthermore, sequential quadratic programming or interior point methods are used to solve the optimization problem of model predictive control, thereby obtaining the optimal control sequence of the control variables.

[0012] Furthermore, the regulating variable is the current value used to adjust the acid flow rate.

[0013] Furthermore, the pH value of the gallium desorption solution was used as a quality indicator.

[0014] The method for precise dynamic control of the gallium enrichment and impurity removal process of the present invention has the following advantages over the prior art: To address the challenge of identifying mismatches between historical prediction models and current dynamic operating conditions, an online operating condition identification method based on Gaussian process regression uncertainty estimation is proposed. By providing real-time operating condition switching probability estimates through Gaussian process regression, this method overcomes the susceptibility to noise interference inherent in traditional error threshold judgments, enabling accurate identification of operating condition changes and timely triggering of prediction model updates.

[0015] To address the scarcity of high-value samples in the initial learning phase of predictive models for new operating conditions, a rapid concept drift correction method is proposed. By constructing auxiliary samples with pseudo-labels using a small amount of new operating condition data and historical predictive models, rapid updates to the new operating condition model are achieved.

[0016] Finally, to address the system stability issues during the model update transition phase, an adaptive constraint-tightening model predictive control method is proposed. By introducing an adaptive scaling factor related to the model confidence level, the boundary constraints of model predictive control are dynamically tightened and relaxed, overcoming the risk of fixed constraints being too conservative or aggressive during the transition period, and achieving a dynamic balance between stability and control accuracy throughout the gallium enrichment and impurity removal process. Attached Figure Description

[0017] Figure 1 This application provides a precise control framework for the dynamic operating conditions of the gallium enrichment and impurity removal process.

[0018] Figure 2 This is a schematic diagram illustrating the adaptive opportunity constraint contraction predictive control principle of an embodiment of this application.

[0019] Figure 3These are the experimental results of online control of the gallium enrichment and impurity removal process in the embodiments of this application. The sub-figures (a), (b), (c), (d), and (e) correspond to the output quantity, control tracking error, model prediction error, control variable, and operating condition switching frequency, respectively.

[0020] Figure 4 Comparison of anti-interference results for working condition identification: SNR is the signal-to-noise ratio. Subgraph (a) shows the F1 score under different noise intensities, and subgraph (b) shows the false trigger rate under different noise intensities.

[0021] Figure 5 The following are the model update comparison results of the embodiments of this application; where subgraph (a) is the root mean square error of the updated model; and subgraph (b) is the prediction determination coefficient of the updated model.

[0022] Figure 6 This is the online control result of the method described in the embodiments of this application and other comparative methods. Detailed Implementation

[0023] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.

[0024] To address the challenge of dynamic operating condition control during gallium enrichment and purification processes due to fluctuations in raw material composition, this embodiment provides a method for precise dynamic operating condition control in the gallium enrichment and purification process, such as... Figure 1 As shown, it consists of three parts: working condition identification, model update, and predictive control, which work together to achieve precise control of the gallium enrichment and impurity removal process.

[0025] First, an initial Gaussian prediction model is constructed based on historical operating condition data for model predictive control and operating condition monitoring. When an operating condition change is detected, data collection begins for model updates. During the transition phase, model mismatch leads to conceptual drift in the prediction results, resulting in a decline in model predictive control performance. To reduce control fluctuations during the transition period, tighter constraints are applied to the input of the control variables based on the prediction uncertainty estimate. After collecting a small number of samples, the model update module obtains the drift matrix using a conceptual drift correction method and combines it with the original sample set to generate a new operating condition dataset with pseudo-labels. Subsequently, the model is reconstructed based on the new dataset, and operating conditions are monitored again. At this point, the uncertainty caused by model mismatch in the predictive model has been largely eliminated, and the model predictive controller adaptively relaxes the input constraint boundaries of the control variables to ensure precise control.

[0026] Step 1, Predictive Modeling and Operating Condition Monitoring: An initial Gaussian prediction model is constructed based on historical operating condition data of the gallium enrichment and impurity removal process, and is used for model predictive control and operating condition monitoring.

[0027] To address the difficulty in identifying mismatches between historical models and current dynamic operating conditions, this invention proposes an online operating condition identification method based on Gaussian process regression uncertainty estimation. This method provides real-time operating condition switching probability estimates through Gaussian process regression, overcoming the susceptibility to noise interference in traditional error threshold judgments. It achieves accurate identification of operating condition changes and timely triggers the prediction model update process. Specifically, the implementation steps are as follows: Step 1.1: Construct a prediction model for the impurity removal process.

[0028] During the offline training phase, the gallium enrichment and impurity removal process was collected under operating conditions. The Gaussian process prediction model was trained using historical data. During online identification, an input vector containing impurity removal control variables and key solution component indicators (such as pH value) is constructed based on process time delay information. .in Represents the control variables in the impurity removal process. This represents the key quality indicators of the solution after impurity removal. and These are the delay orders for the input and output, respectively.

[0029] Through the trained model Obtain the predicted mean of the system output at the current time. and prediction variance : ; Step 1.2: Define the dynamic confidence interval.

[0030] Define confidence radius ,in The standard normal distribution is at a significance level. The critical value is then determined. This leads to the dynamic confidence interval that varies with the uncertainty of the impurity removal process. In actual impurity removal production, due to possible changes in operating conditions, the actual measured values ​​of impurity removal indicators may vary. Compared with the model's predicted mean The deviation can be categorized into three situations: Case 1: Actual value Located within the confidence interval centered on the predicted value, i.e. This indicates that the current prediction model matches the actual impurity removal process, the reaction mechanism has not changed significantly, and the probability of a change in operating conditions is low.

[0031] Scenario 2: Actual value Below the lower bound of the confidence interval, i.e. This indicates that the impurity removal model does not match the actual system, the impurity removal reaction characteristics may have changed, and the further the actual value is from the lower boundary of the confidence interval, the greater the probability of the operating condition switching.

[0032] Case 3: Actual value Above the upper bound of the confidence interval, i.e. Similar to Case 2, the further the actual value is from the upper boundary of the confidence interval, the greater the probability of the corresponding operating condition switching.

[0033] Based on the above analysis, this paper selects the actual value. The degree of deviation from the confidence interval is used to quantify the probability of a change in operating conditions.

[0034] Taking scenario 2 as an example, the probability of switching operating conditions is calculated as follows: ; in, To standardize the residuals, It is the standard normal cumulative distribution function.

[0035] Step 1.3: Estimation of operating condition switching probability.

[0036] Finally, considering scenarios 1, 2, and 3 together, the probability of operating condition switching is... The complete expression is: ; To further reduce the false triggering rate caused by liquid level fluctuations or bubble interference, this embodiment of the invention sets a judgment threshold. Only when Only when this happens is it determined that the operating conditions of the gallium enrichment and impurity removal process have undergone a substantial change, thereby triggering the subsequent prediction model update process.

[0037] Step 2, Fast Concept Shift Correction and Prediction Model Update: When a change in operating condition is detected, the concept shift correction method is used to obtain the correction matrix, and the correction matrix is ​​used to map the historical operating condition dataset to generate a dataset with pseudo-labels; and the Gaussian prediction model for the new operating condition is trained using the sample data of the operating condition transition phase and the dataset with pseudo-labels.

[0038] To address the problem of scarce high-value samples in the initial stage of learning new operating conditions, this invention proposes a rapid concept drift correction method. This method utilizes the mapping relationship between prior knowledge of historical models and a small amount of new data to expand the sample and enable rapid model adaptation.

[0039] Step 2.1: Establish a parameter evolution model based on sparse affine transformation.

[0040] Considering the continuity of the gallium enrichment and impurity removal process, it is assumed that the evolution of model parameters between adjacent operating conditions follows a sparse affine transformation law. Specifically, the parameter vector under the new operating condition... With the old operating condition parameter vector The following relationship exists between them: ; in Used to characterize global physical changes during the impurity removal process, such as the overall activity decay of the agent or changes in system gain. It is a sparse incremental vector, and the number of its non-zero elements is much smaller than the dimension of the parameter vector. It is used to characterize local characteristic drift caused by sensor zero-point drift or fluctuations in specific impurity components.

[0041] Step 2.2: Construct a drift correction solution problem based on sparse regression.

[0042] Substituting the above parameter evolution relationship into the system model of the impurity removal process, the output relationship under the new operating condition is derived: ; in , and This is a nonlinear mapping basis function. Considering the old operating conditions, the predicted output is... The above relationship can be rewritten as relating the correction vector. The linear regression form: ; in . use Based on the sparsity property, the sparse regression algorithm is used to solve for the optimal correction matrix. : ; in, This is the regularization parameter. In practical applications, it can be applied to the basis functions. Perform a simplified approximation, such as directly taking... To reduce computational complexity.

[0043] Step 2.3: Rapid reconstruction of the model based on pseudo-labels.

[0044] Obtain the optimal correction matrix Subsequently, it was used as a bridge connecting historical knowledge with new working conditions, utilizing the large amount of data accumulated under historical working conditions. Perform feature mapping to generate pseudo-labels that fit the dynamic characteristics of the current operating conditions. : ; Subsequently, the generated historical data with pseudo-labels was merged with a small amount of newly collected real working condition data to construct an augmented dataset. Based on this augmented dataset, the Gaussian prediction model was improved. Retraining is performed to enable the impurity-removing prediction model to quickly adapt to and accurately correct new operating conditions under conditions of scarce samples.

[0045] Step 3: Model predictive control and operating condition monitoring under the new operating conditions: Use the Gaussian prediction model under the new operating conditions to perform model predictive control and operating condition monitoring for the gallium enrichment and impurity removal process under the new operating conditions.

[0046] Considering that although the prediction model is rapidly updated during the initial switching phase of the enrichment and impurity removal process, the new prediction model has not yet fully converged to the current operating conditions. This makes traditional predictive control prone to causing the amount of impurity removal reagent added to reach the physical boundary due to model errors, leading to input saturation or even system oscillation, thus affecting the stability of the impurity removal reaction. Therefore, in a better embodiment, the uncertainty of model prediction is introduced into the control framework, proposing an adaptive chance constraint contraction method. By dynamically adjusting the constraint boundary of the control input, a balance between system stability and control accuracy is achieved. A schematic diagram of the algorithm principle is shown below. Figure 2 As shown.

[0047] Step A1: Construct an opportunity constraint model based on risk coefficients. Transform the hard constraint boundary of the control variables into an opportunity constraint form to explicitly address the impact of prediction uncertainty on constraint satisfaction.

[0048] Specifically, the control input constraints of the impurity removal process are expressed in probabilistic form: ; in These are the boundary constraint coefficients. To constrain the upper limit, Let be the risk coefficient. This formula indicates that the probability of the control input satisfying the constraints must be greater than the set risk coefficient.

[0049] Step A2: Establish the mapping relationship between the operating condition switching probability and the risk coefficient. This is to map the operating condition switching probability output by the operating condition identification module... Mapping to the risk coefficient interval, this paper uses an affine transformation strategy to calculate the risk coefficient at the current time. : ; in, and For mapping parameters (e.g., take) The mechanism works by: when When the temperature rises, it indicates that the impurity removal process is in the transition phase of operating condition switching, and the uncertainty increases significantly. The system automatically adjusts the temperature accordingly. Tighten constraints; when the model update is complete and the operating conditions tend to stabilize, As it approaches zero, the system automatically reduces... To relax constraints and restore the normal predictive control operating mode.

[0050] Step A3: Calculate the boundary contraction of the control input. Based on the risk coefficient and the uncertainty of the model prediction, calculate the boundary contraction of the control input. The predicted standard deviation of the output is used in Gaussian process regression. Characterizing the uncertainty of the impurity removal process, combined with the inverse cumulative distribution function of the standard normal distribution. The dynamic compression parameters are calculated as follows: ; in, The adjustment parameters are used to map the output uncertainty to the input domain. Based on this, the physical boundary of the impurity removal agent addition amount is dynamically corrected to obtain the compressed control constraints: ; This ensures that during the transition period when the risk of model mismatch is high, the control system has reserved sufficient safety margin to prevent excessive or insufficient addition of reagents due to prediction errors.

[0051] Step A4: Solving the rolling optimization problem with adaptive tightening constraints. Combining the uncertainty modeling and constraint tightening strategies described above, the predictive control optimization problem with an adaptive chance-constrained tightening mechanism in the transition phase can be formulated as follows: ; in, Reference values ​​for setting key indicators of impurity removal. For predicted values, To control the incremental amount of impurity remover, , Weighting factor and These are the prediction step size and the control step size, respectively. Optimization objective. The aim is to find an optimal control sequence. This will make the future The cumulative tracking error within each prediction step and Minimize the control increment within each control step, ultimately controlling the sequence. The first It acts on the system, and the above process repeats in the next moment.

[0052] The aforementioned optimization problem is a stochastic nonlinear optimization problem with time-varying probabilistic constraints. This invention transforms the probabilistic constraints into time-varying contractile constraints through a chance-constraint deterministic transformation technique. The solution method often employs numerical methods such as sequential quadratic programming and interior-point methods. In this embodiment, the sequential quadratic programming method is used to solve the rolling optimization problem. Its good convergence performance and real-time computation characteristics make it suitable for online control, thereby achieving precise control of the gallium enrichment and impurity removal process while ensuring the system's stable operation during the transition phase.

[0053] To verify the effectiveness of the method of the present invention, an experiment was conducted on the key process of enrichment and impurity removal—pH neutralization precipitation process. The control variable was the current value used to adjust the acid flow rate, in mA, and the control output was the pH value of the gallium desorption solution during the enrichment and impurity removal process.

[0054] Online control experiment results are as follows Figure 3 As shown. In the initial stage of control, the model matches the actual process, the model prediction is accurate, and the control error is minimal. After 25 minutes, the operating conditions change dynamically, the model mismatches, and the prediction error suddenly increases. At this point, the probability of operating condition monitoring is close to 1. The change in operating conditions is detected, triggering a model update and shrinking the control boundary, reducing overshoot in the transition phase. After the model update is completed, the model and operating conditions are re-adapted, and the control input boundary adaptively relaxes to restore the normal high-precision control mode.

[0055] To further verify the superiority of the method of this invention, comparisons were made with existing advanced methods in terms of working condition identification, rapid model update, and online control. The results of working condition identification, rapid model update, and online control are as follows: Figure 4 As shown in Figures 5 and 6.

[0056] Figure 4 This indicates that the method of the present invention has a higher accuracy in identifying operating conditions and a stronger ability to resist noise interference.

[0057] Figure 5 This demonstrates that the model update method of the present invention can achieve rapid model updates using fewer samples compared to other advanced methods.

[0058] Figure 6 The results in Table 1 show that the method of the present invention has better overall control performance in the pH neutralization process of gallium enrichment and impurity removal. The average control accuracy (MAE1) of the method of the present invention is improved by more than 8.2%, the overshoot is reduced by more than 40.71%, and the steady-state error (SSE) is reduced by more than 47.46%.

[0059]

[0060] Therefore, the method of the present invention can be used for precise control of the gallium enrichment and impurity removal process, which can significantly improve control accuracy and reduce control overshoot. Specifically, compared with the existing advanced methods, the average control accuracy is improved by more than 8.2%, the overshoot is reduced by more than 40.71%, and the steady-state error is reduced by more than 47.46%.

[0061] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, these changes or improvements should fall within the scope of protection claimed in this application.

Claims

1. A method for precise dynamic control of gallium enrichment and impurity removal process, characterized in that, include: An initial Gaussian prediction model was constructed based on historical operating data of the gallium enrichment and impurity removal process, and it was used for model predictive control and operating condition monitoring. When a change in operating condition is detected, a concept drift correction method is used to obtain a correction matrix, and the correction matrix is ​​used to map the historical operating condition dataset to generate a dataset with pseudo-labels. Using sample data from the transition phase of the operating condition and a dataset with pseudo-labels, a Gaussian prediction model for the new operating condition is trained. A Gaussian prediction model for the new operating conditions is used to perform model prediction control and operating condition monitoring for the gallium enrichment and impurity removal process under the new operating conditions.

2. The method for precise dynamic control of the gallium enrichment and impurity removal process according to claim 1, characterized in that, The method for using a Gaussian prediction model for operational condition monitoring is as follows: The control variables for the gallium enrichment and impurity removal process at a preset time delay and the quality indicators of the solution after impurity removal are obtained, input into the Gaussian prediction model, and the quality indicators at the current time are output. Predicted mean and prediction variance ; Based on predicted mean and prediction variance And based on the current moment of the gallium enrichment and impurity removal process Actual values ​​of quality indicators Calculate the probability of switching operating conditions at the current moment. : ； In the formula, The standard normal cumulative distribution function is... The standard normal distribution is at a significance level. The critical value below, For the current moment The standardized residuals of the quality indicators; Let be the confidence radius. ; Probability of operating condition switching With the judgment threshold Comparison, only when At that time, it was determined that the operating conditions of the gallium enrichment and impurity removal process had undergone a substantial change, thereby triggering subsequent model updates.

3. The method for precise dynamic control of the gallium enrichment and impurity removal process according to claim 1, characterized in that, A concept drift correction method is used to obtain a correction matrix, and this correction matrix is ​​then used to map historical operating condition datasets, including: Step 2.1, establish the prediction model based on the parameter evolution model of sparse affine transformation: ； In the formula, Let T be the parameter vector for the new operating condition. For the parameter vector of the old working condition T-1, Used to characterize global physical changes during the impurity removal process; It is a sparse incremental vector, and the number of its non-zero elements is much smaller than the dimension of the parameter vector. It is used to characterize the local characteristic drift during the impurity removal process; Step 2.2: Substitute the parameter evolution model into the prediction model of the impurity removal process to obtain the output relationship of the new operating condition T: ； In the formula, For the impurity removal process, a predictive model is provided under the new operating condition T. For the basis function vector, For the dataset of the new working condition T; time step The basis function vector is represented as , To regulate the variable at time 1 Forward The nonlinear mapping basis function corresponding to each time delay For quality indicators at any time Forward The nonlinear mapping basis function corresponding to each time delay and These are the time delay orders for the control variables and quality indicators, respectively; Step 2.3, considering the prediction model of the old working condition T-1 for the dataset of the new working condition T. The predicted output is , The prediction model for the old working condition T-1 will be used to predict the actual quality indicators of the new working condition T. Rewritten as about the correction vector The linear regression form: ； In the formula, ; Step 2.4, using Based on the sparsity property, the sparse regression algorithm is used to solve for the optimal correction matrix. : ； in, For regularization parameters; Step 2.5, using the optimal correction matrix The large amount of data accumulated from the old operating condition T-1 Perform feature mapping to generate pseudo-labels that fit the dynamic characteristics of the current operating condition T. : ； Step 2.6: Combine the historical dataset with pseudo-labels with the collected small sample dataset of the real new working conditions T. Merge and build an enhanced dataset .

4. The method for precise dynamic control of the gallium enrichment and impurity removal process according to claim 1, characterized in that, During the transition phase of the gallium enrichment and impurity removal process, the constraint boundaries of the control variables are dynamically adjusted in the model predictive control process.

5. The method for precise dynamic control of the gallium enrichment and impurity removal process according to claim 4, characterized in that, The constraint boundaries of the control variables are dynamically adjusted, specifically as follows: Step A1: Express the constraint boundaries of the control variables in probabilistic form: ； In the formula, For a moment The regulatory variables, To regulate the variable at time 1 Boundary constraint coefficients, To regulate the variable at time 1 The upper limit of the constraint, Risk coefficient; Step A2: Based on the probability of operating condition switching And an affine transformation strategy is used to calculate the current time. risk factor : ； in, and For mapping parameters; Step A3: Use the prediction standard deviation output by the Gaussian prediction model Characterizing the uncertainty of the impurity removal process, combined with the inverse cumulative distribution function of the standard normal distribution. Calculate the dynamic contraction parameters of the control variables. : ； in, The adjustment parameters are used to map the uncertainty of the Gaussian prediction model output to the input domain; Step A4, using dynamic compression parameters The constraint boundaries of the control variables are dynamically adjusted to obtain the tightened constraint conditions for the control variables: 。 6. The method for precise dynamic control of the gallium enrichment and impurity removal process according to claim 5, characterized in that, By dynamically adjusting the constraint boundaries of the control variables, the gallium enrichment and impurity removal process is subjected to model predictive control. The optimization problem is formulated as follows: ； In the formula, the optimization objective is... The aim is to find an optimal control sequence. This will make the future The cumulative tracking error within each prediction step and The control increment within each control step is minimized. For quality indicators at any time The set reference value, Gaussian prediction model Quality indicators at any time The predicted value, To regulate the variable at time 1 The control increment, and As a weighting factor, and For the time index of the sample point, and These are the prediction step size and the control step size, respectively. and These are the time delay orders for the control variables and quality indicators, respectively; and These are the upper and lower limits of hard constraints for the control variables.

7. The method for precise dynamic control of the gallium enrichment and impurity removal process according to claim 6, characterized in that, The optimal control sequence of the control variables is obtained by solving the optimization problem of model predictive control using sequential quadratic programming or interior point method.

8. The method for precise dynamic control of the gallium enrichment and impurity removal process according to claim 1, characterized in that, The control variable is the current value used to regulate the acid flow rate.

9. The method for precise dynamic control of the gallium enrichment and impurity removal process according to claim 1, characterized in that, The pH value of the gallium desorption solution was used as a quality indicator.

Images