A dynamic multi-objective parameter optimization method and system for a rare earth extraction process

Abstract

The application discloses a rare earth extraction process dynamic multi-target parameter optimization method and system, and belongs to the technical field of automatic optimization control of a rare earth extraction process. Working condition data is collected by online detection instruments arranged in extraction tanks and washing tanks at all levels of a rare earth extraction production line, is uploaded to an optimization computer after time alignment is completed according to inter-stage material transmission lag steps, and a long short-term memory network dynamic proxy model adapting to inter-stage series time sequence characteristics is constructed. Rare earth ion extraction rates and product purities at all levels are predicted on line and in rolling. An optimization period matching a process dynamic response rate and production operation constraints are set. A multi-objective evolutionary algorithm based on decomposition is used to solve a Pareto optimal solution set. Optimal operation parameters are selected according to a purity priority rule and are issued to an execution mechanism. The application adapts to strong nonlinearity, large lag and slow time variation of process characteristics of the rare earth extraction, and can improve matching degrees of operation parameters and actual working conditions.

Description

Technical Field

[0001] This invention relates to the field of automated optimization control technology for rare earth extraction processes, and in particular to a method and system for dynamic multi-objective parameter optimization in rare earth extraction processes. Background Technology

[0002] Rare earth elements are important strategic resources, and solvent extraction is currently the mainstream process for separating and purifying rare earth elements. Industrial extraction production lines typically employ a cascade extraction architecture with dozens or even hundreds of extraction tanks and washing tanks connected in series. The production process involves multiphase material mass transfer and inter-stage material circulation. The operating conditions are affected by a combination of factors, including fluctuations in feed composition, changes in the degree of saponification of the organic phase, and temperature drift in the tanks, exhibiting strong nonlinearity, large hysteresis, and slow time-varying characteristics. At present, most rare earth extraction production lines still rely on operators to manually adjust operating parameters based on offline detection results. Manual adjustments suffer from strong hysteresis and poor consistency, easily leading to fluctuations in product purity, high extractant consumption, and unstable extraction recovery rates. Some production lines using conventional closed-loop control can only achieve fixed-value adjustments of single parameters, making it difficult to adapt to the dynamic changes in the extraction process and to simultaneously optimize multiple production objectives.

[0003] Existing multi-objective optimization methods for chemical processes suffer from several limitations in their applicability to rare earth extraction scenarios. Conventional mechanistic modeling methods require the calibration of numerous parameters related to extraction reactions and mass transfer, which is challenging and difficult to adapt to the slow time-varying characteristics of the extraction process, leading to a decline in prediction accuracy over long-term operation. Ordinary data-driven proxy models fail to consider the temporal characteristics of the series connection between rare earth extraction tank stages and the lag effect of inter-stage material transport, resulting in significant deviations between predictions and actual production conditions. Conventional multi-objective optimization algorithms do not adapt to the dimensional differences of different optimization objectives, resulting in poor uniformity of the Pareto solution set. Furthermore, many methods employ static optimization logic, with optimization cycles not matching the dynamic response rate of the extraction process. The optimized operating parameters are no longer suitable for the current operating conditions when they are deployed. Some optimization methods also fail to consider the stringent quality requirements of rare earth products when selecting solutions from the solution set, easily leading to the selection of optimal solutions that do not meet product purity compliance requirements.

[0004] Most existing optimization schemes do not consider the actual operating status of the actuators. After the optimization parameters are issued, the execution tracking is not monitored. When the actuators experience malfunctions such as jamming or excessive deviation, it is difficult to detect them in time, which can easily cause production fluctuations. In addition, the online update strategies of most proxy models do not take into account the time-varying laws of the rare earth extraction process. Either the update frequency is too high and it is easy to fit the measurement noise, or the update is too slow and it cannot track the drift of the operating conditions, which further affects the stability of the optimization effect. Summary of the Invention

[0005] This invention proposes a dynamic multi-objective parameter optimization method and system for rare earth extraction processes to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for dynamic multi-objective parameter optimization in a rare earth extraction process, comprising the following steps: S1. Real-time data collection of operating conditions is achieved by online monitoring instruments deployed in the extraction and washing tanks at all levels of the rare earth extraction production line. After the data collected by each monitoring instrument is aligned with the lag step of material transfer between extraction stages, it is uploaded to the process optimization computer via industrial fieldbus. S2. A dynamic proxy model for the rare earth extraction process is constructed based on historical operating data. This model is adapted to the serial time characteristics between rare earth extraction tank stages. It uses the number of extraction tank stages as an index and takes the time sequence of feed flow rate, extractant flow rate and stirring speed of each stage for three consecutive optimization cycles as input. It uses a long short-term memory deep neural network to output the rare earth ion extraction rate and product purity of each stage for the next N optimization cycles, which are matched with the material residence time corresponding to the total number of stages of the extraction production line. The model can perform online rolling prediction of the extraction rate and product purity of each stage. S3. A dynamic multi-objective optimization problem is constructed with the optimization objectives of maximizing the purity of rare earth products, maximizing the rare earth extraction recovery rate, and minimizing the consumption of extractant. The optimization period is set to 1.2 to 1.5 times the average residence time of the material in a single extraction stage to match the dynamic response rate of the extraction process. Constraints are set on the upper and lower limits of the liquid level in each extraction tank, the range of stirring speed, the extraction ratio, and the feed flow rate adjustment rate. A decomposition-based multi-objective evolutionary algorithm is used to collaboratively optimize the feed flow rate, extractant flow rate, and stirring speed at each stage under all constraints to obtain the Pareto optimal solution set. S4. Select a compromise optimal solution from the Pareto optimal solution set according to the preset product purity priority rule as the optimal combination of operating parameters for this optimization cycle. The product purity priority rule is to prioritize the Pareto optimal solution with the highest extraction recovery rate under the premise that the product purity is not lower than the preset minimum purity threshold. The preset minimum purity threshold is the minimum purity requirement of the corresponding rare earth product national standard plus 0.2 percentage points. S5. The optimal combination of operating parameters is output to the feed regulating valve, extractant regulating valve and stirring frequency converter of each extraction tank through the industrial fieldbus as the new control set value, while monitoring the tracking deviation between the actual action value and the set value of each actuator. S6. At the end of each optimization cycle, obtain the deviation between the actual online extraction analysis data and the predicted data of the dynamic agent model, and use the deviation as a model correction signal to perform online incremental updates on the network parameters of the long short-term memory deep neural network. The training sample set for incremental updates only retains the working condition data of the most recent 30 optimization cycles.

[0007] Furthermore, it also includes: when executing the decomposition-based multi-objective evolutionary algorithm in S3, a goal normalization preprocessing step is first performed before weighted Chebyshev aggregation, decomposing the multi-objective optimization problem into multiple single-objective sub-problems. The specific process is as follows: Before each round of optimization, the normalized upper and lower limits of each optimization objective are updated based on the actual working data of the most recent three optimization cycles. Then, min-max normalization is performed on all optimization objectives to convert the actual value of each objective into a dimensionless value in the interval [0,1]. In the formula, For the first An optimization objective is defined in the decision variable vector. The objective function value is below. , The updated numbers are respectively The upper and lower limits of the value of an optimization objective. For the normalized first The optimization objective is dimensionless; for the... An ideal reference point for an optimization objective The dimensionless ideal reference is obtained by using the same normalization method. point; Then, weighted Chebyshev aggregation is used to obtain the objective function of the subproblem: In the formula, For the first The weighted Chebyshev aggregation objective function values ​​for each subproblem are dimensionless. To optimize the total number of targets, For the first The first sub-problem The decision variable vector consists of weights for each optimization objective, where the sum of all weights is one and dimensionless. The variables include three categories: feed flow rate, extractant flow rate, and stirring speed for each extraction tank.

[0008] Furthermore, it also includes: when performing online incremental updates on the long short-term memory deep neural network in S6, a three-stage learning rate adjustment strategy is adopted according to the magnitude of the current deviation. The deviation fluctuation range corresponding to the deviation is set according to the extraction rate deviation range corresponding to the concentration detection error ±0.5g / L under normal operating conditions of the rare earth extraction process.

[0009] Furthermore, the online detection instruments described in S1 include visible-ultraviolet fiber optic probes installed in each extraction tank and washing tank for online detection of rare earth ion concentration in the aqueous phase; capacitive interface meters installed in the clarification chamber of each extraction tank for online detection of the interface position between the organic phase and the aqueous phase; Coriolis mass flow meters installed at the outlet of the extractant storage tank and the outlet of the feed liquid storage tank for online detection of mass flow rate; and torque sensors installed on each stage of the stirring motor for online detection of stirring torque to indirectly reflect the degree of uniformity of material mixing in the tank.

[0010] Furthermore, the long short-term memory deep neural network described in S2 includes an input layer, three stacked long short-term memory hidden layers, a fully connected output layer, and a dropout regularization layer following each long short-term memory layer. The number of neurons in the input layer is equal to the sum of the feed flow rate dimension, the extractant flow rate dimension, and the stirring speed dimension. The number of neurons in each long short-term memory hidden layer is equal to four times the total number of stages in the extraction tank. The fully connected output layer contains two neurons corresponding to the rare earth ion extraction rate prediction value and the product purity prediction value, respectively. The dropout probability of the dropout regularization layer is set to 0.2 during the model training phase and to zero during the online prediction phase.

[0011] Furthermore, the actuators described in S5 include electric regulating valves installed on the feed pipes of each extraction tank, pneumatic diaphragm regulating valves installed on the extractant pipes of each extraction tank, and frequency converters installed on each stirring motor. Each actuator receives the optimal operating parameter setpoints via 4 to 20 mA analog signals or Modbus communication protocol based on RS485 bus and completes closed-loop tracking control in the local proportional-integral-derivative controller. The tracking deviation is calculated by the local proportional-integral-derivative controller transmitting the actual valve position feedback value and speed feedback value back to the process optimization computer at the end of each control cycle.

[0012] Furthermore, the visible-ultraviolet fiber optic spectroscopy probe described in S1 is installed at the outlet of the mixing chamber and the aqueous phase outlet of the clarification chamber of each extraction tank. The part of the probe tip that contacts the extractant is sealed with a sapphire optical window to withstand the strongly acidic working environment inside the extraction tank. The spectral data is transmitted via fiber optic to the spectrometer host installed in the control room of the extraction workshop for wavelength scanning and analysis. The process optimization computer obtains the absorption spectral data of each sampling point in the visible to ultraviolet wavelength range from the spectrometer host and uses the partial least squares regression method to convert the absorption spectrum into the corresponding rare earth ion concentration value.

[0013] Furthermore, a dynamic multi-objective parameter optimization system for rare earth extraction processes includes the following modules: The data acquisition module is used to collect operating data in real time through online detection instruments deployed in the extraction tanks and washing tanks at all levels of the rare earth extraction production line. After aligning the collected data with the lag step of material transfer between extraction process stages, the data is uploaded to the process optimization computer through the industrial fieldbus at a unified sampling period. The proxy model module is used to construct a long short-term memory deep neural network dynamic proxy model based on historical operating data. This model is adapted to the serial time characteristics between rare earth extraction tank stages. It uses the number of extraction tank stages as an index, takes the time sequence of feed flow rate, extractant flow rate and stirring speed of each stage for three consecutive optimization cycles as input, and outputs the rare earth ion extraction rate and product purity of each stage for the next N optimization cycles that match the material residence time corresponding to the total number of stages of the extraction production line. The module also performs online rolling prediction of the extraction rate and product purity of each stage. The multi-objective optimization module aims to maximize the purity of rare earth products, maximize the extraction recovery rate, and minimize the extractant consumption. The optimization cycle is set to 1.2 to 1.5 times the average residence time of the material in each extraction stage to match the dynamic response rate of the extraction process. Under constraints of liquid level, stirring speed, extraction ratio, and adjustment rate, a decomposition-based multi-objective evolutionary algorithm is used to collaboratively optimize the feed flow rate, extractant flow rate, and stirring speed at each stage to obtain a Pareto optimal solution set. The optimal combination of operating parameters is selected according to a preset product purity priority rule. This rule prioritizes the Pareto optimal solution with the highest extraction recovery rate, provided that the product purity is not lower than the minimum purity requirement of the corresponding rare earth product national standard by 0.2 percentage points. The execution output module is used to output the optimal combination of operating parameters to the feed regulating valve, extractant regulating valve and stirring frequency converter of each extraction tank via industrial fieldbus as control setpoints. It also monitors the tracking deviation between the actual action value and the setpoint of each actuator. When the deviation exceeds the limit, it triggers a fault alarm and suspends subsequent optimization. At the end of each optimization cycle, it obtains the deviation between the actual online extraction analysis data and the predicted data of the dynamic agent model. The deviation is used as a model correction signal to perform online incremental updates to the network parameters of the long short-term memory deep neural network. The training sample set for incremental updates only retains the operating condition data of the most recent 30 optimization cycles, which is adapted to the time-varying characteristics of the slow drift of the saponification degree of the organic phase in the rare earth extraction process.

[0014] Furthermore, the multi-objective optimization module incorporates an objective normalization unit, a parameter dynamic update subunit, and a sub-problem aggregation unit, which decompose the multi-objective problem according to the following logic: First, the parameter dynamic update subunit automatically retrieves the actual working condition statistics of the most recent 3 optimization cycles before each round of optimization and updates the normalization upper and lower limits of each optimization objective; then, the objective normalization unit performs min-max normalization processing on all optimization objectives, converting the actual value of each objective into a dimensionless value in the [0,1] interval.

[0015] Furthermore, each actuator in the execution output module includes an electric regulating valve installed on the feed pipeline of each extraction tank, a pneumatic diaphragm regulating valve installed on the extractant pipeline of each extraction tank, and a variable frequency speed controller installed on each stirring motor. Each actuator receives the optimal operating parameter setting value through a 4 to 20 mA analog signal or a Modbus communication protocol based on RS485 bus and realizes closed-loop tracking control in the local proportional-integral-derivative controller.

[0016] Compared with existing technologies, the beneficial effects of this invention are: This invention designs an optimization scheme specifically for the unique properties of rare earth extraction processes, adapting to the temporal characteristics, large hysteresis, and slow time-varying characteristics of cascade extraction. This improves the matching degree between optimized operating parameters and the actual operating state of the extraction process. During the data acquisition phase, this invention performs temporal alignment of detection data at each stage, eliminating data temporal misalignment caused by material transport lag between stages, providing an accurate foundation of operating condition data for subsequent modeling and optimization. The constructed dynamic proxy model combines the inter-stage sequential temporal characteristics of cascade extraction, employing a multi-cycle temporal input and a multi-cycle output structure matching material residence time. This enables more accurate prediction of the changing trends of extraction rates and product purity at each stage after adjusting operating parameters, and the rolling prediction method can also keep up with the dynamic changes in operating conditions in a timely manner.

[0017] The optimization cycle set in this invention matches the average residence time of the extracted single-stage material, avoiding production fluctuations caused by excessively high optimization frequency and preventing untimely responses to changes in operating conditions due to excessively low optimization frequency. The optimization process simultaneously considers multiple constraints such as liquid level, stirring speed, extraction ratio, and adjustment rate, ensuring that the obtained operating parameters meet the actual operational requirements of on-site production and avoid any unexecutable situations. The optimization process first adapts optimization objectives of different dimensions before performing polymerization decomposition, improving the uniformity of the Pareto solution set distribution and providing more options for subsequent solution selection. The product purity-priority solution selection rule ensures that the selected operating parameters meet the stringent requirements for product purity, thereby maximizing the extraction recovery rate and meeting the actual needs of rare earth extraction production.

[0018] This invention synchronously monitors the tracking deviation of the actuator after the optimization parameters are issued. When the deviation is too large, it can promptly issue an alarm and suspend optimization, avoiding production problems caused by continuing optimization under fault conditions. The online incremental update strategy of the proxy model combines the characteristics of slow drift in the saponification degree of the organic phase of rare earth extraction, retaining only recent operating data as training samples. Combined with a learning rate adjustment strategy corresponding to the magnitude of the deviation, it can not only track the slow changes in operating conditions in a timely manner, but also avoid measurement noise in model fitting, thus maintaining the prediction accuracy of the model over a long period of time and improving the long-term stability of the optimization effect. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the dynamic multi-objective parameter optimization method for rare earth extraction proposed in this invention; Figure 2 This is a flowchart of the process data acquisition and timing alignment proposed in this invention; Figure 3 This is a flowchart of the prediction and incremental update process of the dynamic agent model proposed in this invention; Figure 4 This is a flowchart of the multi-objective optimization problem decomposition and collaborative solution proposed in this invention; Figure 5 This is a flowchart of the closed-loop control parameter execution and tracking control proposed in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figures 1 to 5 This invention proposes a dynamic multi-objective parameter optimization method for a rare earth extraction process, comprising the following steps: S1. Real-time data collection of operating conditions is achieved through online monitoring instruments deployed in the extraction and washing tanks of each stage of the rare earth extraction production line. This includes at least the concentration of rare earth ions in the aqueous and organic phases of each stage, the temperature of each tank, the flow rate of extractant and feed liquid, and the speed and current of the stirring motor. After the data collected by each monitoring instrument is aligned with the lag step of material transfer between stages of the extraction process, the data is uploaded to the process optimization computer via the industrial fieldbus. S2. A dynamic proxy model for the rare earth extraction process is constructed based on historical operating data. This model is a long short-term memory deep neural network that adapts to the serial time characteristics between rare earth extraction tank stages. It takes the time sequence of feed flow rate, extractant flow rate and stirring speed of each stage for three consecutive optimization cycles as input, and outputs the rare earth ion extraction rate and product purity of each stage for the next N optimization cycles that match the material residence time corresponding to the total number of stages of the extraction production line. The model is used to make online rolling predictions of the extraction rate and product purity of each stage. S3. A dynamic multi-objective optimization problem is constructed with the optimization objectives of maximizing the purity of rare earth products, maximizing the rare earth extraction recovery rate, and minimizing the consumption of extractant. The optimization period is set to 1.2 to 1.5 times the average residence time of the material in a single extraction stage to match the dynamic response rate of the extraction process. Constraints are set on the upper and lower limits of the liquid level in each extraction tank, the range of stirring speed, the extraction ratio, and the feed flow rate adjustment rate. A decomposition-based multi-objective evolutionary algorithm is used to collaboratively optimize the feed flow rate, extractant flow rate, and stirring speed at each stage under all constraints to obtain the Pareto optimal solution set. S4. Select a compromise optimal solution from the Pareto optimal solution set according to the preset product purity priority rule as the optimal combination of operating parameters for this optimization cycle. The product purity priority rule is to prioritize the Pareto optimal solution with the highest extraction recovery rate under the premise that the product purity is not lower than the preset minimum purity threshold. The preset minimum purity threshold is the minimum purity requirement of the corresponding rare earth product national standard plus 0.2 percentage points. S5. The optimal combination of operating parameters is output to the feed regulating valve, extractant regulating valve and stirring frequency converter of each extraction tank through the industrial fieldbus as the new control set value. At the same time, the tracking deviation between the actual action value and the set value of each actuator is monitored. When the tracking deviation exceeds the preset execution deviation threshold, the actuator fault alarm is triggered and the next optimization cycle is paused. S6. At the end of each optimization cycle, obtain the deviation between the actual online extraction analysis data and the predicted data of the dynamic agent model. Use the deviation as a model correction signal to perform online incremental updates on the network parameters of the long short-term memory deep neural network. The training sample set for incremental updates only retains the working condition data of the most recent 30 optimization cycles to adapt to the time-varying characteristics of the slow drift of the saponification degree of the organic phase in the rare earth extraction process, so as to maintain the model's adaptive tracking ability to the time-varying characteristics of the extraction process.

[0022] This invention also includes: when executing the decomposition-based multi-objective evolutionary algorithm in S3, a target normalization preprocessing step is first performed before weighted Chebyshev aggregation, decomposing the multi-objective optimization problem into multiple single-objective sub-problems. The specific process is as follows: Before each round of optimization, the normalization upper and lower limits of each optimization objective are updated based on the actual working data of the last three optimization cycles. Then, min-max normalization processing is performed on all optimization objectives to convert the actual value of each objective into a dimensionless value in the interval [0,1]. In the formula, For the first An optimization objective is defined in the decision variable vector. The objective function value is below. , The updated numbers are respectively The upper and lower limits of the value of an optimization objective. For the normalized first The optimization objective is dimensionless; for the... An ideal reference point for an optimization objective The dimensionless ideal reference is obtained by using the same normalization method. Points; then, using weighted Chebyshev aggregation, the objective function for the subproblem is obtained: ; In the formula, For the first The weighted Chebyshev aggregation objective function values ​​for each subproblem are dimensionless. To optimize the total number of targets, For the first The first sub-problem The decision variable vector consists of weights for each optimization objective, where the sum of all weights is one and dimensionless. The variables include three categories: feed flow rate, extractant flow rate, and stirring speed of each extraction tank. By normalizing and aggregating as described above, multiple objectives with different optimization directions and dimensions are uniformly transformed into a minimization scalar problem. This improves the search efficiency of the Pareto front and the uniformity of the solution set distribution by utilizing the information sharing mechanism among neighborhood subproblems.

[0023] The present invention also includes: when performing online incremental updates on the long short-term memory deep neural network in S6, a three-stage learning rate adjustment strategy is adopted according to the current deviation amount, wherein the deviation fluctuation range corresponding to the deviation amount is set according to the extraction rate deviation range corresponding to the concentration detection error ±0.5g / L under normal working conditions of rare earth extraction process. When the deviation exceeds the upper limit of the preset deviation fluctuation range, a fast correction is performed using three times the preset base learning rate. When the deviation is within the preset deviation fluctuation range, a smooth correction is performed using the preset base learning rate. When the deviation is below the lower limit of the preset deviation fluctuation range, the learning rate is reduced to one-tenth of the preset base learning rate to prevent the model from overfitting to measurement noise.

[0024] In this invention, the online detection instruments described in S1 include visible-ultraviolet fiber optic spectroscopy probes installed in each extraction tank and washing tank for online detection of rare earth ion concentration in the aqueous phase; capacitive interface gauges installed in the clarification chamber of each extraction tank for online detection of the interface position between the organic phase and the aqueous phase; Coriolis mass flow meters installed at the outlet of the extractant storage tank and the outlet of the feed liquid storage tank for online detection of mass flow rate; and torque sensors installed on each stage of the stirring motor for online detection of stirring torque to indirectly reflect the degree of uniformity of material mixing in the tank.

[0025] In this invention, the long short-term memory deep neural network described in S2 includes an input layer, three stacked long short-term memory hidden layers, a fully connected output layer, and a dropout regularization layer located after each long short-term memory layer. The number of neurons in the input layer is equal to the sum of the feed flow rate dimension, the extractant flow rate dimension, and the stirring speed dimension. The number of neurons in each long short-term memory hidden layer is equal to four times the total number of stages in the extraction tank. The fully connected output layer contains two neurons corresponding to the rare earth ion extraction rate prediction value and the product purity prediction value, respectively. The dropout probability of the dropout regularization layer is set to 0.2 during the model training stage and to zero during the online prediction stage.

[0026] In this invention, the actuators described in S5 include electric regulating valves installed on the feed pipes of each extraction tank, pneumatic diaphragm regulating valves installed on the extractant pipes of each extraction tank, and frequency converters installed on each stirring motor. Each actuator receives the optimal operating parameter setting value through a 4 to 20 mA analog signal or a Modbus communication protocol based on RS485 bus and completes closed-loop tracking control in the local proportional-integral-derivative controller. The tracking deviation is calculated by the local proportional-integral-derivative controller transmitting the actual valve position feedback value and speed feedback value back to the process optimization computer at the end of each control cycle.

[0027] In this invention, one visible-ultraviolet fiber optic spectroscopy probe is installed at the outlet of the mixing chamber and the aqueous phase outlet of the clarification chamber of each extraction tank. The part of the probe front end that contacts the extractant is sealed with a sapphire optical window to withstand the strongly acidic working environment inside the extraction tank. The spectral data is transmitted via fiber optic to the spectrometer host installed in the control room of the extraction workshop for wavelength scanning and analysis. The process optimization computer obtains the absorption spectral data of each sampling point in the visible to ultraviolet wavelength range from the spectrometer host and uses the partial least squares regression method to convert the absorption spectrum into the corresponding rare earth ion concentration value.

[0028] This invention also discloses a dynamic multi-objective parameter optimization system for rare earth extraction processes, comprising the following modules: The data acquisition module is used to collect operating data in real time through online detection instruments deployed in the extraction tanks and washing tanks at all levels of the rare earth extraction production line. After aligning the collected data with the lag step of material transfer between extraction process stages, the data is uploaded to the process optimization computer through the industrial fieldbus at a unified sampling period. The proxy model module is used to construct a long short-term memory deep neural network dynamic proxy model based on historical operating data. This model is adapted to the serial time characteristics between rare earth extraction tank stages. It uses the number of extraction tank stages as an index, takes the time sequence of feed flow rate, extractant flow rate and stirring speed of each stage for three consecutive optimization cycles as input, and outputs the rare earth ion extraction rate and product purity of each stage for the next N optimization cycles that match the material residence time corresponding to the total number of stages of the extraction production line. The module also performs online rolling prediction of the extraction rate and product purity of each stage. The multi-objective optimization module aims to maximize the purity of rare earth products, maximize the extraction recovery rate, and minimize the extractant consumption. The optimization cycle is set to 1.2 to 1.5 times the average residence time of the material in each extraction stage to match the dynamic response rate of the extraction process. Under constraints of liquid level, stirring speed, extraction ratio, and adjustment rate, a decomposition-based multi-objective evolutionary algorithm is used to collaboratively optimize the feed flow rate, extractant flow rate, and stirring speed at each stage to obtain a Pareto optimal solution set. The optimal combination of operating parameters is selected according to a preset product purity priority rule. This rule prioritizes the Pareto optimal solution with the highest extraction recovery rate, provided that the product purity is not lower than the minimum purity requirement of the corresponding rare earth product national standard by 0.2 percentage points. The execution output module is used to output the optimal combination of operating parameters to the feed regulating valve, extractant regulating valve and stirring frequency converter of each extraction tank via industrial fieldbus as control setpoints. It also monitors the tracking deviation between the actual action value and the setpoint of each actuator. When the deviation exceeds the limit, it triggers a fault alarm and suspends subsequent optimization. At the end of each optimization cycle, it obtains the deviation between the actual online extraction analysis data and the predicted data of the dynamic agent model. The deviation is used as a model correction signal to perform online incremental updates to the network parameters of the long short-term memory deep neural network. The training sample set for incremental updates only retains the operating condition data of the most recent 30 optimization cycles, which is adapted to the time-varying characteristics of the slow drift of the saponification degree of the organic phase in the rare earth extraction process.

[0029] In this invention, the multi-objective optimization module incorporates an objective normalization unit, a parameter dynamic update subunit, and a sub-problem aggregation unit. It decomposes the multi-objective problem according to the following logic: First, the parameter dynamic update subunit automatically retrieves the actual operating data from the last three optimization cycles before each optimization round, updating the normalization upper and lower limits of each optimization objective. Then, the objective normalization unit performs min-max normalization on all optimization objectives, converting the actual value of each objective into a dimensionless value in the [0,1] interval. Through the above normalization and aggregation method, multiple objectives with different optimization directions and dimensions are uniformly converted into a minimization scalar problem, thereby utilizing the information sharing mechanism among neighboring sub-problems to improve the search efficiency of the Pareto front and the uniformity of the solution set distribution.

[0030] In this invention, each actuator in the execution output module includes an electric regulating valve installed on the feed pipeline of each extraction tank, a pneumatic diaphragm regulating valve installed on the extractant pipeline of each extraction tank, and a variable frequency speed controller installed on each stirring motor. Each actuator receives the optimal operating parameter setpoint through a 4 to 20 mA analog signal or a Modbus communication protocol based on RS485 bus and implements closed-loop tracking control in the local proportional-integral-derivative controller. At the end of each optimization control cycle, the process optimization computer reads the actual valve position feedback value and the actual speed feedback value from the local controller of each actuator and compares them with the setpoint to calculate the tracking deviation, thereby determining whether to trigger a fault alarm and suspend subsequent optimization calculations.

[0031] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for dynamic multi-objective parameter optimization in a rare earth extraction process, characterized in that, Includes the following steps: S1. Real-time data collection of operating conditions is achieved by online monitoring instruments deployed in the extraction and washing tanks at all levels of the rare earth extraction production line. After the data collected by each monitoring instrument is aligned with the lag step of material transfer between extraction stages, it is uploaded to the process optimization computer via industrial fieldbus. S2. A dynamic proxy model for the rare earth extraction process is constructed based on historical operating data. This model is adapted to the serial time characteristics between rare earth extraction tank stages. It uses the number of extraction tank stages as an index and takes the time sequence of feed flow rate, extractant flow rate and stirring speed of each stage for three consecutive optimization cycles as input. It uses a long short-term memory deep neural network to output the rare earth ion extraction rate and product purity of each stage for the next N optimization cycles, which are matched with the material residence time corresponding to the total number of stages of the extraction production line. The model can perform online rolling prediction of the extraction rate and product purity of each stage. S3. A dynamic multi-objective optimization problem is constructed with the optimization objectives of maximizing the purity of rare earth products, maximizing the rare earth extraction recovery rate, and minimizing the consumption of extractant. The optimization period is set to 1.2 to 1.5 times the average residence time of the material in a single extraction stage to match the dynamic response rate of the extraction process. Constraints are set on the upper and lower limits of the liquid level in each extraction tank, the range of stirring speed, the extraction ratio, and the feed flow rate adjustment rate. A decomposition-based multi-objective evolutionary algorithm is used to collaboratively optimize the feed flow rate, extractant flow rate, and stirring speed at each stage under all constraints to obtain the Pareto optimal solution set. S4. Select a compromise optimal solution from the Pareto optimal solution set according to the preset product purity priority rule as the optimal combination of operating parameters for this optimization cycle. The product purity priority rule is to prioritize the Pareto optimal solution with the highest extraction recovery rate under the premise that the product purity is not lower than the preset minimum purity threshold. The preset minimum purity threshold is the minimum purity requirement of the corresponding rare earth product national standard plus 0.2 percentage points. S5. The optimal combination of operating parameters is output to the feed regulating valve, extractant regulating valve and stirring frequency converter of each extraction tank through the industrial fieldbus as the new control set value, while monitoring the tracking deviation between the actual action value and the set value of each actuator. S6. At the end of each optimization cycle, obtain the deviation between the actual online extraction analysis data and the predicted data of the dynamic agent model, and use the deviation as a model correction signal to perform online incremental updates on the network parameters of the long short-term memory deep neural network. The training sample set for incremental updates only retains the working condition data of the most recent 30 optimization cycles.

2. The method for dynamic multi-objective parameter optimization in a rare earth extraction process according to claim 1, characterized in that, Also includes: When executing the decomposition-based multi-objective evolutionary algorithm in S3, a goal normalization preprocessing step is first performed, followed by weighted Chebyshev aggregation, which decomposes the multi-objective optimization problem into multiple single-objective sub-problems. The specific process is as follows: Before each round of optimization, the normalized upper and lower limits of each optimization objective are updated based on the actual working data of the most recent three optimization cycles. Then, min-max normalization is performed on all optimization objectives to convert the actual value of each objective into a dimensionless value in the interval [0,1]. In the formula, For the first An optimization objective is defined in the decision variable vector. The objective function value is below. , The updated numbers are respectively The upper and lower limits of the value of an optimization objective. For the normalized first The optimization objective is dimensionless; for the... An ideal reference point for an optimization objective The dimensionless ideal reference is obtained by using the same normalization method. point; Then, weighted Chebyshev aggregation is used to obtain the objective function of the subproblem: In the formula, For the first The weighted Chebyshev aggregation objective function values ​​for each subproblem are dimensionless. To optimize the total number of targets, For the first The first sub-problem The decision variable vector consists of weights for each optimization objective, where the sum of all weights is one and dimensionless. The variables include three categories: feed flow rate, extractant flow rate, and stirring speed for each extraction tank.

3. The method for dynamic multi-objective parameter optimization in a rare earth extraction process according to claim 1, characterized in that, Also includes: In S6, when performing online incremental updates on the long short-term memory deep neural network, a three-stage learning rate adjustment strategy is adopted based on the current deviation amount. The deviation fluctuation range corresponding to the deviation amount is set according to the extraction rate deviation range corresponding to the concentration detection error ±0.5g / L under normal operating conditions of rare earth extraction process.

4. The method for dynamic multi-objective parameter optimization in a rare earth extraction process according to claim 1, characterized in that, The online detection instruments described in S1 include visible-ultraviolet fiber optic probes installed in each extraction tank and washing tank for online detection of rare earth ion concentration in the aqueous phase; capacitive interface gauges installed in the clarification chamber of each extraction tank for online detection of the interface position between the organic phase and the aqueous phase; Coriolis mass flow meters installed at the outlet of the extractant storage tank and the outlet of the feed liquid storage tank for online detection of mass flow rate; and torque sensors installed on each stage of the stirring motor for online detection of stirring torque to indirectly reflect the degree of uniformity of material mixing in the tank.

5. The method for dynamic multi-objective parameter optimization in a rare earth extraction process according to claim 2, characterized in that, The long short-term memory deep neural network described in S2 includes an input layer, three stacked long short-term memory hidden layers, a fully connected output layer, and a dropout regularization layer following each long short-term memory layer. The number of neurons in the input layer is equal to the sum of the feed flow rate dimension, the extractant flow rate dimension, and the stirring speed dimension. The number of neurons in each long short-term memory hidden layer is equal to four times the total number of stages in the extraction tank. The fully connected output layer contains two neurons corresponding to the rare earth ion extraction rate prediction value and the product purity prediction value, respectively. The dropout probability of the dropout regularization layer is set to 0.2 during the model training phase and to zero during the online prediction phase.

6. The method for dynamic multi-objective parameter optimization in a rare earth extraction process according to claim 1, characterized in that, The actuators described in S5 include electric regulating valves installed on the feed pipes of each extraction tank, pneumatic diaphragm regulating valves installed on the extractant pipes of each extraction tank, and frequency converters installed on each stirring motor. Each actuator receives the optimal operating parameter setpoints via 4 to 20 mA analog signals or Modbus communication protocol based on RS485 bus and completes closed-loop tracking control in the local proportional-integral-derivative controller. The tracking deviation is calculated by the local proportional-integral-derivative controller sending the actual valve position feedback value and speed feedback value back to the process optimization computer at the end of each control cycle.

7. The method for dynamic multi-objective parameter optimization in a rare earth extraction process according to claim 4, characterized in that, The visible-ultraviolet fiber optic spectroscopy probe described in S1 is installed at the outlet of the mixing chamber and the aqueous phase outlet of the clarification chamber of each extraction tank. The part of the probe in contact with the extractant is sealed with a sapphire optical window to withstand the strongly acidic working environment in the extraction tank. The spectral data is transmitted via fiber optic to the spectrometer host installed in the control room of the extraction workshop for wavelength scanning and analysis. The process optimization computer obtains the absorption spectral data of each sampling point in the visible to ultraviolet wavelength range from the spectrometer host and uses the partial least squares regression method to convert the absorption spectrum into the corresponding rare earth ion concentration value.

8. A dynamic multi-objective parameter optimization system for a rare earth extraction process, applied to the dynamic multi-objective parameter optimization method for a rare earth extraction process as described in any one of claims 1 to 7, characterized in that, Includes the following modules: The data acquisition module is used to collect operating data in real time through online detection instruments deployed in the extraction tanks and washing tanks at all levels of the rare earth extraction production line. After aligning the collected data with the lag step of material transfer between extraction process stages, the data is uploaded to the process optimization computer through the industrial fieldbus at a unified sampling period. The proxy model module is used to construct a long short-term memory deep neural network dynamic proxy model based on historical operating data. This model is adapted to the serial time characteristics between rare earth extraction tank stages. It uses the number of extraction tank stages as an index, takes the time sequence of feed flow rate, extractant flow rate and stirring speed of each stage for three consecutive optimization cycles as input, and outputs the rare earth ion extraction rate and product purity of each stage for the next N optimization cycles that match the material residence time corresponding to the total number of stages of the extraction production line. The module also performs online rolling prediction of the extraction rate and product purity of each stage. The multi-objective optimization module aims to maximize the purity of rare earth products, maximize the extraction recovery rate, and minimize the extractant consumption. The optimization cycle is set to 1.2 to 1.5 times the average residence time of the material in each extraction stage to match the dynamic response rate of the extraction process. Under constraints of liquid level, stirring speed, extraction ratio, and adjustment rate, a decomposition-based multi-objective evolutionary algorithm is used to collaboratively optimize the feed flow rate, extractant flow rate, and stirring speed at each stage to obtain a Pareto optimal solution set. The optimal combination of operating parameters is selected according to a preset product purity priority rule. This rule prioritizes the Pareto optimal solution with the highest extraction recovery rate, provided that the product purity is not lower than the minimum purity requirement of the corresponding rare earth product national standard by 0.2 percentage points. The execution output module is used to output the optimal combination of operating parameters to the feed regulating valve, extractant regulating valve and stirring frequency converter of each extraction tank via industrial fieldbus as control setpoints. It also monitors the tracking deviation between the actual action value and the setpoint of each actuator. When the deviation exceeds the limit, it triggers a fault alarm and suspends subsequent optimization. At the end of each optimization cycle, it obtains the deviation between the actual online extraction analysis data and the predicted data of the dynamic agent model. The deviation is used as a model correction signal to perform online incremental updates to the network parameters of the long short-term memory deep neural network. The training sample set for incremental updates only retains the operating condition data of the most recent 30 optimization cycles, which is adapted to the time-varying characteristics of the slow drift of the saponification degree of the organic phase in the rare earth extraction process.

9. The dynamic multi-objective parameter optimization system for rare earth extraction process according to claim 8, characterized in that, The multi-objective optimization module has a built-in objective normalization unit, a parameter dynamic update sub-unit, and a sub-problem aggregation unit. It decomposes the multi-objective problem according to the following logic: First, the parameter dynamic update sub-unit automatically pulls the actual working condition statistics of the most recent 3 optimization cycles before each round of optimization and updates the normalization upper and lower limits of each optimization objective; then, the objective normalization unit performs min-max normalization processing on all optimization objectives and converts the actual value of each objective into a dimensionless value in the range [0,1].

10. The dynamic multi-objective parameter optimization system for rare earth extraction process according to claim 8, characterized in that, Each actuator in the execution output module includes an electric regulating valve installed on the feed pipeline of each extraction tank, a pneumatic diaphragm regulating valve installed on the extractant pipeline of each extraction tank, and a variable frequency speed controller installed on each stirring motor. Each actuator receives the optimal operating parameter setting value through a 4 to 20 mA analog signal or a Modbus communication protocol based on RS485 bus and realizes closed-loop tracking control in the local proportional-integral-derivative controller.

Images