Temperature precision self-learning control method for chemical impurity removal of metal gallium refining process

Abstract

The application discloses a temperature precision self-learning control method for chemical impurity removal in a metal gallium refining process, which predicts the temperature error by long and short scale statistics, and judges the working condition change according to the ratio; in the working condition transition stage, the BRB-FC method is used to obtain the control quantity to control the chemical impurity removal temperature; wherein, the temperature deviation and the change rate are taken as the input, the control quantity is taken as the output, the relative dominance of each input and rule under different temperature error forms is described through the input and rule weight, the rule conclusion is expanded into the confidence distribution of the multi-grade output through the confidence of the output control quantity, and the ER fusion output control quantity; after the working condition is switched, the temperature prediction model is adaptively updated in two stages: stage I only updates the low-dimensional dense parameters, and stage II performs sparse re-identification to complete model structure correction, so as to perform model prediction control on the chemical impurity removal. The application can provide consistent and repeatable temperature process conditions for the chemical impurity removal in the metal gallium refining process.

Description

Technical Field

[0001] This invention belongs to the field of industrial control technology and relates to a precise self-learning control method for temperature control in the chemical impurity removal process of gallium refining. Background Technology

[0002] With the rapid development of third-generation semiconductors and next-generation information technology, gallium and its compounds have become key materials for manufacturing semiconductors, high-frequency communication devices, optoelectronic devices, and 5G chips. High-purity gallium, as an upstream raw material, directly determines the performance ceiling of downstream devices and the security of the supply chain. Research shows that even trace impurities such as Cu, Pb, Mg, and Zn can significantly reduce device performance. Therefore, industrial production typically requires further increasing the purity of primary gallium from 2N–3N to 4N–5N or even higher. High-purity gallium production generally begins with the extraction of crude gallium from the gallium-containing mother liquor in the alumina process through adsorption, impurity removal, and electrolysis. This crude gallium then enters the refining stage for acid washing, crystal freezing, and packaging. Taking the patent with publication number CN103031450A as an example, crude gallium and pure hydrochloric acid are stirred in a single device and maintained at 40–50°C. Combined with ultrasonic vibration, light metal and heavy metal impurities can be initially separated. Subsequently, heavy metals are precipitated in the temperature range of −1°C–30°C by multiple freezing crystals. After acid washing, stirring and ultrasonication are repeated at 40–50°C to further separate sponge gallium into layers and obtain a product with a purity of 4N–5N.

[0003] Although the aforementioned process routes are relatively mature, temperature control remains a key challenge in achieving stable and compliant chemical impurity removal in the refining stage of actual industrial processes. On the one hand, the reaction rate and phase separation behavior of chemical impurity removal are highly sensitive to temperature windows; temperature fluctuations can easily cause reaction rate drift, impurity re-incorporation, or localized overcooling and solidification, thereby weakening the impurity removal effect and process stability. On the other hand, the coupled superposition of multiple factors such as impurity load, stirring efficiency, steam pressure, and cooling conditions causes the system to be in a state of frequent disturbances and dynamic operating condition switching for a long time. Traditional control methods based on fixed parameters or a single strategy often struggle to simultaneously meet the requirements of "rapidly sensing changes in operating conditions—stable temperature control during the transition period—rapid recovery of accuracy under new operating conditions." After a change in operating conditions, prediction errors often exhibit strong heterogeneity in amplitude and distribution, making commonly used dynamic threshold triggering mechanisms (dynamic sliding window threshold error triggering methods) susceptible to transient spikes, leading to false or missed triggers. Temperature deviations during the transition phase exhibit significant time-varying and non-constant characteristics, making control actions prone to deviation and pushing the system into areas with insufficient historical coverage. Without sufficient information support, deterministic reasoning (fuzzy control methods based on general rule bases) is prone to over-adjustment and amplified temperature fluctuations. Furthermore, most changes in operating conditions manifest primarily as drift in parameters such as heat release intensity and heat transfer coefficient, with only a few cases exhibiting structural differences due to changes in the dominant reaction path. This makes it difficult to simultaneously achieve both rapid parameter correction to restore temperature control accuracy and reliable model reconstruction to avoid long-term deviations under limited real-time computing resources. Summary of the Invention

[0004] To address the challenge of precise temperature control for chemical purification under dynamic operating conditions, this invention provides a self-learning temperature control method for precise chemical purification in gallium refining processes. This method enables the chemical purification temperature to operate stably within the neighborhood of the set value throughout the entire process and under multiple operating conditions, providing consistent and repeatable temperature process conditions for stable purification performance.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0006] A precise self-learning temperature control method for chemical purification in gallium refining processes includes:

[0007] By analyzing the temperature prediction errors of statistical temperature prediction models at both long and short scales, and based on the ratio of the statistical values ​​of the temperature prediction errors at the two scales, the changes in the working conditions of chemical purification during the gallium refining process can be determined.

[0008] During the transition phase after the change of operating conditions, a fuzzy control method based on a confidence rule base is adopted to obtain the control quantity for temperature control of chemical impurity removal in the gallium refining process. The temperature deviation and the rate of change of deviation are used as control inputs and the control quantity is used as output. The relative dominance of each input and rule under different temperature error modes is characterized by the input attribute weights and rule weights. The confidence of the output control quantity is used to expand the rule conclusions into a multi-level output confidence distribution. Then, the multi-rule fusion is performed through evidence reasoning to output the control quantity.

[0009] After switching operating conditions, new operating condition data is collected, and the temperature prediction model based on sparse identification is updated in two stages: in the first stage, the dictionary structure is kept unchanged, and only the low-dimensional dense parameters in the sparse matrix are updated; if the temperature error obtained in the first stage is still abnormal, the second stage is triggered to perform sparse re-identification to obtain the temperature prediction model under the new operating conditions.

[0010] An adaptively updated temperature prediction model was used to perform model predictive control on the chemical impurity removal process in gallium refining.

[0011] Furthermore, the statistical methods for calculating the temperature prediction error statistics at both long and short scales are as follows:

[0012] Step A1, obtain the online sampling time. Measured temperature value Temperature prediction values ​​output by the temperature prediction model Calculate time Temperature prediction error :

[0013] ;

[0014] Step A2, using exponentially weighted moving average to recursively calculate the time. Short-timescale error energy statistics and long-scale error energy statistics :

[0015] ;

[0016] ;

[0017] in, These are the short-scale smoothing factor and the long-scale smoothing factor, respectively.

[0018] Step A3: The ratio of the temperature prediction error statistics for the two time scales is used as the trigger indicator for changes in operating conditions.

[0019] ;

[0020] If the indicator is triggered If the value exceeds the trigger threshold, the operating condition is determined to have changed.

[0021] Furthermore, a fuzzy control method based on a confidence rule base is used to obtain the control quantity, specifically including:

[0022] Step B1: Based on the weights of the two control inputs, namely the chemical impurity removal temperature deviation and the rate of change of deviation, and the membership degree to each fuzzy language level, calculate the antecedent matching degree of each rule:

[0023] ;

[0024] In the formula, For the first Rules Prefix matching degree, For the first One control input For fuzzy language levels membership degree For the first Rule number 1 One control input The weights; For the first One control input In the rules The level of fuzzy language in;

[0025] Step B2, combining the rules weight Match with the predecessor Calculation rules Normalized activation weights :

[0026] ;

[0027] in, For the total number of rules, Index of the rules for summation;

[0028] Step B3, when the first Once a rule is activated, it is executed according to the rules. Output fuzzy language level Give confidence level and its normalized activation weights By fusing the confidence distributions of each rule through an evidence-based reasoning mechanism, the output fuzzy language level is obtained. confidence level :

[0029] ;

[0030] ;

[0031] in, To incorporate the normalization factor, Representation rules Total confidence level assigned to all output levels; and For rules Output fuzzy language level , Confidence level, and To output the index of fuzzy language levels, For rules The number of confidence scores for the output fuzzy language level; The number of fuzzy rules;

[0032] Step B4: Based on the confidence level of each output fuzzy language level and its membership function in the output control domain, the fuzzy set of the output control quantity is obtained as follows:

[0033] ;

[0034] in, To output fuzzy language level In the output control domain The corresponding membership function;

[0035] Step B5, fuzzy set of output control quantity Defuzzification is performed to obtain the output control quantity:

[0036] ;

[0037] in, express Output control quantity at any given time.

[0038] Furthermore, the parameters of the fuzzy control output control quantity , Represents the control input weights. Represents the rule weight. The weights representing the output fuzzy language level are uniformly obtained through offline optimization using an adaptive evolution optimization algorithm based on the projection covariance matrix, where the objective function is:

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] in, This is a performance loss term, measuring the output control quantity. Reference deviation, For use in optimizing parameters The number of samples, for Reference control quantity at any given time. For the regularization part, These are the regularization components of the fuzzy rule's input weights, rule weights, and output fuzzy language level weights, respectively. These are the corresponding regularization coefficients.

[0044] Furthermore, in the first stage, the dictionary structure remains unchanged, and only the low-dimensional dense parameters in the sparse matrix are updated, specifically including:

[0045] Step C1: Let the temperature prediction model obtained before the update based on the sparse identification method be:

[0046] ;

[0047] in, For temperature prediction models Predicted temperature at any given time; These are the state variables and control variables for chemical impurity removal, respectively. The state variable refers to temperature. A dictionary of candidate functions At sample points The value at; It is a sparse matrix, used to extract from... Key terms were selected to characterize the temperature prediction model, and its non-zero elements depicted the dominant dynamic mechanism under the current operating conditions.

[0048] Step C2: Convert the sparse matrix Decompose into column vectors, and divide any 1st column vector into column vectors. The column vectors are denoted as ,extract Non-zero item construction support set :

[0049] ;

[0050] in, Represents column vectors The indexes of each row element in the dataset are the indices of the candidate mechanism items; the support set Used to characterize the set of candidate mechanism terms that are activated. Indicates support set The number of rows, i.e. the number of candidate mechanism terms that are activated;

[0051] Step C3: Maintain the support set Unchanged, from Unit Array Extracting support sets The corresponding rows of the index form the selection matrix. ,in It is a dictionary of candidate functions Size;

[0052] Step C4: Convert the high-dimensional candidate regression vector Projected onto support set The corresponding low-dimensional subspace yields a low-dimensional compact regression vector. and sparse column vectors Projection as low-dimensional compact parameters :

[0053] ;

[0054] ;

[0055] Step C5: Perform recursive least squares fast update under low-dimensional dense representation. .

[0056] Furthermore, step C5 performs a recursive least squares fast update under the low-dimensional dense representation. include:

[0057] Step C5.1: Use low-dimensional dense regression vectors and low-dimensional dense parameters Calculate the temperature prediction model for Prediction error at time :

[0058] ;

[0059] In the formula, Indicates the first A temperature state quantity at time... The value;

[0060] Step C5.2: Calculate the current time. Gain vector :

[0061] ;

[0062] Step C5.3: Recursive Update Timing parameters Covariance Matrix :

[0063] ;

[0064] in, This is a forgetting factor used to enhance the adaptability of parameters to new operating conditions.

[0065] Furthermore, the criterion for determining whether the temperature error obtained in the first stage remains abnormal is: the long-term error energy statistic recorded during the switching of operating conditions is denoted as... Determine the time frame within the data collection window. Time Scale Error Energy Statistics Does it meet the requirements? , This represents the margin coefficient of the reference threshold. If it is met, it indicates that the predicted temperature error within the data collection window is continuously abnormal.

[0066] Furthermore, the second stage involves sparse re-identification to obtain a temperature prediction model under the new operating conditions, specifically including:

[0067] Using the dataset within the data collection window This includes the actual temperature and the control quantity, which are then substituted into the candidate function dictionary. A regression problem is formed using the temperature sequence and the control sequence:

[0068] ;

[0069] in, For the regression objective, the dataset In The actual temperature data of each sample point constitutes the data. ;

[0070] The optimal sparse matrix is ​​obtained by using sparse regression. :

[0071] ;

[0072] In the formula, is the regularization coefficient; thus, the temperature prediction model under the new operating conditions is obtained.

[0073] Furthermore, the specific controlled temperatures include the internal temperature of the vessel and the equivalent temperature of the heat-traced side.

[0074] Furthermore, the control quantity specifically refers to the opening degree of the steam valve.

[0075] To address the challenge of precise temperature control in chemical purification under dynamic operating conditions, this invention proposes a self-learning temperature control method for precise chemical purification in gallium refining processes. This method utilizes a collaborative approach between Model Predictive Control (MPC) – a model expert – and Belief Rule Base-Fuzzy Control (BRB-FC) – a knowledge expert. It uses the relative significance of prediction errors across long and short time scales to perceive changes in operating conditions in real time. During transition phases, the knowledge expert outputs appropriate control actions through a confidence-weighting mechanism. Simultaneously, the two-stage self-learning update significantly accelerates the convergence and updating of the model under new operating conditions, thereby ensuring high-precision and stable control throughout the entire process and under multiple operating conditions.

[0076] First, to address the problem of difficulty in timely and stable identification of changes in operating conditions caused by the coupling of multiple factors, this invention proposes a dynamic error triggering mechanism with long and short time scales. The "relative significance of abrupt changes" is measured by the ratio of short-scale exponential recursion to long-scale exponential recursion, thereby achieving accurate and consistent triggering timing even in the stage of strong error heterogeneity.

[0077] Secondly, to address the problem of insufficient information during the transition phase of operating conditions leading to overly aggressive control actions and increased temperature fluctuations, a Belief Rule Base-Fuzzy Control (BRB-FC) method is proposed. This method explicitly characterizes dominance and uncertainty through attribute weights, rule weights, and output confidence distributions, and utilizes Evidence Reasoning (ER) adaptive fusion to achieve "high-confidence dominance and low-confidence deweighting," thus maintaining a smooth temperature transition before the model update is complete.

[0078] Finally, to address the challenge of balancing "efficient and real-time parameter correction capability" and "reliable structural reconstruction capability" in online model updates for new operating conditions, a two-stage adaptive model update strategy of "parameter fine-tuning - structural reconstruction" is proposed. This strategy involves recursively updating the model on a fixed structural support set and determining whether sparse re-identification update dynamics are triggered, thereby ensuring both update efficiency and structural reliability under resource constraints.

[0079] Through the above-mentioned technical means, the present invention can achieve stable operation of chemical impurity removal temperature within the set value range under the conditions of full process and multiple working conditions, providing consistent and repeatable temperature process conditions for the stability of impurity removal effect. Attached Figure Description

[0080] Figure 1 This is the overall framework of the expert hybrid self-learning control method for the chemical purification temperature in the gallium refining process, as described in the embodiments of this application.

[0081] Figure 2 This is a flowchart illustrating the principle of the dynamic error triggering mechanism on long and short time scales in this application.

[0082] Figure 3 This is a flowchart illustrating the principle of the two-stage adaptive update strategy of "parameter fine-tuning - structural reconstruction" in the embodiments of this application.

[0083] Figure 4 These are the simulation results of the temperature control of chemical impurity removal in the gallium refining process according to embodiments of this application, wherein... Figure 4 (a) and Figure 4 (b) are the normalized temperature control effect diagram and the control quantity fluctuation diagram of the actual reaction in state 1, respectively. Detailed Implementation

[0084] 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.

[0085] This invention provides a precise self-learning temperature control method for chemical purification in gallium refining processes. First, a dynamic error triggering mechanism is constructed across long and short timescales. This mechanism sensitively captures transient and spike disturbances during operating condition transitions through short-scale exponential recursion, and smoothly tracks background bias and long-term error evolution through long-scale exponential recursion. The ratio of these two factors enables stable identification and consistent triggering of operating condition changes. Second, a BRB-FC method is proposed, using attribute / rule weights to characterize dominance and output confidence distribution to depict uncertainty. An ER (Extreme Error Detection) mechanism is used to achieve "high-confidence dominance and low-confidence deweighting," ensuring stable temperature during the transition phase before the prediction model update is complete. Finally, a two-stage adaptive update strategy for the prediction model—"parameter fine-tuning—structural reconstruction"—is proposed. First, prediction accuracy is rapidly restored through recursion on a fixed structural support set. If the error remains abnormal, sparse re-identification is triggered to complete the structural update, achieving rapid recovery of the prediction model and reconstruction of closed-loop temperature control performance under new operating conditions. This ensures that the chemical purification temperature operates stably within the neighborhood of the set value throughout the entire process and under multiple operating conditions. The overall architecture is as follows: Figure 1 As shown.

[0086] Step 1: Monitor operating conditions based on dynamic error changes over both short and long time scales.

[0087] To address the issue that during chemical purification and temperature control, the temperature prediction error after a change in operating condition is subject to abrupt changes in distribution and strong heterogeneity in energy and structure due to the combined effects of mechanistic mismatch bias, closed-loop unsteady-state response, and transient disturbances during switching. This leads to the problem that traditional dynamic thresholds are easily "statistically contaminated" by abnormal samples and inconsistent triggering timing. This embodiment proposes a long-short timescale dynamic error triggering mechanism. This mechanism utilizes the continuous forgetting characteristic of exponential recursion to reduce the dependence of operating condition change triggering judgment on window boundaries and statistical update rhythm: short-timescale error energy statistics sensitively characterize the abrupt change in error energy response, while long-timescale error energy statistics smoothly represent the historical background error level. The ratio of these two statistics measures the "relative significance of the abrupt change," thereby achieving timely, accurate, and consistent triggering of operating condition changes. The algorithm is illustrated below. Figure 2 As shown.

[0088] First, a temperature prediction model for model prediction is constructed in the temperature control loop of gallium refining chemical purification, with the chemical purification reaction temperature as the controlled variable, and the online sampling time is set. The measured temperature value is The prediction model provides a one-step temperature prediction of... The temperature prediction error is defined as follows:

[0089] ;

[0090] and with error energy As a basic input to the trigger statistic, it enhances the sensitivity to sudden biases and transient spikes.

[0091] Then, construct the short-timescale error energy statistic. and long-scale error energy statistics The exponentially weighted moving average (EWMA) is used for recursion:

[0092] ;

[0093] ;

[0094] in, For short-scale and long-scale smoothing factors, usually taken as... . It is highly sensitive to sudden changes in error energy, and is used to quickly capture transient and unsteady peaks during operating condition switching. The mechanism is used to smoothly track the slow evolution of bias and long-term error levels caused by mechanistic mismatch, and to suppress the contamination of background estimation by short-term anomalous spikes.

[0095] Finally, the ratio of the temperature prediction error statistics at the two time scales is used to measure the relative significance of changes in operating conditions, serving as a trigger indicator for these changes.

[0096] ;

[0097] When operating conditions change, the short-timescale error energy statistic rises before the long-timescale one, triggering the indicator. It will deviate significantly from the steady-state level; compared to directly... or Set a dynamic threshold. By mitigating the strong heterogeneity in the absolute magnitude and distribution of prediction errors through relative changes in the "short-to-long" approach, the triggering timing remains consistent under different disturbance intensities and transition patterns. The final triggering determination for the change in operating conditions is as follows:

[0098] ;

[0099] in, To trigger the tag, This is the trigger threshold.

[0100] In this embodiment, the short-scale smoothing factor Long-scale smoothing factor and trigger threshold This is mapped to a target parameter with clear statistical meaning:

[0101] (1) Expected detection step size :

[0102] ;

[0103] Among them, the expected detection step size It is used to characterize the response speed of short-scale EWMA to sudden error changes, and is derived from the relationship of the equivalent window length of EWMA. The smaller the value, the fewer effective samples are retained by the rapid measurement, the more sensitive it is to mutations, and the more agile the triggering. The larger the value, the stronger the smoothing effect and the better the noise reduction, but the more conservative the triggering.

[0104] (2) Baseline estimation accuracy :

[0105] ;

[0106] Among them, baseline estimation accuracy The upper bound of the estimation uncertainty of the long-term baseline error energy variance level is used to constrain the long-scale EWMA, which is approximately approximated when the operating conditions remain unchanged and the model is correct. Its relative standard deviation can be approximated. A smaller value indicates a desired baseline stability and less susceptibility to transient contamination; therefore, a smaller value is required. ; The larger the baseline, the faster the tracking speed, but the lower the stability.

[0107] (3) Allowable false alarm rate :

[0108] ;

[0109] Among them, the allowable false alarm rate Used to give trigger threshold Statistical significance setting: When the operating conditions remain unchanged, the short-timescale error energy statistic can be adjusted. Approximately considered as based on Short-term variance estimate for each sample, long-term error energy statistic Approximately considered as based on The "long-term variance estimate" for each sample is then the ratio statistic. Approximately obey distributed. The smaller the threshold The higher the value, the more conservative the triggering and the fewer false alarms; The larger the threshold, the lower the threshold, making the trigger more sensitive, but increasing the risk of false alarms.

[0110] Step 2: Use BRB-FC to control the temperature during the transition phase.

[0111] When the gallium refining chemical purification process experiences operating condition changes or strong disturbances, the temperature prediction model often suffers short-term mismatch. The temperature prediction error exhibits significant time-varying and non-constant characteristics under closed-loop control, causing the control quantity to deviate and making the system state more likely to touch areas not covered by historical operating conditions. Insufficient available information makes the control more prone to overly aggressive decisions, exacerbating control error fluctuations. To address this, this step proposes the BRB-FC method. This method characterizes the relative dominance of each input and rule under different error patterns through input attribute weights and rule weights. It expands the rule conclusions into a multi-level output confidence distribution using output confidence levels, and achieves multi-rule fusion and "high confidence dominance, low confidence deweighting" through ER (Extreme Error Recovery), thereby stabilizing the chemical purification temperature before the temperature prediction model is updated.

[0112] First, the input and membership are constructed based on the temperature deviation of chemical purification. and the rate of change of deviation As a control input:

[0113] ;

[0114] ;

[0115] in It is the target temperature value;

[0116] For each input Set several language terms (NB, NS, Z, PS, PB), and use membership functions to obtain the membership degrees:

[0117] ;

[0118] in, Indicates input For the first The membership degree of each language item.

[0119] Then, calculate the rule matching degree and activation degree to construct the BRB rules:

[0120] ;

[0121] in, Representing the Rule 1 For rules The weight, For the weights of the input attributes, and Rules respectively Two inputs and Fuzzy language levels, For rules Fuzzy language level of output control quantity Confidence level, For rules middle Quantity, Indicates to All cases are listed and grouped into a set. (Regarding the rules...) Calculate the antecedent matching degree:

[0122] ;

[0123] Combined with the rules Obtain the rule activation degree and normalize it:

[0124] ;

[0125] in, Describe the strength of the rule's contribution in the current transition state. This represents the total number of rules.

[0126] Finally, ER fusion is performed and a control quantity is output. Once a rule is activated, it affects the output level. Give confidence level And it allows for incompleteness. Let the normalized activation weight of this rule be... Then the confidence distributions of each rule can be fused into the overall output confidence score through the ER mechanism. Specifically, we first define the fusion normalization factor. for:

[0127] ;

[0128] Then, the fused first The confidence level for each output level is:

[0129] ;

[0130] in, Representation rules The total confidence level assigned to each output level, and the uncertainty of the rule output corresponding to the unassigned portion; The larger the value, the stronger the evidence for the rule, and the greater the contribution of its confidence distribution to the fusion result. The result obtained after ER fusion... This will naturally reflect the characteristics of "confidence concentration when multiple rules are consistent, and confidence dispersion while retaining uncertainty when there are conflicts." Based on this, the output fuzzy sets are synthesized: Let each output level... In the control range above corresponding membership function Then the output fuzzy set after ER fusion can be written as:

[0131] ;

[0132] right Defuzzification is performed to obtain the control output:

[0133] ;

[0134] In this embodiment, parameters Offline optimization is performed using the Projection Covariance Matrix Adaptive Evolution Strategy (P-CMA-ES) optimization algorithm. The objective function is:

[0135] ;

[0136] in, This is a performance loss term, measuring the output control quantity. Reference deviation, For use in optimizing parameters The number of samples, For the output control quantity of BRB-FC, To provide high-quality data for manual historical output control variables. For the regularization part, it constrains different rules to have a consistent perception of the importance of the same feature, learns the causal structure shared across operating conditions, and encourages the diversity of rule sets. When there is sufficient evidence, it is believed that the confidence distribution is sharp and prevents a few rules from monopolizing the market.

[0137] Step 3: The two-stage model of "parameter fine-tuning - structural reconstruction" adaptively updates the prediction model.

[0138] To address the challenges of frequent and complex temperature changes during the chemical purification process in gallium refining, existing online updates to prediction models often struggle to simultaneously balance real-time performance and reliability. On one hand, most changes manifest as variations in parameters such as heat transfer coefficients and exothermic intensity; directly triggering full re-identification would incur high computational costs and lengthy data collection wait times. On the other hand, some conditions involve changes in the acid pickling reaction state, leading to increases or decreases in effective kinetic terms. If only parameter recursive fine-tuning is performed, structural mismatch in the prediction model will result in persistent and amplified prediction errors. Therefore, this step employs a two-stage adaptive model update strategy: parameter fine-tuning followed by structural reconstruction. First, temperature prediction accuracy is rapidly restored recursively on a given structural support set. Then, sparse re-identification is triggered when error statistics remain abnormal to correct the structure. This enables rapid and robust self-learning modeling support for chemical purification temperatures in computationally limited industrial settings. The algorithm flowchart is shown below. Figure 3 As shown.

[0139] (1) First stage: parameter fine-tuning. That is, the dictionary structure remains unchanged, and only the low-dimensional dense parameters in the sparse matrix are updated.

[0140] After the trigger mechanism determines the change in operating conditions, record the time scale at this moment. for The candidate function dictionary structure remains unchanged, and only the low-dimensional dense parameters are updated rapidly on the current support set to obtain a new sparse matrix.

[0141] Step C1: Let the old temperature prediction model obtained based on the Sparse Identification of Nonlinear Dynamical Systems (SINDy) be:

[0142] ;

[0143] in, These are system state variables and control variables, respectively. A dictionary of candidate functions At sample points The value at; It is a sparse matrix, used to extract from... The key terms selected in the model characterize the temperature prediction model, and their non-zero elements describe the dominant dynamic mechanism under the current operating conditions. sparse matrix Multiplying the i-th column gives the i-th column. The state equation obtained from the first... The value at time k+1 of each state.

[0144] Step C2: Convert the sparse matrix Decompose into several column vectors Extract the non-zero terms to construct the corresponding support set:

[0145] ;

[0146] Among them, superscript Indicates the first The index of each state equation, the support set Used to characterize the set of candidate mechanism terms activated in this model system, i.e., the first... In each output dimension, which dictionary terms have non-zero coefficients?

[0147] Step C3: Construct the selection matrix. The first stage parameter update assumes the structure of the prediction model remains unchanged, which is equivalent to the hypothesis support set. It remains unchanged in a short period of time. To reduce the computational complexity of online calculations, from Unit Array Extracting support sets The corresponding rows of the index form the selection matrix. ,in It represents the total number of dictionary entries for candidate functions, i.e., the size of the dictionary.

[0148] Step C4: Obtain the complete high-dimensional candidate regression vector Project the vectors onto the low-dimensional subspace corresponding to the current support set to obtain a low-dimensional dense regression vector; and project the sparse parameters into low-dimensional dense parameters as well.

[0149] ;

[0150] ;

[0151] The projection operation described above is equivalent to retaining only the candidate function terms selected under the current operating conditions, thereby transforming the original high-dimensional sparse regression problem into a low-dimensional dense parameter estimation problem.

[0152] Step C5: Perform a fast update using Recursive Least Squares (RLS) on this low-dimensional dense representation. .

[0153] Step C5.1: Calculate the prediction error:

[0154] ;

[0155] in, From the current moment Input construct, and parameters For the previous moment The estimated value is used to reflect the causal time series of the recursive estimate.

[0156] Step C5.2: Calculate the gain vector:

[0157] ;

[0158] Step C5.3: Recursively update the effective parameters and covariance matrix:

[0159] ;

[0160] in, This is a forgetting factor used to enhance the adaptability of parameters to new operating conditions. The core idea of ​​Phase I is to rapidly fine-tune only the key parameters while keeping the support set unchanged, thereby restoring the model's prediction accuracy with extremely low online computational cost.

[0161] (2) Second stage: structural reconstruction. That is, to obtain the temperature prediction model under the new operating conditions by performing sparse re-identification based on the collected data.

[0162] If the first phase is completed and data has been collected, the window is open. Subsequently, the long- and short-scale error statistics remained abnormal. , As a margin factor for the reference threshold, this embodiment differs from the trigger threshold in the first part. If the values ​​are the same, it indicates that there may be a missing effective dynamic term or a change in the dominant mechanism under the current working condition; at this time, the second stage I: structural reconstruction is triggered. (Using a window...) Data Substitute into the dictionary matrix containing commonly used candidate function terms. Each candidate function term is a "dictionary atom" in the dictionary, forming a regression problem with the temperature sequence and the control sequence:

[0163] ;

[0164] in, For the regression objective, the dataset In The actual temperature data of each sample point constitutes the data. Finally, sparse regression is used to obtain the optimal sparse matrix. :

[0165] ;

[0166] This leads to a temperature prediction model for the new operating conditions.

[0167] Step 4: Implement model predictive control for the new operating conditions.

[0168] The updated temperature prediction model from step 3 is then injected back into the model expert—model predictive control—to restore rolling optimization control. During the temperature prediction model update, the knowledge expert maintains transitional stability, thereby achieving "rapid recovery at a controllable cost + reliable reconstruction when necessary," ensuring that the chemical purification temperature operates stably within the setpoint neighborhood under various operating conditions.

[0169] To verify the effectiveness of the method of the present invention, an experiment was designed for verification:

[0170] This study focuses on a gallium refining chemical purification field device, which features typical characteristics such as stirring and mixing, steam heating, and resistance temperature measurement. The process is equivalently represented at the control level as a lumped-parameter thermodynamic system consisting of "mixing temperature inside the vessel - thermal potential on the heating side." In actual operation, the temperature evolution inside the vessel is mainly determined by two channels: first, the thermal inertia and heat dissipation of the material inside the vessel lead to temperature self-holding and slow decline; second, changes in the steam valve opening cause changes in the thermal potential on the heating side, which drive the temperature response inside the vessel via jacket-vessel heat transfer coupling. Factors such as fluctuations in steam supply pressure, thermal disturbances introduced by acid / water addition, and exothermic / endothermic reactions cause nonlinearity and operational condition correlation in the temperature response, resulting in different dynamic gains and time constants for the same control action under different batches, different loading amounts, or different heat transfer conditions. To improve the numerical stability of model identification and control optimization, and to avoid parameter scale inconsistencies caused by differences in the dimensions of different physical quantities, the experiment performed uniform interval scaling and dimensionless processing on the state variables of the chemical purification temperature control process. Specifically, the temperature inside the vessel... Equivalent temperature of the heat-traced side Perform linear mappings within their respective operational intervals to construct normalized states. This ensures that the control quantity falls within a bounded interval under typical operating conditions, thereby maintaining a reasonable amplitude for the nonlinear function term and improving the condition number for regression fitting. This corresponds to the percentage of the steam valve opening on site.

[0171] Based on the above mechanism, this experiment uses temperature and valve position command data collected on-site to perform equivalent fitting on the process in the discrete time domain. In order to characterize the switching of working conditions in real production, this experiment further identifies parameter sets in different batches or different working condition intervals, so that the dynamic differences caused by changes in the amount of material, attenuation of heat transfer due to scaling, or fluctuations in steam capacity in the actual process can be reproduced in the simulation environment. Specifically, it is manifested in the change of set value and the adjustment of internal parameter structure.

[0172] The control results of this experiment are as follows: Figure 4 As shown, Figure 4 (a) The vertical axis represents the actual reaction temperature. The normalized value within the operable range, which is the first state variable. The normalized value, Figure 4 (b) The vertical axis represents the control quantity. This refers to the percentage of the steam valve opening on site; the setpoint sequence is as follows: The RMSE of the control error was 0.061874, the RMSE of the prediction error was 0.054969, and the variance of the control variable was 0.005984. The experiment verified the setpoint tracking and disturbance scenarios covering multiple temperature rise / fall switching and changes in internal parameter structure. The experiment maintained stable tracking and consistent control performance throughout the entire process, fully demonstrating the accuracy, reliability, and engineering applicability of the method of this invention in temperature control under various operating conditions during the chemical purification process of gallium refining.

[0173] This invention presents an expert-driven hybrid self-learning temperature control scheme for chemical purification in gallium refining. It integrates the rolling optimization capabilities of model experts with the transition robustness of knowledge experts, achieving precise temperature control that is "triggerable by operating condition changes, stable during transition phases, and self-learning of model parameters." This scheme is applicable to the automatic adjustment of steam heating and cooling circuits in the chemical purification reaction of gallium refining processes. It can maintain stable temperature operation under long-term disturbances and frequent switching conditions, and can be further extended to other temperature-sensitive process control scenarios in gallium production.

[0174] 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 precise self-learning temperature control method for chemical impurity removal in gallium refining processes, characterized in that, include: By analyzing the temperature prediction errors of statistical temperature prediction models at both long and short scales, and based on the ratio of the statistical values ​​of the temperature prediction errors at the two scales, the changes in the working conditions of chemical purification during the gallium refining process can be determined. During the transition phase after the change of operating conditions, a fuzzy control method based on a confidence rule base is adopted to obtain the control quantity for temperature control of chemical impurity removal in the gallium refining process. The temperature deviation and the rate of change of deviation are used as control inputs and the control quantity is used as output. The relative dominance of each input and rule under different temperature error modes is characterized by the input attribute weights and rule weights. The confidence of the output control quantity is used to expand the rule conclusions into a multi-level output confidence distribution. Then, the multi-rule fusion is performed through evidence reasoning to output the control quantity. After switching operating conditions, new operating condition data is collected, and the temperature prediction model based on sparse identification is updated in two stages: in the first stage, the dictionary structure is kept unchanged, and only the low-dimensional dense parameters in the sparse matrix are updated; if the temperature error obtained in the first stage is still abnormal, the second stage is triggered to perform sparse re-identification to obtain the temperature prediction model under the new operating conditions. The adaptively updated temperature prediction model is used for model predictive control of chemical impurity removal in the gallium refining process. The statistical method for the temperature prediction error statistics at both long and short scales is as follows: Step A1, obtain the online sampling time. Measured temperature value Temperature prediction values ​​output by the temperature prediction model Calculate time Temperature prediction error : ； Step A2, using exponentially weighted moving average to recursively calculate the time. Short-timescale error energy statistics and long-scale error energy statistics : ； ； in, These are the short-scale smoothing factor and the long-scale smoothing factor, respectively. Step A3: The ratio of the temperature prediction error statistics for the two time scales is used as the trigger indicator for changes in operating conditions. ； If the indicator is triggered If the value exceeds the trigger threshold, the operating condition is determined to have changed.

2. The precise self-learning temperature control method for chemical purification in gallium refining processes according to claim 1, characterized in that, The control quantity is obtained using a fuzzy control method based on a confidence rule base, specifically including: Step B1: Based on the weights of the two control inputs, namely the chemical impurity removal temperature deviation and the rate of change of deviation, and the membership degree to each fuzzy language level, calculate the antecedent matching degree of each rule: ； In the formula, For the first Rules Prefix matching degree, For the first One control input For fuzzy language levels membership degree For the first Rule number 1 One control input The weights; For the first One control input In the rules The level of fuzzy language in; Step B2, combining the rules weight Match with the predecessor Calculation rules Normalized activation weights : ； in, For the total number of rules, Index of the rules for summation; Step B3, when the first Once a rule is activated, it is executed according to the rules. Output fuzzy language level Give confidence level and its normalized activation weights By fusing the confidence distributions of each rule through an evidence-based reasoning mechanism, the output fuzzy language level is obtained. confidence level : ； ； in, To incorporate the normalization factor, Representation rules Total confidence level assigned to all output levels; and For rules Output fuzzy language level , Confidence level, and To output the index of fuzzy language levels, For rules The number of confidence scores for the output fuzzy language level; The number of fuzzy rules; Step B4: Based on the confidence level of each output fuzzy language level and its membership function in the output control domain, the fuzzy set of the output control quantity is obtained as follows: ； in, To output fuzzy language level In the output control domain The corresponding membership function; Step B5, fuzzy set of output control quantity Defuzzification is performed to obtain the output control quantity: ； in, express Output control quantity at any given time.

3. The precise self-learning temperature control method for chemical purification in gallium refining processes according to claim 2, characterized in that, Parameters of fuzzy control output control quantity , Represents the control input weights. Represents the rule weight. The weights representing the output fuzzy language level are uniformly obtained through offline optimization using an adaptive evolution optimization algorithm based on the projection covariance matrix, where the objective function is: ； ； ； ； in, This is a performance loss term, measuring the output control quantity. Reference deviation, For use in optimizing parameters The number of samples, for Reference control quantity at any given time. For the regularization part, These are the regularization components of the fuzzy rule's input weights, rule weights, and output fuzzy language level weights, respectively. These are the corresponding regularization coefficients.

4. The precise self-learning temperature control method for chemical purification in gallium refining processes according to claim 2, characterized in that, In the first stage, the dictionary structure remains unchanged, and only the low-dimensional compact parameters in the sparse matrix are updated, specifically including: Step C1: Let the temperature prediction model obtained before the update based on the sparse identification method be: ； in, For temperature prediction models Predicted temperature at any given time; These are the state variables and control variables for chemical impurity removal, respectively. The state variable refers to temperature. A dictionary of candidate functions At sample points The value at; It is a sparse matrix, used to extract from... Key terms were selected to characterize the temperature prediction model, and its non-zero elements depicted the dominant dynamic mechanism under the current operating conditions. Step C2: Convert the sparse matrix Decompose into column vectors, and divide any 1st column vector into column vectors. The column vectors are denoted as ,extract Non-zero item construction support set : ； in, Represents column vectors The indexes of each row element in the dataset are the indices of the candidate mechanism items; the support set Used to characterize the set of candidate mechanism terms that are activated. Indicates support set The number of rows, i.e. the number of candidate mechanism terms that are activated; Step C3: Maintain the support set Unchanged, from Unit Array Extracting support sets The corresponding rows of the index form the selection matrix. ,in It is a dictionary of candidate functions Size; Step C4: Convert the high-dimensional candidate regression vector Projected onto support set The corresponding low-dimensional subspace yields a low-dimensional compact regression vector. and sparse column vectors Projection as low-dimensional compact parameters : ； ； Step C5: Perform recursive least squares fast update in low-dimensional dense representation. .

5. The precise self-learning temperature control method for chemical purification in gallium refining processes according to claim 4, characterized in that, Step C5 performs a recursive least squares fast update in the low-dimensional dense representation. include: Step C5.1: Use low-dimensional dense regression vectors and low-dimensional dense parameters Calculate the temperature prediction model for Prediction error at time : ； In the formula, Indicates the first A temperature state quantity at time... The value; Step C5.2: Calculate the current time. Gain vector : ； Step C5.3: Recursive Update Timing parameters Covariance Matrix : ； in, This is a forgetting factor used to enhance the adaptability of parameters to new operating conditions.

6. The precise self-learning temperature control method for chemical purification in gallium refining processes according to claim 4, characterized in that, The criterion for determining whether the temperature error obtained in the first stage remains abnormal is: the long-term error energy statistic recorded during the switching of operating conditions is denoted as... Determine the time frame within the data collection window. Time Scale Error Energy Statistics Does it meet the requirements? , This represents the margin coefficient of the reference threshold. If it is met, it indicates that the predicted temperature error within the data collection window is continuously abnormal.

7. The precise self-learning temperature control method for chemical purification in gallium refining processes according to claim 4, characterized in that, The second stage involves sparse re-identification to obtain a temperature prediction model under the new operating conditions, specifically including: Using the dataset within the data collection window This includes the actual temperature and the control quantity, which are then substituted into the candidate function dictionary. A regression problem is formed using the temperature sequence and the control sequence: ； in, For the regression objective, the dataset In The actual temperature data of each sample point constitutes the data. ; The optimal sparse matrix is ​​obtained by using sparse regression. : ； In the formula, is the regularization coefficient; thus, the temperature prediction model under the new operating conditions is obtained.

8. The precise self-learning temperature control method for chemical purification in gallium refining processes according to claim 1, characterized in that, The specific controlled temperatures include the internal temperature of the vessel and the equivalent temperature of the heat-traced side.

9. The precise self-learning temperature control method for chemical purification in gallium refining processes according to claim 1, characterized in that, The control quantity specifically refers to the opening degree of the steam valve.

Images